Optimized and Efficient Color Prediction Algorithm

electronics
Article
Optimized and Efficient Color Prediction Algorithms Using
Mask R-CNN
Rajesh Kannan Megalingam *, Balla Tanmayi, Gadde Sakhita Sree, Gunnam Monika Reddy,
Inti Rohith Sri Krishna and Sreejith S. Pai
Department of Electronics and Communication Engineering, Amrita Vishwa Vidyapeetham,

Amritapuri 690525, India
* Correspondence: rajeshm@am.amrita.edu
Abstract: Color cognizant capability has a significant impact in service robots for object detection
based on color, traffic signal interpretation for autonomous vehicles, etc. Conventional clustering
algorithms such as K-means and mean shift can be used for predicting the dominant color of an
image by mapping the pixels from RGB to HSV and clustering them based on HSV values, thereby
picking the cluster with the most pixels as the dominant color of the image, but these approaches
are not solely dedicated to the same outcome. This research’s goal is to introduce novel techniques
for predicting the dominant color of objects in images, as well as pixel extraction concepts, which
allow these algorithms to be more time and efficiency optimized. This investigation appraises
propriety of integrating object detection and color prediction algorithms. We introduce a dominant
color prediction color map model and two new algorithms: average windowing and pixel skip. To
predict objects in an image prior to color prediction, we combined the Mask R-CNN framework with
our proposed techniques. Verification of our approach is done by creating a benchmark dataset of
200 images and comparing color predicted by algorithms with actual color. The accuracy and runtime
of existing techniques are compared with those of the proposed algorithms to prove the superiority
of our algorithms. The viability of the proposed algorithms was demonstrated by scores of 95.4%
accuracy and color prediction time of 9.2 s for the PXS algorithm and corresponding values of 93.6%
and 6.5 s for the AVW algorithm.
Citation: Megalingam, R.K.; Keywords: average windowing (AVW); pixel skip (PXS); dominant color prediction color map
Tanmayi, B.; Sree, G.S.; Reddy, G.M.;
(DCPCM); HSV (hue saturation value); Mask R-CNN; object detection; color prediction
Krishna, I.R.S.; Pai, S.S. Optimized
and Efficient Color Prediction
Algorithms Using Mask R-CNN.
Electronics 2023, 12, 909. https://
1. Introduction
doi.org/10.3390/electronics12040909
Proliferating deployment of robotics and autonomous applications in diverse aspects
Academic Editor: Chiman Kwan
of everyday life in recent times has aroused keen interest of researchers in the domain
Received: 16 December 2022 of detection and identification of objects. Such attributes outfit robots with a capability
Revised: 25 January 2023 to mimic human behavior. Deep learning techniques have become popular these days
Accepted: 26 January 2023 for implementing object detection. Some of the applications of deep learning techniques
Published: 10 February 2023 apart from object detection include facial recognition [1,2], gesture recognition [3,4], health
care [5,6], image enhancement [7], etc. Object detection assumes a dominant role in the field
of automation, harmonized with increased precision and simplification of work. Object
detection gadgets are conspicuous in video surveillance, security cameras, and self-driving
Copyright: © 2023 by the authors. systems [8]. Object detection models introduced in the past decade include region-based
Licensee MDPI, Basel, Switzerland.
convolutional neural networks (R-CNN), Fast R-CNN, you only look once (YOLO), and
This article is an open access article
Mask R-CNN [9]. The deployed models are enhanced versions of primitive CNN models.
distributed under the terms and
CNN models are also widely used for object recognition and classification [7]. YOLO
conditions of the Creative Commons
and R-CNN are esteemed for their accurate detection of objects [10,11]. The pretrained
Attribution (CC BY) license (https://
model employed in Mask R-CNN is used for detecting objects in the input image and as
creativecommons.org/licenses/by/
4.0/).
an input for our proposed algorithms that are discussed in the later sections. Dominant
Electronics 2023, 12, 909. https://doi.org/10.3390/electronics12040909 https://www.mdpi.com/journal/electronics

Electronics 2023, 12, 909 2 of 25
color extraction from an image is a frequently used technique in applications such as image
segmentation, medical image analysis, image retrieval, and image mixing [12]. Although
many algorithms are available for detection of the dominant color of an object, they are
based on clustering techniques that are deficient in time complexity or in accuracy in
color prediction. Hence, we have introduced two algorithms based on new techniques for
predicting the dominant color of an object in the image. The two algorithms, AVW and
PXS, use averaging and skipping techniques respectively, to select pixels from an input
image. These algorithms have huge importance in terms of time as most of the pixels in the
image are eliminated while predicting color (which will be discussed in the later sections)
to save time without compromising on accuracy. The DCPCM model predicts the color of
all input pixels given by the proposed algorithms.
Initially, YOLO was selected for object detection, as it is one of the fastest object
detection models [10]. But YOLO does not provide a mask along with the bounding box for
the detected object, which results in reduced accuracy, as some background pixels inside
the bounding box that are not part of the object may contribute to the color prediction and
may disrupt the final color predicted. Isolation of the pixels contributing to the background
of the object, enclosed within the masked portion of AVW and PXS algorithms, facilitate
increased accuracy in the predictions of the (dominant) color of an object. Concurrently,
the runtime for predicting the object’s color is significantly diminished as the number of
pixels considered for prediction gets reduced. Unlike other object detection algorithms,
Mask R-CNN, outfitted with a segmentation mask for each object detected, is preferred for
integration with our proposed algorithms purposes. Mask R-CNN has also been found to
be more efficient in detecting the type of vehicles, for video surveillance and self-driving
systems [13], where color prediction of objects can be an added advantage. Nevertheless,
the AVW and PXS algorithms are flexible enough to be unified with any application that
requires color prediction. The canny edge detection approach is used to identify the
nonuniformity percentage for every object detected in an image. This technique detects
a wide variety of edges in images by noise suppression [14], yet it is applicable only to
grayscale images, returning a binary matrix of dimensions, stating whether the respective
pixel contributes to an edge or not. Average pixel width (APW) is the average of the entire
binary matrix. APW is more likely to give the percentage of nonuniformity in an image [15],
which is used to model the algorithms propounded in Section 5.
HSV is a color model of cylindrical shape that transforms the red green blue (RGB)
color model into easily comprehensible dimensions [16]. Hue determines the angle of color
on the HSV and RGB circles. Saturation represents the amount of color being used, and
value represents the level of brightness of the respective color. Supervened by its balanced
color perception capability, HSV color space is DCPCM model [17]. Elaborated in Section 3,
the entire HSV color space, marshalled in the proposed model, is mapped into 15 basic
color classes to provide a better understanding and readability of the DCPCM’s output.
The existing techniques of color prediction are based on clustering. Clustering refers to the
process of grouping the available data points (pixels in the context of color prediction) into
different clusters based on their similarity with certain parameters such as color, features,
etc. Clustering is commonly applied to predict the dominant color of an image [18,19];
K-means, mean shift, and fuzzy C-means are a few popular clustering models [20]. Hence,
this calls for a simple dedicated algorithm for prediction of the dominant color in an image.
These clustering algorithms excel in various machine learning applications [21–26].
An edge detector scheme, adaptable for real-time applications, is presented in [27];
the results presented in [27] show that the proposed method of edge detection outperforms
other existing methods in terms of accuracy. A model that discerns humans from the
background in a video and identifies the actions performed by him or her was proposed
in [28,29]. It introduced a scheme to discriminate between red and black grapes using
image detection and color prediction schemes. A hierarchical interactive refinement net-
work, which facilitates effective preservation of edge structures while detecting salient
objects, was illustrated in [30]. An efficient method that exploited mean-shift clustering,
Electronics 2023, 12, 909 3 of 25
was elucidated in [31], for the discovery of common visual patterns (CVPs) between two
images. An enhanced version of Mask R-CNN for the detection of cucumber in green-
houses, was demonstrated in [32]. Recently [33], reports were published on an application
of HSV combined with K-means clustering in the detection of food spoilage. A comparative
analysis of the execution times between the software implementation and register-transfer
level (RTL) design of the Sobel edge detection algorithm was presented in [34]. A scheme
for identification of an object, based on recognition of its color and contour, through digital
image processing techniques was presented in [35]. A technique for the segregation of ob-
jects in an image based on their color, alleviates the complexities of boundary detection [36].
A hybrid method that extracts the spatial information and color for identification of the
moving objects in a video was presented [37].
We summarize below the limitations of the clustering algorithms in color prediction.
1. The clustering algorithms are not exclusively designed for color prediction. They also
excel in various machine learning applications;
2. The K-value is not easy to predict in K-means algorithm, which does not go well with
global clusters. In addition, the time complexity of K-means increases if the datasets
are bigger;
3. Mini batch K-means algorithm has lower accuracy than K-means;
4. Time complexity is a major disadvantage in mean shift algorithm;
5. GMM takes more time to converge and hence slower than K-means;
6. Fuzzy C-means requires greater number of iterations for better results.
The purpose of this study is to address the limitations of clustering algorithms in
color prediction. We propose a model that is exclusively used for color prediction and
not for any other purpose. In addition, we propose two new algorithms that can extract
dominant colors accurately in less time. In view of this, a dominant color prediction color
map (DCPCM) model and two new algorithms, average windowing (AVW) and pixel skip
(PXS), are presented in this work. To predict the color of the object in an image, we need to
detect the object first. For this purpose, we integrate the Mask R-CNN algorithm with our
proposed techniques to predict the objects in the image prior to color prediction. Rather
than considering all the pixels of the object, the AVW algorithm takes the average of the
section/window of the pixels from the image at a time and repeats the averaging for the
entire object in the image. Because of this, the overall time for color prediction decreases.
At the same time, it maintains good accuracy because the window size is chosen according
to the uniformity of the color in the object. On the other hand, the PXS algorithm skips the
pixels selectively based on the uniformity of the color in the object. This results in less time
for color prediction without compromising much accuracy.
The main contributions of this research work are listed below:
1. A color prediction model called DCPCM that is exclusively designed by uniquely
categorizing HSV domain values into 15 different color classes, to be used only for
color prediction and not for any other purpose. The main aim of the given model
is to reduce the color spectrum into 15 commonly used colors, thereby reducing the
complexity and the runtime.
2. Two new algorithms called AVW and PXS to selectively extract pixels from an image,
using precomputed formulae, thereby reducing the runtime and maintaining the
accuracy;
3. Integration of Mask-RCNN algorithm with the proposed techniques for identifying
the objects in the image prior to pixel extraction and dominant color prediction;
4. Creation of benchmark data set with 200 images of single object, multi-object, uni-
formly colored, multicolored and of various sizes, to test the proposed algorithms and
compare them with the existing clustering techniques stated above.
The rest of the paper is organized as follows. Section 2 summarizes the entire working
of the system using an architecture diagram. The technical aspects of the work start from
Section 3, by presenting the DCPCM model, which is then followed by Section 4, describing
from Section 3, by presenting the DCPCM model, which is then followed by Section 4,
describing AVW and PXS algorithms in a detailed manner. Section 5 presents the experi-
mental results followed by Section 6, which deals with proving the supremacy of AVW
Electronics 2023, 12, 909 and PXS over the existing techniques for color prediction. Section 7 provides discussion
4 of 25
on the results and concludes the work.
2. System
AVW Architecture
and PXS algorithms in a detailed manner. Section 5 presents the experimental results
Theby
followed system architecture
Section diagram
6, which deals shown in
with proving theFigure 1 presents
supremacy theand
of AVW workflow
PXS overofthe
the
proposed AVW and PXS algorithms that predict the object class using the Mask R-CNN
existing techniques for color prediction. Section 7 provides discussion on the results and
framework,
concludes thealong
work.with its color property. Consider an input image that contains two
bowls, a knife, and a spoon, as shown in Figure 1. When the proposed algorithm
2. System Architecture
AVW/PXS is executed, this image is assigned and stored in the form of a matrix. Subse-
The system
quently, architecture
the matrix diagram
triggers the objectshown
detectionin Figure
function1 presents the workflow
that predicts the objectsofalong
the
proposed AVW and PXS algorithms that predict the object class using the
with other parameters like bounding boxes, and mask color. If objects are detected in theMask R-CNN
framework,
image, thenalong with its color
the proposed property.store
algorithms Consider an inputofimage
the location thati.e.,
objects, contains two bowls,
its bounding box
acoordinates,
knife, and a along
spoon,with
as shown in Figure 1. When the proposed algorithm AVW/PXS
its masked portion. After prediction, the masked pixels are fil- is
executed, this image is assigned and stored in the form of a matrix. Subsequently,
tered out and sent as an input to the algorithm block, i.e., AVW or PXS. Simultaneously, the matrix
triggers the object
preprocessing detection
happens for function
the giventhat predicts
object the objects
to calculate a fewalong with other
parameters parameters
required for the
like bounding boxes, and mask color. If objects are detected in the image, then the proposed
algorithms using canny edge detection.
algorithms store the location of objects, i.e., its bounding box coordinates, along with its
masked portion. After prediction, the masked pixels are filtered out and sent as an input to
the algorithm block, i.e., AVW or PXS. Simultaneously, preprocessing happens for the given
object to calculate a few parameters required for the algorithms using canny edge detection.
Figure 1. System architecture of dominant color prediction using AVW and PXS algorithms. The
Figureimage
output 1. System
in thearchitecture
given figureof detected
dominant5 objects
color prediction
along withusing AVW and color.
the dominant PXS algorithms. The
The detection
output image in the given figure detected 5 objects along with the dominant color. The detection
results in the output image (left to right) are as follows: gray dining table (0.974); cyan bowl (0.961);
results in the output image (left to right) are as follows: gray dining table (0.974); cyan bowl (0.961);
orange bowl (0.989); pink spoon (0.975); and red knife (0.989). The output format is as follows:
orange bowl (0.989); pink spoon (0.975); and red knife (0.989). The output format is as follows: dom-
dominant
inant color color of predicted
of the the predicted object,
object, predicted
predicted objectobject class (confidence
class (confidence score
score of the of the predicted
predicted object).
Each detected object is surrounded by a bounding box.
object). Each detected object is surrounded by a bounding box.
The
TheDCPCM
DCPCMmodel
modelisisintegrated
integratedwith
withboth
bothAVW
AVWandandPXS
PXSalgorithms
algorithmsthat
thatmap
mappixels
pixels
tototheir respective color class among the fifteen predetermined colors. After the executionof
their respective color class among the fifteen predetermined colors. After the execution
of AVW/PXS, the output is displayed in a separate window that contains the objects
AVW/PXS, the output is displayed in a separate window that contains the objects sur-
surrounded by their bounding boxes, the object’s class, and its dominant color. This process
rounded by their bounding boxes, the object’s class, and its dominant color. This process
repeats for every object detected by the Mask R-CNN framework. If no objects are found in
repeats for every object detected by the Mask R-CNN framework. If no objects are found in
the image, then the algorithm simply displays “No object detected” and the same input
image is shown as the output in the Jupyter notebook. The DCPCM model skips the
color prediction step if the object detected is a person to avoid controversies as the model
can predict the costume color as the color of a person. This architecture is helpful in
classification of similar objects based on color that have wide applications in the service
robot sector.
Electronics 2023, 12, 909 5 of 25
3. Dominant Color Prediction Color Map (DCPCM) Model

There are many shades of different colors available in each color space. Some of the
color spaces are HSV, cyan magenta yellow key (CMYK), and RGB. Generally exposed
to various color shades or trained, human beings are capable of predicting the color of
an object merely by a glance at an object, unlike a machine that perceives an image in
the form of a pixel constituted of three channels: red, green and blue. In order to predict
the dominant color of an object, algorithms are required to iterate through all the pixels
of an object image, and maintain counts of the various colors, ensuing the prediction of
the dominant color. In order to generalize the HSV color space, all the available color
shades are classified into 15 color classes: red, green, blue, yellow, cyan, magenta, orange,
yellow-green, green-cyan, cyan-blue, violet, pink, gray, white, and black. Fifteen colors
are chosen because they are the ones most widely used that include primary, secondary,
tertiary, and neutral colors. HSV values range from [0,360], [0,100], [0,100]. Hence, the
range of each pixel is 9 × 7 × 7 bits. The DCPCM model extracts each pixel in an object
and maps it into one of the 15 predetermined color classes.
The saturation (S) and value (V) parameters of the HSV color model are split into three
ranges—low, medium, and high. These ranges are shown in Equations (1) and (2).

 Low i f Se[0, 13]
S1 ( S ) = Medium i f Se[14, 50] Ranges of “S” parameter (1)
High i f Se[51, 100]

S1 is the name of the range in which “S” belongs.


 Low i f Ve[0, 15]
V1 (V ) = Medium i f Ve[16, 75] Ranges of “V” parameter (2)
High i f Ve[76, 100]

V 1 is the name of the range in which “V” belongs.

The main objective of the DCPCM model is to predict colors for various combinations
of the low, medium, and high ranges. The nine combinations available from different
ranges are shown in Table 1. Among these nine combinations of “S” and “V” parameters,
for five combinations the color is treated as black, gray, or white, whereas the colors for
the remaining four combinations are based on values of the “H” parameter. These four
combinations are shown in Equations (3)–(6).
Table 1. Combinations of “S” and “V” parameters.
Saturation (S1 ) Value (V 1 ) Color

Low Low Black
Low Medium Gray
Low High White
Medium Low Black
Medium Medium Only Tertiary colors (3)
Medium High Secondary and Tertiary colors (4)
High Low Black
High Medium Primary and Tertiary colors (5)
High High All colors (6)
if H (◦ )e[15, 45]

 Orange
Yellow − Green if H (◦ )e[46, 110]




Green − Cyan if H (◦ )e[111, 175]


Color = (3)

 Cyan − Blue if H (◦ )e[176, 240]
Violet if H (◦ )e[241, 310]





Pink otherwise
Electronics 2023, 12, 909 6 of 25
Predictions for the combination S = “medium” and V = “medium”.
if H (◦ )e[15, 40]


 Orange
Yellow if H (◦ )e[41, 60]








 Green − Cyan if H (◦ )e[111, 150]



Color = Cyan if H (◦ )e[151, 190] (4)
Cyan − Blue if H (◦ )e[191, 225]




Violet if H (◦ )e[226, 280]




if H (◦ )e[281, 330]

 Magenta



 Pink otherwise
Predictions for the combination S = “medium” and V = “high”.
if (H (◦ )e[0, 10] or H (◦ )e[331, 360])



 Red
Orange if H (◦ )e[11, 50]








if H (◦ )e[76, 135]



 Green
Color = Green − Cyan if H (◦ )e[136, 165] (5)
Cyan − Blue if H (◦ )e[166, 210]




Blue if H (◦ )e[211, 265]




if H (◦ )e[266, 295]




 Violet
 Pink otherwise
Predictions for the combination S = “high” and V = “medium”.

 Low i f Se[0, 13]
S1 (S) = Medium i f Se[14, 50] Ranges of “S” parameter (6)
High i f Se[51, 100]

Predictions for the combination S = “high” and V = “high”.
4. AVW and PXS Algorithms

This section presents the terminologies used in the proposed algorithm first, followed
by relevant details of the AVW and PXS algorithms. The all-pixel approach for predicting
the dominant color of the object using the DCPCM model is discussed first. Later, we move
on to present the AVW algorithm and the PXS algorithm.
A detailed list of symbols used for modeling AVW and PXS is given in Table 2.
Table 2. Symbols used for modeling AVW and PXS.
Saturation (S1 ) Value (V 1 )

M Length of the Original Image.
N Width of the Original Image.
This represents the Masked Color of the Original Image. This notation
[Mc ]1 × 3 implies that Im is a matrix of dimension 1 × 3 (since it has 3 channels i.e.,
R, G, B).
[Io ]m × n Original Image Data. This notation implies that Io is a matrix of size m × n.
[Im ]m × n Masked Image Data. This notation implies that Im is a matrix of size m × n.
R Variable that iterates through rows of [Io ] and [Im ].
C Variable that iterates through columns of [Io ] and [Im ].
Dc Dominant color of the image.
Matrix representing the result of DCPCM model for the respective pixel.
[K]m × n × 15
The number 15 in the notation represents the predetermined colors.
T Length of list containing masked pixels.
Electronics 2023, 12, 909 7 of 25
4.1. Preprocessing Required for AVW and PXS Algorithms

This section lists the terminologies used while defining the algorithms AVW and PXS.
The terminologies and definitions presented here are common to both AVW and PXS. Let
(y1, x1), and (y2, x2) be the top-left and bottom-right coordinates of the bounding box for
an object predicted in the input image after the image passes through the Mask R-CNN
block shown in Figure 1.
Let the object currently predicted be referred to as the current object.
1. Masked Color, represents the RGB values of the mask chosen by Mask R-CNN, for the
current object;
2. Masked_Image, is the image obtained after application of the respective mask, Mask
R-CNN, to the Original_Image;
3. OriginalImageData, is the matrix that stores the pixel representations of the Origi-
nal_Image in terms of RGB channels.
4. MaskImageData, is the matrix that stores the pixel representations of the Masked_Image
in terms of RGB channels.
5. Color_Set, is an array that stores the predetermined colors, by allocations of an index
to each of the 15 colors and represented as [CSET ]1*15 . This notation implies that the
Color_Set, is a 1*15 matrix i.e., a 1-D array that stores all the 15 colors. [CSET ] = [“Red”,
“Orange”, “Yellow”, “Yellow-Green”, “Green”, “Green-Cyan”, “Cyan”, “Cyan-Blue”,
“Blue”, “Violet”, “Magenta”, “Black”, “Pink”, “White”, “Gray”].
6. R2H (r, g, b), represents a function that takes in the RGB values of a pixel and returns
the respective HSV value.
7. C_P (h, s, v), represents a function that takes in the HSV values of a pixel and returns
the respective color class. This pertains to implementation of the DCPCM model,
discussed in the previous section.
8. [Gscale ], is the grayscale representation of an original image [Io ].
9. Edgecanny (Am ), represents the canny edge detector function, discussed previously in
Section 1. It gives us the APW i.e., average number of pixels contributing to edges
in the given matrix Am , where Am is the resultant output, after the [Gscale ] passes
through the edge_filter.
10. Edge_filter filters out the pixels contributing to the unmasked portion of the current
object. Since the uniformity is based on the edges, the pixel contributing to the
unmasked portion inside the bounding box is converted to “white”. This operation
will exclude the unmasked portion for uniformity prediction while maintaining the
null spaces in the input matrix.
4.2. All-Pixel Approach

Let [A]1*15 be the 1-D array where each index corresponds to the respective color class
and [K] be the matrix representing the result of the DCPCM model for the respective pixel.
The steps listed below are preceded by the Mask R-CNN block presented in the system
architecture in Figure 1. The steps given below are followed for every object detected in the
input image.
Step-1: Initialize array A to zero.
Step-2: Iterate through the current_object and increment the respective color of the pixel in
[A], if the pixel is inside the bounding box of the current_object and the respective
pixel of the masked image is equal to the Masked_Color.
Equations (7) and (8), show a mathematical representation of [A] and [K] for the
all-pixel approach. At this stage, the pixel value should be converted from RGB to HSV
color space, using R2H (r, g, b).
y2 x2
[ A] = ∑ ∑ K [r ][c] if Im [r][c] = [ Mc ]. (7)
r =y1 c= x1
Electronics 2023, 12, 909 8 of 25

1, CSET [i] = C_P(h, s, v)
K [r ][c][i ] = , ∀ i e [0, 15), where h, s, v = R2H ( Io [r ][c]). (8)
0, CSET [i] 6= C_P(h, s, v)
Step-3: Iterate through [A] and find the color with the maximum count which gives us the
Dc (dominant color) of the image. Equation (9), represents the mathematical form
of Step-3.
\
Dc = CSET [i ], where A[i ] > A[ j], ∀ ( j e [0, 15)) ( j! = i ) . (9)
The time complexity of all-pixel approach is O (m × n), where m and n are the width
and height of the given object, while the space complexity is O (m × n × 15).
4.3. Average Windowing (AVW) Algorithm

The AVW algorithm decreases the time of the all-pixel approach discussed in Section 4.2
to a greater extent. A window of size “f ” is used for implementation of this algorithm. This
window slides through the pixels and predicts the average of the RGB channels, at each stop.
If the window size is too high, the average of the pixel values may get disrupted and the
prediction of dominant color may tend to lose accuracy, whereas if the window size is too
small, the run time does not vary much compared to the all-pixel approach. Hence, the value
of “f ” should be chosen precisely. The concept of uniformity is used to predict “f ” that is
discussed later in Section 4.3.1. The value of f can vary according to the uniformity factor of
the image i.e., the value of “f ” increases with an increase in the uniformity factor and vice
versa. This algorithm reduces the time complexity to O(T/f) where T is the list of pixels in
masked portion of the object and “f ” is the window size, but the space complexity remains
the same i.e., O(m × n × 15).
A detailed explanation of the AVW algorithm is presented below. The steps listed are
preceded by the Mask R-CNN block presented in the system architecture in Figure 1. The
steps given below are followed for every object detected in the input image.
Step-1: Initialize [A] to zero.
Step-2: Convert the input image into grayscale and store the data into [Gscale].
Step-3: Iterate through the image and store the pixels from the OriginalImageData that con-
tribute to the masked portion of the MaskedImageData and are inside the bounding
box coordinates of the current_object in a list (L) represented in Equation (10).
[ L] = Io [r ][c]i f Im [r ][c] = [ Mc], ∀ {r e W, y1 <= r <= y2 }& {c e W, x1 <= r <= x2 }. (10)
Step-4: Perform edge detection and predict “f ” (Detailed in Section 4.3.1).
Step-5: Iterate through the list and compute the average of the pixels in each window
and increment the respective count in [A]. The mathematical representation of
[A], and the result array [K], for AVW are shown in Equations (11) and (12). The
representation of average of the HSV values for a particular window is shown in
Equation (13).
y2 x2
[ A] = ∑ ∑ [K] if Im [r][c] = [ Mc ]. (11)
r=y1 c=x1

1, CSET [i] = C_P(h, s, v)
K [r ][c][i ] = , ∀ i e[0, 15). (12)
0, CSET [i] 6= C_P(h, s, v)
j+ f
1 T
[h, s, v] = [ ∗ ∑ R2H ( L[i ])]; j = n ∗ f , n 0, . (13)
f i= j
f
Electronics 2023, 12, 909 9 of 25
Step-6: Iterate through [A], and find the color with the maximum count, equivalent to
the Dc (dominant color) of the image. The extraction of dominant color from the
Color-set, is represented in Equation (14).
\
Dc = CSET [i ], where A[i ] > A[ j], ∀ ( j e [0, 15)) ( j! = i ) . (14)
4.3.1. Predicting Window Size (“f ”)

Initially, the input image is converted into a grayscale image and stored in [Gscale ] array.
For each object predicted, the respective bounding box from the image is considered and the
unmasked pixels are changed to the same color (i.e., white) so that they do not contribute
to the nonuniformity of color in the image. This modified array is sent as an input for the
edge detection algorithm. This gives an APW value (refer point-9 in Section 4.1) which is
used for predicting “f ”. APW of the current_object is more likely to predict the uniformity
of an object. If APW is 0, the object is 100% uniform with respect to color. The object is
considered as 100% nonuniform with respect to color if the APW is 100.
Let “k” be the number of pixels contributing to edges.
APW
k= × T. (15)
100
Let us assume that the k-pixels are uniformly divided in the object (average case). If
we consider that 1 window has 1 pixel contributing to edge (assuming that 1 value does
not disrupt the average value), then, f = T/k.
100
According to Equation (15), f = T × APW × T , which on simplification yields Equation (16).
100
f = . (16)
APW
4.3.2. Pseudo Code of AVW

The pseudo code for the AVW function is shown below. It takes in the list (L) containing
the masked pixels, along with the window size “f ” and returns the array A, which contains
the count of each color class for the input pixels.
AVW (L,f ): //L is the list containing the extracted pixels of the object and “f ” is the
window size
j=0
while “j” < length (L)
var h, s, v <- (0,0,0) //Defining variables h, s and v.
for every “f ” pixels. //Loop runs “f ” times
h, s, v <- (h + L[j][0], s + L[j][1], v + L[j][2])
increment “j”
end for
//At this point, h, s, and v variables contains sum of the current “f ” pixel values.
var h_avg,s_avg,v_avg <- (avg(h),avg(s),avg(v))
var l <- color prediction (h_avg,s_avg,v_avg) //l is the color predicted for the given values of
h_avg,s_avg and v_avg.
A[l]++//A is the array containing count of each of the 15 colors. At this step, we are inceasing the
count of the color class “l” in the array A.
end while
return A
4.3.3. Corner Cases for AVW

Case-1: Object is completely uniform: In this case, the average of all the pixels in the object
will give us the dominant color. So, “f ” can be as high as possible.
Proof: According to Equation (16), f = 100/APW.
Image is completely uniform ⇒APW = 0.
So, APW = 0 ⇒f → ∞.
Electronics 2023, 12, 909 10 of 25
Case-2: Object is 50% uniform: This is the case where every alternate pixel is an edge pixel.
Hence, the average of two pixels should be considered for color prediction if
according to the assumption that every window should have one edge pixel.
Proof: f = 100/APW.
Image is 50% uniform ⇒ APW = 50.
Hence, f = 2.
Case-3: Uniformity is above 50%: When uniformity is above 50%, the value of f tends to
decrease below 2. Since the window size is a positive integer; “f ” becomes 1
for uniformity greater than 50%. Hence, the algorithm changes into the all-pixel
approach.
Proof: Since f = 1, every window contains only one pixel whose average results
in the same value. Hence, all pixels are considered for prediction and the
threshold of APW for implementing the AVW algorithm is 50. For all other
values, the algorithm turns into the all-pixel approach.
4.4. Pixel Skip (PXS) Algorithm

The all-pixel approach that considers all the pixels to find the dominant color in the
current object image has a higher execution time to predict the output. We therefore propose
an algorithm named PXS, where pixels are selectively skipped, based on the uniformity
factor of the object. Without affecting accuracy, this algorithm consumes less time than the
all-pixel approach.
PXS algorithm compares two pixels of the current_object that are inside the masked
portion and separated by (S-2) pixels. If the compared pixels are of the same color, then
all the (S-2) pixels are skipped and are assumed to be the same color as their corner pixels.
Otherwise, the (S-2) pixels are skipped and not considered for prediction of the dominant
color. The prediction accuracy and runtime of the algorithm mainly depends on the “S”
factor. If the “S” factor increases, then the number of pixels skipped in the object increases,
which may lead to decrease in accuracy. If the “S” factor decreases, then the time difference
between all-pixel and PXS is minimal. Hence, the value of “S” should be chosen precisely.
The value of “S” is predicted using the uniformity approach, mentioned previously. A brief
description of selecting “S” will be discussed later in this section. This algorithm reduces
the time complexity to O(T/S) where “T” is the list of pixels in masked portion of the object
and “S” is the skip size but the space complexity remains the same i.e., O(m × n × 15).
A detailed explanation of the PXS algorithm is given below. The steps listed below are
preceded by the Mask R-CNN block presented in the system architecture in Figure 1. The
steps given below are followed for every object detected in the input image.
Step-1: Initialize [A] to zero.
Step-2: Convert the image into grayscale and store the data into [Gscale ].
Step-3: Iterate through the image and store the pixels from the OriginalImageData that con-
tribute to the masked portion of the MaskedImageData and are inside the bounding
box coordinates of current_object in a list(L) shown in Equation (17).
[L] = Io [r ][c] i f Im [r ][c] = [ Mc], ∀ {r e W, y1 <= r <= y2 } & {c e W, x1 <= r <= x2 }. (17)
Step-4: Perform edge detection and predict the PXS factor “S”. The number of pixels
skipped is considered to be (S-2).
Step-5: Let “p” be the current pixel. If [p] = [Mc ] and C_P(p) = C_P (p − S + 1), skip the
middle pixels assuming that they are of the same color and increment the count of
the respective color by “S”. If, C_P(p) 6= C_P (p − S + 1), skip the middle pixels
without considering them for color prediction. If the pixel does not belong to the
masked region, discard the current pixel and move to the next pixel and repeat the
Electronics 2023, 12, 909 11 of 25
same step. The mathematical representation of [A] and [K] for PXS is shown in
Equations (18) and (19).
y2 x2
x2 − x1

[ A] = ∑ ∑ [K], I f Im [r][c] = [ Mc ]; n e 0, S . (18)
r =y1 c= x1+n∗S


 S, ( CSET [i ] = C P(h,s,v)

 (K [r ][c] = K[r ][c − s + 1]))

 T
K [r ][c][i ] =
 1, ((CSET [i] = C_P(h, s, v)) ∀ i e [0, 15); (19)
(K [r ][c] 6= K[r ][c − s + 1]))

 T


0, CSET [i] 6= CP(h,s,v)

Step-6: Iterate through [A] and find the color with the maximum count that gives us the Dc
(dominant color) of the image. The extraction of dominant color from the Color_set,
is shown in Equation (20).
\
Dc = CSET [i ], where A[i ] > A[ j], ∀ ( j e [0, 15)) ( j! = i ). (20)
4.4.1. Predicting Skip Size (S)

As noted in the case of AVW, “k” is considered to be the number of edges in the image,
and the edges are considered to be uniformly distributed throughout the image (average
case). Unlike AVW, rather than taking the average of the entire window, the pixels inside
the window (between edge pixels) are assumed to be of similar color, as no edge in between
implies that the pixels are uniform.
Hence,
100
S= . (21)
APW
Since, (S-2) is the number of pixels to be skipped, the value of (S-2) should be a
non-negative integer. Hence, considering the corner cases,
100
(S − 2) = Max(0, ). (22)
APW
4.4.2. Pseudo Code for PXS

The pseudo code given below indicates the PXS function. It takes in the list (L)
containing the masked pixels along with the skip size “S” and returns the array A that
contains the count of each color class for the input pixels.
def.PXS (L,s): //L is the list containing the extracted pixels of the object and s is the skip size
for “j” in length(L)
var x = color_prediction(L[j])
var y = color_prediction(L[j + s-1])
if x = y,
A[x] = A[x] + s
else
A[x]++
A[y]++
J=j+s
end for
return A
Electronics 2023, 12, 909 12 of 25
4.4.3. Corner Cases

Case-1: Object is completely uniform: In this case, if the first and last pixel colors are equal
for an object, then the dominant color will be of the first and last pixel color; else
the dominant color will be either first or last pixel color. So, S → ∞.
Proof: S = 100/APW. (According to Equation (21))
Image is completely uniform ⇒APW = 0.
So, APW = 0 ⇒ S → ∞.
Case-2: CUniformity is greater than or equal to 50%: When uniformity is greater than or equal
to 50%, the value of “S” tends to be a negative value. Since, the number of pixels
to be skipped is a non-negative integer (according to Equation (22)), (S-2) becomes
0, for uniformity greater than or equal to 50%. Hence, the algorithm transitions
into the all-pixel approach.
Proof: Since (S-2) = 0, every pixel is compared with the corresponding pixel. If the
compared pixels belong to the same or different color each pixel will be considered
for maximum color prediction of an object. Hence, the threshold of APW for
implementation of the PXS algorithm is 50.
For all other values greater than or equal to 50 of APW, the algorithm becomes an
all-pixel approach.
5. Experimental Results
A benchmark data set was created with 200 images, each of a different size, to test the
proposed algorithms. This data set is composed of objects with different categories like
single-color objects, multicolored objects, and objects with various sizes with respect to
the image size. The proposed algorithms AVW and PXS are tested with the benchmark
data set, with three iterations each, where the accuracy of prediction and the average of the
runtime are noted.
The testing process was automated such that when we start the iteration, it fetches
an image from the benchmark data set and triggers the object detection framework. The
output from the Mask R-CNN block presented in Figure 1 is directed to the AVW or PXS
algorithms which display a picture with a bounding box, color and class of all detected
objects in the image. Simultaneously, the result data, i.e., color of the predicted objects
and the time taken for predictions, are stored in an Excel file. This is repeated until all the
images in the data set are completed. The Excel file created is used to perform statistics,
such as average time taken, accuracy and, standard deviation of time. The accuracies of
the AVW and PXS algorithms, using the benchmark data set, are calculated manually. The
percentage of correct predictions of the objects, in each image of the benchmark data set,
is factored in to calculate the overall accuracy. The entire testing process stated above is
implemented in Jupyter Notebook in MacOS with 1.4 GHz Quad-Core Intel Core i5 CPU
and Intel Iris Plus Graphics 645 1536 MB GPU properties.
Object color predictions, using PXS and AVW algorithms, resulted in two output
images, the extracted image and the predicted image. Figures 2a and 3a, show the processed
images using AVW and PXS respectively, where bounding boxes are drawn around each
object, and also the color and object name are displayed together on top of the bounding box.
For the image containing two cups in Figure 2, the color predicted by both the algorithms
is identical, but the time consumed to make the predictions is not.
Electronics 2023, 12, x FOR PEER REVIEW 13 of 26
Electronics 2023, 12, 909 13 of 25

box. For the image containing two cups in Figure 2, the color predicted by both the algo-
rithms
box. Foristhe
identical,
image but the timetwo
containing consumed
cups in to make2,the
Figure thepredictions is not.by both the algo-
color predicted
rithms is identical, but the time consumed to make the predictions is not.
(a) Processed image. (b) Extracted image.

(a) Processed image.
Figure 2. [Predicted and extracted] output images by the(b) Extracted
proposed image.
AVW algorithm.
Figure 2. [Predicted and extracted] output images by the proposed AVW algorithm.
Figure 2. [Predicted and extracted] output images by the proposed AVW algorithm.
(a)Processed image. (b)Extracted image.

(a)Processed image.
Figure (b)Extracted image.
Figure 3.
3. [Predicted
[Predicted and
and extracted] output images
extracted] output images by
by the
the proposed PXS
proposed PXS algorithm.
algorithm.
Figure 3. [Predicted and extracted] output images by the proposed PXS algorithm.
Figures
Figures2b 2bandand3b 3bshow
showthe theextracted
extractedimages
imagesofofAVW AVWand andPXS,
PXS,respectively,
respectively,which
which
present
present the
the pixels
Figures 2b andcontributing
pixels 3b show thetoextracted
contributing tocolor
colorprediction.
images ofTo
prediction. display
AVW
To andthe
display PXS,extracted
the output,
respectively,
extracted an an
which
output,
image of RGB values (235, 235, 235) is created with the same size as
present the pixels contributing to color prediction. To display the extracted output, anThe
image of RGB values (235, 235, 235) is created with the same size as the
the original
original image.
image.
The
imageextracted
extracted image
of RGB image of
of the
values the 235,
AVW
(235, AVW algorithm
algorithm
235) shown
is created shown in
theFigure
in Figure
with 2bobtained
2b is
same sizeisas
obtained by by assigning
assigning
the original image.the
average color
the average of
color the corresponding
of the corresponding window
window to all the
to all pixels inside
the pixels that
inside particular window.
that particular
The extracted image of the AVW algorithm shown in Figure 2b is obtained by assigning
For the
window. PXS algorithm,
Forcolor
the PXS the pixels
algorithm, that contributed
the pixels to the
that contributedcolor prediction
to the are
colorthat represented
prediction are by
the average of the corresponding window to all the pixels inside particular
predicted color, while pixels that are discarded by algorithm are represented as background
represented
window. Forbythepredicted color, while
PXS algorithm, pixelsthat
the pixels thatcontributed
are discarded by color
to the algorithm are repre-
prediction are
RGB value, (235, 235, 235), as shown in Figure 3b.
sented as background
represented by predicted RGB value,
color, (235,pixels
while 235, 235),
that as shown
are in Figure
discarded by 3b.
algorithm are repre-
Figure 4 presents the flow of color prediction, in different stages of AVW and PXS
sentedFigure 4 presents RGB
as background the flow of (235,color235,
prediction, in different stages 3b.of AVW and PXS
algorithms. Stage 1 is the value, original image 235),
whichasundergoes
shown in Figure
color prediction. Stage 2
algorithms. 4Stage
Figurethe 1 is the
presents the original
flow of image which undergoes color prediction. Stageand
2 repre-
represents extracted images forcolor
bothprediction,
the algorithms,in different
with the stages
finalof AVW
predicted PXS
output
sents the
algorithms. extracted
Stage 1 images
is the for
originalboth the
image algorithms,
which undergoeswith the
colorfinal predicted
prediction.
obtained at stage 3. Some of the results of AVW and PXS algorithms are shown separately output
Stage 2 ob-
repre-
tained
sents at
the stage 3.
extracted
in Figures 5 and 6. Some
images of the
forresults
both of
the AVW and
algorithms, PXS algorithms
with the final are shown
predicted separately
output ob-
in Figures
tained 5 and3.6.Some of the results of AVW and PXS algorithms are shown separately
at stage
in Figures 5 and 6.
Electronics
Electronics 2023, 12, 12,
2023, x FOR
x FOR PEER
PEER REVIEW
REVIEW 14 14 of 26
of 26
Electronics 2023, 12, 909 14 of 25
Figure
Figure
Figure 4.Different
Different
4. 4.Different stages
stages
stages involved
involved
involvedininincolor
color
color prediction,
prediction, i.e.,
prediction, i.e., extraction
extraction
i.e., extraction ofthe
ofof
thethe pixels
pixels contributing
contributing
pixels tototo
contributing
color
color prediction
prediction andand
the the final
final output,
output, using
using AVWAVWandand
PXS PXS algorithms.
algorithms.
color prediction and the final output, using AVW and PXS algorithms.
Figure
Figure 5.Color
5. 5.Color
Figure Colorprediction
prediction
prediction results
results
results (final
(final
(final output
output
output with
thethe
with
with thecolor
of of
color
color ofthe
the thedetected
detected
detected objects)
objects)
objects) using
using
using AVWAVW
AVW
algorithm.
algorithm.
algorithm.

Electronics 2023, 12, 909 15 of 25
Figure 6. Color prediction results (final output with the color of the detected objects) using PXS
algorithm.
algorithm.
algorithm.
Figures 7 and 8 present the standard deviation of runtime in each iteration plots for
AVW Figures
and PXS7algorithms, respectively,
and 8 present when
the standard tested with
deviation the benchmark
of runtime data set plots
in each iteration with for
Figures 7 andalgorithms,
8 present the standard deviation of runtime in benchmark
each iteration plots for
respect to time. The maximum deviation in time for AVW is nearly 4.82 s, as shown inwith
AVW and PXS respectively, when tested with the data set
AVW and
respect PXS algorithms,
to time. maximum respectively,
The maximum when
deviation tested
in is
time with the
for 5.82
AVWs, is benchmark data set with
Table 3, whereas deviation for PXS nearly asnearly
shown4.82 s, as 3.
in Table shown in
respect
Table 3,towhereas
time. The maximum
maximum deviation
deviation forinPXS
timeisfor AVW
nearly is nearly
5.82 4.82 s,inasTable
s, as shown shown3. in
Table 3, whereas maximum deviation for PXS is nearly 5.82 s, as shown in Table 3.
Figure 7. Standard deviation of time taken in each iteration of AVW.


Electronics 2023, 12, 909 16 of 25
Figure 8.
Figure Standard deviation
8. Standard deviation of
of time
time taken
taken in
in each
each iteration
iteration of
of PXS.
PXS.
Table 3.
Table 3. Normalized
Normalized standard
standard deviation
deviation times
times for
for AVW
AVWand
andPXS
PXSschemes.
schemes.
Algorithm
Algorithm Minimum
Minimum Maximum
Maximum Mean
Mean
AVW
AVW 0.0084
0.0084 4.82 4.82 0.54 0.54
PXS
PXS 0.0069
0.0069 5.82 5.82 0.75 0.75
The computed accuracies for PXS and AVW for the benchmark data set are 95.4%
and 93.6%. In orderaccuracies
The computed to evaluateforthe
PXSperformance
and AVW for ofthe
algorithms
benchmark with respect
data set areto95.4%
time, and
the
average
93.6%. Inreduction in time isthe
order to evaluate compared
performanceto theofall-pixel approach.
algorithms The evaluated
with respect reductions
to time, the average
in time forin
reduction AVW
timeand PXS are approximately
is compared to the all-pixel 62% and 44%.
approach. The evaluated reductions in time
for AVW and PXS are approximately 62% and 44%.
6. Comparisons of Color Prediction Schemes
6. Comparisons of Color Prediction Schemes
As stated earlier, clustering techniques can be used for predicting the color of an ob-
ject inAs stated
the image earlier, clustering
[38,12]. techniques
Clustering can beby
is performed used for predicting
grouping the color data
all the available of anpoints
object
in the
into image [12,38].
different clusters.Clustering
The decision is performed by grouping
of which cluster all the
is allocated to aavailable
data pointdata points
depends
into different clusters. The decision of which cluster is allocated to a data
on the distance between the cluster’s centroid and the data point. Clustering methods can point depends
on the distance between the cluster’s centroid and the data point. Clustering methods
be applied in the context of predicting the dominant color of the image. The RGB channels
can be applied in the context of predicting the dominant color of the image. The RGB
are separately grouped into various clusters; the centroids of the clusters yield the domi-
channels are separately grouped into various clusters; the centroids of the clusters yield
nant colors in the image. To compare the prediction accuracy and the runtime of PXS and
the dominant colors in the image. To compare the prediction accuracy and the runtime of
AVW for the benchmark data set, conventional clustering algorithms are tested with the
PXS and AVW for the benchmark data set, conventional clustering algorithms are tested
same benchmark data set, over three iterations. The device used for testing the clustering
with the same benchmark data set, over three iterations. The device used for testing the
algorithms is the same as the one used for testing AVW and PXS, i.e., MacOS with 1.4 GHz
clustering algorithms is the same as the one used for testing AVW and PXS, i.e., MacOS
Quad-Core Intel Core i5 CPU and Intel Iris Plus Graphics 645 1536 MB GPU. The mean of
with 1.4 GHz Quad-Core Intel Core i5 CPU and Intel Iris Plus Graphics 645 1536 MB GPU.
the three iterations is considered for comparison of time with AVW and PXS. Each of these
The mean of the three iterations is considered for comparison of time with AVW and PXS.
comparisons are shown in different graphs and discussed in this section.
Each of these comparisons are shown in different graphs and discussed in this section.
6.1.
6.1. All-Pixel
All-Pixel Approach
Approach
Color
Color prediction was
prediction was initially
initially tested
tested with
with the
the benchmark
benchmark data data set using the
set using all-pixel
the all-pixel
approach,
approach, where all the pixels in the object portion are considered for the prediction. The
where all the pixels in the object portion are considered for the prediction. The
accuracy
accuracy for
for this
this approach
approach is is 94.5%.
94.5%. The
The time
time comparisons
comparisons of of the
the all-pixel
all-pixel approach
approach withwith
the
the AVW
AVW andand PXS
PXS algorithms
algorithms areare shown
shown in in Figure
Figure 9.
9. The
The average
average time
time consumed
consumed by the
by the
all-pixel approach for
all-pixel approach for predictions
predictionsof ofthe
theoutput
outputisis17.3
17.3s,s,versus
versus6.5
6.5s by
s by AVW,
AVW, and and
9.29.2
s bys
by
PXS PXS algorithms
algorithms as as shown
shown inin Table
Table 4. 4.
As Asthetheall-pixel
all-pixelapproach
approachaccounts
accountsfor for all
all the
the pixels
pixels
for
for color
color prediction,
prediction, the
the time
time drained
drained is is indelibly
indelibly higher
higher compared
compared to to either
either of
of AVW
AVW or or
PXS schemes.
PXS schemes.
Electronics 2023, 12, 909 17 of 25
Figure 9. Comparison of color prediction time of all-pixel, AVW and PXS schemes.
Figure 9. Comparison of color prediction time of all-pixel, AVW and PXS schemes.
Table 4. Comparison of normalized times for all-pixel, AVW and PXS schemes.
Table 4. Comparison of normalized times for all-pixel, AVW and PXS schemes.
Standard
Algorithm Minimum Maximum Mean Standard
Algorithm Minimum Maximum Mean Deviation
Deviation
All-Pixel 0.30 174.23 17.32 27.10
All-Pixel
AVW 0.30
0.12 174.23
59.23 17.32
6.52 27.10
10.24
AVWPXS 0.12
0.17 59.23
82.27 6.52
9.16 10.24
13.73
PXS 0.17 82.27 9.16 13.73
6.2. K-Means Clustering
6.2. K-Means Clustering
In the domain of machine learning, K-means is one of the well-known unsupervised
In the domain
clustering of machine
algorithms. learning, K-means
The algorithm, at random is one of the“k”
chooses well-known
data points,unsupervised
called means.
clustering
It then classifies the other data points into their nearest mean, by checking themeans.
algorithms. The algorithm, at random chooses “k” data points, called It
Euclidean
then classifies the other data points into their nearest mean, by checking
distance, and updates the mean value to the average of the cluster after each iteration of the Euclidean
distance,
K-means. and updates
This processtheismean value
repeated to the
until the average
centroidsofofthethecluster after
clusters each equal
become iteration
for of
two
K-means. This process
simultaneous is repeated
iterations. until the centroids
Many researchers of the clusters
have proposed become
modified equal
versions of for two
K-means
simultaneous
approach, toiterations.
improve theMany researchers
efficiency have proposed
of K-means [39–41].modified
Hence, inversions
order toofevaluate
K-means the
approach,
performance to improve the efficiency
of the proffered of K-means
algorithms, [39–41].algorithm
the K-means Hence, inwas order to evaluate
tested the
for dominant
performance of the with
color prediction proffered algorithms,data
the benchmark the K-means algorithm
set. A plot was tested
of the time for dominant
comparison between
color prediction with the benchmark data set. A plot
K-means and the AVW, PXS algorithms is displayed in Figure 10. of the time comparison between K-
means It andcanthe
beAVW, PXS
inferred algorithms
from is displayed
Figure 10, that AVW in Figureoutpaces
handily 10. K-means clustering in
predicting the dominant color of objects. K-means draws less time than PXS. The accuracy
of K-means for the benchmark data set is 84.1%, which is less compared to AVW and PXS.
Hence, both the AVW and PXS algorithms outperform K-means clustering on accurate
predictions of the dominant color of objects. The average time taken by the K-means for
predicting the output for the benchmark data set is 8.7 s, versus 6.5 s for AVW. However,
PXS takes a higher time, i.e., 0.5 s higher, compared to K-means as displayed in Table 5.
Electronics 2023, 12, 909 18 of 25
Figure 10. Time comparison plot of K-means, AVW and PXS.

Figure 10. Time comparison plot of K-means, AVW and PXS.
Table 5. Comparison of normalized times for K-means, AVW and PXS schemes.
It can be inferred from Figure 10, that AVW handily outpaces K-means clustering in
Standard
predicting the dominantMinimum
Algorithm color of objects. K-means
Maximumdraws less time
Meanthan PXS. The accuracy
Deviation
of K-means for the benchmark data set is 84.1%, which is less compared to AVW and PXS.
AVW 0.12 59.23 6.52 10.24
Hence, both
PXS the AVW and0.17 PXS algorithms82.27outperform K-means
9.16 clustering on13.73
accurate
predictions of
K-means the dominant 0.20color of objects. The
87.74 average time taken
8.64 by the K-means
12.98 for
predicting the output for the benchmark data set is 8.7 s, versus 6.5 s for AVW. However,
PXS
6.3.takes
Mini aBatch
higher time, i.e.,
K-Means 0.5 s higher, compared to K-means as displayed in Table 5.
Algorithm
The time complexity of K-means increases for larger data sets [42]. Hence, alternative
Table 5. Comparison of normalized times for K-means, AVW and PXS schemes.
clustering algorithms, such as mini batch K-means, are introduced, which mitigate the
spatial complexity of the algorithm. In mini batch K-means, small batches Standard
of data are
Algorithm Minimum Maximum Mean
processed as randomized. Mini batch K-means reduces the number of computationsDeviation of
AVW
distance 0.12in additional decrease
per iteration, ensuing 59.23 in the processing
6.52 time of 10.24
the algorithm.
PXS K-means is implemented
Mini batch 0.17 82.27
inside the 9.16 of the input
masked regions 13.73
image that
Electronics 2023, 12, x FOR PEER REVIEW
yieldK-means
the maximum RGB value 0.20 for each object. Color
87.74 prediction
8.64 is 12.9819 by
accomplished of the
26
conversion of the RGB values into HSV equivalents. Figure 11 shows the time comparison
plot
6.3. for Batch
Mini mini batch K-means
K-Means and the proposed algorithms.
Algorithm
The time complexity of K-means increases for larger data sets [42]. Hence, alternative
clustering algorithms, such as mini batch K-means, are introduced, which mitigate the
spatial complexity of the algorithm. In mini batch K-means, small batches of data are pro-
cessed as randomized. Mini batch K-means reduces the number of computations of dis-
tance per iteration, ensuing in additional decrease in the processing time of the algorithm.
Mini batch K-means is implemented inside the masked regions of the input image that
yield the maximum RGB value for each object. Color prediction is accomplished by the
conversion of the RGB values into HSV equivalents. Figure 11 shows the time comparison
plot for mini batch K-means and the proposed algorithms.
Figure 11. Time comparison plot of mini batch K-means, AVW and PXS.
Figure 11. Time comparison plot of mini batch K-means, AVW and PXS.
AVW performs color prediction in less time than mini batch K-means. The time taken
for mini batch K-means is similar to that of K-means, whereas the accuracy of mini batch
Electronics 2023, 12, 909 19 of 25
AVW performs color prediction in less time than mini batch K-means. The time taken
for mini batch K-means is similar to that of K-means, whereas the accuracy of mini batch
K-means is 83.8%, much less than the accuracy of both AVW and PXS algorithms. The
average time taken by the mini batch K-means for predicting the output for the benchmark
data set is 8.4 s compared to 6.5 s for AVW; however, PXS consumes a prediction time of
9.2 s as presented in Table 6. The effect of higher average time on the overall performance
for PXS can be perceived as a trade-off for its higher accuracy.
Table 6. Comparison of normalized times for mini batch K-means, AVW and PXS schemes.
Standard
Deviation
AVW 0.12 59.23 6.52 10.24
PXS 0.17 82.27 9.16 13.73
Mini Batch K-means 0.24 86.43 8.39 12.55
6.4. Mean-Shift Clustering

The mean-shift scheme continuously iterates data points by calculating their means,
followed by relocation of the cluster’s center to the mean point. This process is repeated
until there is a convergence, when the center of the cluster with the highest number of
data points can be considered the dominant color. The algorithm was integrated with
Mask R-CNN, and evaluated on the benchmark data set against the results with AVW
and PXS. Time complexity, the major snag of the mean-shift algorithm, can be alleviated
by considering selective data points for mean calculation [43]. The time comparisons for
mean-shift and the AVW, PXS algorithms are shown in Figure 12.
Figure 12. Time comparison plot of mean shift, AVW and PXS.
Figure 12. Time comparison plot of mean shift, AVW and PXS.
The accuracy of mean shift for the benchmark data set is 85.2% which is less than that
ofThe
AVW accuracy
and PXS.of Hence,
mean shift
bothfor
AVW theand
benchmark data set the
PXS outperform is 85.2%
meanwhich is less than
shift algorithm in terms
that of AVW and PXS. Hence, both AVW and PXS outperform the mean shift algorithmtime
of time and accuracy for prediction of the dominant color of an image. The average
in taken
terms byof time and accuracy
the mean for prediction
shift for predicting of the dominant
the output color of andata
for the benchmark image. The
set is 33 av-
s—five
erage
timestime takenthan
greater by the mean
that shift for
for AVW andpredicting thetimes
easily three output for the
higher benchmark
than the time data set by
entailed
PXS as indicated in Table 7.
is 33 s—five times greater than that for AVW and easily three times higher than the time
entailed by PXS as indicated in Table 7.
Table 7. Comparison of normalized times for mean shift, AVW and PXS schemes.
Standard
Deviation
AVW 0.12 59.23 6.52 10.24
PXS 0.17 82.27 9.16 13.73
Mean shift 0.30 577.93 32.93 76.66
Electronics 2023, 12, 909 20 of 25
Table 7. Comparison of normalized times for mean shift, AVW and PXS schemes.
Standard
Deviation
AVW 0.12 59.23 6.52 10.24
PXS 0.17 82.27 9.16 13.73
Mean shift 0.30 577.93 32.93 76.66
6.5. Gaussian Mixture Model

The Gaussian mixture model is a well-defined clustering algorithm that is used to
classify a set of given data into clusters having common parameters. It uses Gaussian
curves to assign the data points into different classes that yield a mixed representation of
Gaussians. The Gaussian mixture model considers the probability by which a data point
is assigned to one or more Gaussian curves. It is possible to have data points assigned
under different clusters. It works similar to that of the K-means algorithm, but nets a
better result in the quality and shape of the clusters. This algorithm is integrated21with
of 26
the
DCPCM model for comparing accuracy and time complexity with AVW and PXS, which
are summarized in Figure 13.
Figure 13. Time comparison plot of Gaussian mixture model, AVW and PXS.
Figure 13. Time comparison plot of Gaussian mixture model, AVW and PXS.
From the above plot, it can be deduced that the Gaussian mixture model imposes the
From the
highest above
burden onplot,
timeitcompared
can be deduced that the
to the AVW Gaussian
and mixtureThe
PXS schemes. model imposes
Gaussian mixture
the highest burden on time compared to the AVW and PXS schemes. The Gaussian in
model, when tested with the benchmark data set, delivered an accuracy of 88% mix-
color
ture model, when
prediction. The tested
averagewith
timethedrained
benchmarkby thedata set, delivered
Gaussian mixture anmodel
accuracy of 88% in the
for predicting
output for the benchmark data set is 13.4 s, as demonstrated in Table 8, notably
color prediction. The average time drained by the Gaussian mixture model for predict- higher than
the time consumed by the AVW and PXS algorithms.
ing the output for the benchmark data set is 13.4 s, as demonstrated in Table 8, notably
higher than the time consumed by the AVW and PXS algorithms.
Table 8. Comparison of normalized times for Gaussian mixture model, AVW and PXS schemes.
Table 8. Comparison of normalized times for Gaussian mixture model, AVW and PXS Standard
schemes.
Deviation
Standard
Algorithm
AVW 0.12Minimum 59.23
Maximum 6.52 Mean 10.24
Deviation
PXS 0.17 82.27 9.168 13.73
AVWmodel
Gaussian mixture 0.29 0.12 132.14 59.23 6.52
13.40 10.24
21.52
PXS 0.17 82.27 9.168 13.73
6.6. Fuzzy C-Means model
Gaussian mixture 0.29 132.14 13.40 21.52
Fuzzy C-means is an unsupervised clustering algorithm that analyzes various types
6.6.
ofFuzzy C-Means
data and groups them according to their similarity [44]. It assigns membership value
Fuzzy C-means is an unsupervised clustering algorithm that analyzes various types
of data and groups them according to their similarity [44]. It assigns membership value to
each data point per cluster center, based on the Euclidean distance between the data point
and center of the cluster. Rather than forcing a data point into one cluster, it assigns a data
Electronics 2023, 12, 909 21 of 25
to each data point per cluster center, based on the Euclidean distance between the data
point and center of the cluster. Rather than forcing a data point into one cluster, it assigns
a data point to one or more clusters based on its membership value. Indicating partial
membership and fuzzy partitioning, its value can range between 0 and 1 for different cluster
centers. New cluster centers and membership values of each data point are computed 22 of 26for
each iteration until convergence is encountered. The time comparisons for fuzzy C-means
integrated with the DCPCM model and the proposed algorithms are shown in Figure 14.
Figure 14. Time comparison plot of fuzzy C-means, AVW and PXS.
Figure 14. Time comparison plot of fuzzy C-means, AVW and PXS.
According to the above plot, fuzzy C-means require more time for color prediction
According
than AVW andtoPXS. the The
above plot, fuzzy
accuracy C-means
of color predictionrequire
with more
fuzzy time
C-meansfor color prediction
for the benchmark
than
dataAVW
set isand PXS.AVW
85.9%. The accuracy
and PXSof color prediction
outperformed with
fuzzy fuzzy C-means
C-means for the
with respect bench-
to time and
mark data set
accuracy. is 85.9%.
The average AVW
timesandforPXS outperformed
prediction fuzzy C-means
of the output withC-means
by the fuzzy respect to timethe
with
and accuracy. data
benchmark The average times
set is 22.4 for prediction
s, compared to 6.5ofs the
for output by the
AVW, and 9.2 fuzzy C-means
s for PXS with in
as shown
theTable
benchmark
9. data set is 22.4 s, compared to 6.5 s for AVW, and 9.2 s for PXS as shown
in Table 9.
Table 9. Comparison of normalized times for fuzzy C-means, AVW and PXS schemes.
Table 9. Comparison of normalized times for fuzzy C-means, AVW and PXS schemes.
Standard
Standard
Deviation
AVW 0.12 59.23 6.52 Deviation
10.24
AVW
PXS 0.17 0.12 82.27 59.23 6.52
9.16 10.24
13.73
Fuzzy C-means 1.36 273.48 22.41 38.85
PXS 0.17 82.27 9.16 13.73
Fuzzy C-means 1.36 273.48 22.41 38.85
Table 10 represents the algorithm-specific prediction accuracy of an object’s color and
Table
time 10 represents
reduction comparedthe algorithm-specific prediction
to the all-pixel approach accuracy
for the of an algorithm.
respective object’s color and
Discern-
time reduction compared to the all-pixel approach for the respective algorithm. Discern-
ment of the results in Table 10 reveals the following:
ment
• ofAVW
the results
and PXSin Table 10 reveals
algorithms havethe following:
higher accuracies compared to all other appraised
• AVW and PXS algorithms have higher accuracies compared to all other appraised
algorithms;
• algorithms;
PXS algorithm is the most accurate among all the other clustering algorithms;
• • PXS algorithm
AVW is the
algorithm hasmost accurate
the highest among in
reduction alltime
the along
other with
clustering
decentalgorithms;
prediction accuracy;
• • AVW algorithm has the highest reduction in time along with decent prediction
Negative values of “reduction in time” for mean shift and fuzzy C-means accu-
algorithms
racy;
imply that they require much longer than the all-pixel approach for the color predic-
• Negative values of “reduction in time” for mean shift and fuzzy C-means algorithms
tion task.
imply that they require much longer than the all-pixel approach for the color predic-
tion task.
Electronics 2023, 12, 909 22 of 25
Table 10. Comparison of time reduction and accuracies of all the discussed algorithms.
Reduction in Time
Algorithm Color Prediction Accuracy (%)
Compared to All-Pixel (%)
AVW 93.6 62
PXS 95.4 44.3
K-means 84.1 45.1
Mini batch K-means 83.8 47.5
Mean shift 85.2 −70.7
Fuzzy C-means 85.9 −53.4
Gaussian mixture model 88 22.4
7. Conclusions
Sustained proliferation in the deployment of autonomous robotic devices has stimu-
lated enhanced urgency in their capability of detection and discernment of objects in an
image. This paper has elucidated the design and working of two innovative color predic-
tion algorithms, PXS and AVW, for the extraction of pixels in a faster and more efficient
manner. Accuracy and reliability of the proposed algorithms are appraised by comparison
of the proffered algorithms with conventional approaches—K-means, Gaussian mixture
model, fuzzy C-means, mini batch K-means, and mean-shift clustering algorithms—using
a benchmark data set. The propounded algorithms performed with greater accuracy in an
exceptionally short time span, when AVW and PXS were juxtaposed with popular extant
color prediction algorithms. AVW and PXS algorithms are distinct from each other. PXS
algorithm exhibited enhanced accuracy (95.4%) with a downside of longer prediction time.
The AVW algorithm perceptibly needed less time for prediction of an object’s dominant
color (up to 62% decrease versus the all-pixel approach), incurring a compromised accu-
racy of up to 2%. A notable inference of this study calls for application-specific trade-offs
between latency and prediction accuracy.
For color prediction, either AVW or PXS can be deployed; if time complexity is a
major concern for predicting color, the AVW algorithm can be considered, which consumes
an optimized time returning decent accuracy. If one prefers higher accuracy, PXS can
be chosen for the color prediction, with a little higher prediction time compared to the
existing techniques. Integration of the proposed algorithms with service robots can have
a significant impact in detection of objects based on color in a real-time scenario. Time
consumed to predict the color by these proposed algorithms is notably less than that of
clustering algorithms. These algorithms are also useful for detecting traffic signals for
autonomous vehicles and for detection of damaged foods. Clustering algorithms are ill
suited for real time scenarios, due to their twin problems of less accuracy and longer
prediction times. The proposed DCPCM model can be deficient dealing with equally
shaded multicolored objects as well as certain real-time applications. These apparent
deficiencies call for additional research and development of the DCPCM concept to sustain
higher precision and streamlined performance of AVW and PXS algorithms.
Author Contributions: R.K.M. was responsible for conceptualization; i.e., ideas, formulation or
evolution of overarching research goals and aims. He was also in charge of supervision, i.e., oversight
and leadership responsibility for the research activity planning and execution, including mentorship
external to the core team, not only the development or design of methodology and creation of models,
i.e., the methodology, but also provision of study materials, reagents, materials, patients, laboratory
samples, animals, instrumentation, computing resources, or other analysis tools, i.e., the resources, in
addition to handling the preparation, creation and/or presentation of the published work, specifically,
writing the initial draft (including substantive translation), i.e., writing the original draft. He had
management and coordination responsibility for research activity planning and execution, i.e., project
administration and preparation, creation, and/or presentation of the published work by those from
the original research group, specifically critical review, commentary, or revision, including pre- and
postpublication stages, i.e., writing—review and editing. B.T. conducted management activities to
annotate (produce metadata), scrub data, and maintain research data (including software code, where
Electronics 2023, 12, 909 23 of 25
necessary for interpreting the data itself) for initial use and later reuse, i.e., data curation, applied
statistical, mathematical, computational, and other formal techniques to analyze and synthesize
study data, i.e., formal analysis, and the development and design of the methodology and creation
of models, i.e., methodology. G.S.S. was performing management activities to annotate (produce
metadata), scrub data and maintain research data (including software code, where necessary for
interpreting the data itself) for initial use and later reuse i.e., data curation. Also, applied statistical,
mathematical, computational, and other formal techniques to analyze and synthesize study data
i.e., formal analysis, and the development of design of methodology and creation of models, i.e.,
methodology. G.M.R. carried out the programming, software development, designing computer
programs, implementation of the computer code and supporting algorithms, and testing existing
code components, i.e., software, and performed verification, whether as a part of the activity or
separate, of the overall replication/reproducibility of results/experiments and other research outputs
i.e., validation. I.R.S.K. did the programming, software development, designing computer programs,
implementation of computer code, supporting algorithms, and testing existing code components,
i.e., software, and performed verification, whether as a part of the activity or separate, of the overall
replication/reproducibility of results/experiments and other research outputs i.e., validation. S.S.P.
executed the programming, software development, design of computer programs, implementation
of the computer code and supporting algorithms, and testing of existing code components, i.e.,
software, and carried out the verification, whether as a part of the activity or separate, of the overall
replication/reproducibility of results/experiments and other research outputs, i.e., validation. All
authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The datasets generated during and/or analyzed during the current
study are available from the corresponding author on reasonable request.
Acknowledgments: Members of this research team are grateful to the Department of Electronics and
Communication Engineering and the Humanitarian Technology Labs (HuT Labs) at the Amritapuri
campus of Amrita Vishwa Vidyapeetham, Kollam for providing all the necessary lab facilities and a
highly encouraging work environment, which were key factors in completion of this research project.
Conflicts of Interest: The authors have no competing interests to declare that are relevant to the
content of this article. The authors have no relevant financial or nonfinancial interests to disclose.
References
1. Kim, J.H.; Kim, B.G.; Roy, P.P.; Jeong, D.M. Efficient Facial Expression Recognition Algorithm Based on Hierarchical Deep Neural
Network Structure. IEEE Access 2019, 7, 41273–41285. [CrossRef]
2. Jeong, D.; Kim, B.-G.; Dong, S.-Y. Deep Joint Spatiotemporal Network (DJSTN) for Efficient Facial Expression Recognition. Sensors
2020, 20, 1936. [CrossRef] [PubMed]
3. Kim, J.-H.; Hong, G.-S.; Kim, B.-G.; Dogra, D.P. deepGesture: Deep learning-based gesture recognition scheme using motion
sensors. Displays 2018, 55, 38–45. [CrossRef]
4. Manju, K.; Aditya, G.; Ruben, G.C.; Verdú, E. Gesture Recognition of RGB and RGB-D Static Images using Convolutional Neural
Networks. Int. J. Interact. Multimed. Artif. Intell. 2019, 5, 22–27. [CrossRef]
5. Qummar, S.; Khan, F.G.; Shah, S.; Khan, A.; Shamshirband, S.; Rehman, Z.U.; Khan, I.A.; Jadoon, W. A Deep Learning Ensemble
Approach for Diabetic Retinopathy Detection. IEEE Access 2019, 7, 150530–150539. [CrossRef]
6. Shamshirband, S.; Mahdis, F.; Dehzangi, A.; Chronopoulos, A.T.; Alinejad-Rokny, H. A review on deep learning approaches in
healthcare systems: Taxonomies, challenges, and open issues. J. Biomed. Inform. 2021, 113, 103627. [CrossRef]
7. Pillai, M.S.; Chaudhary, G.; Khari, M.; Crespo, R.G. Real-time image enhancement for an automatic automobile accident detection
through CCTV using deep learning. Soft Comput. 2021, 25, 11929–11940. [CrossRef]
8. Pathak, A.R.; Pandey, M.; Rautaray, S. Application of Deep Learning for Object Detection. Procedia Comput. Sci. 2018, 132,
1706–1717. ISSN-1877-0509. [CrossRef]
9. Liu, L.; Ouyang, W.; Wang, X.; Fieguth, P.; Chen, J.; Liu, X.; Pietikäinen, M. Deep Learning for Generic Object Detection: A Survey.
Int. J. Comput. Vis. 2020, 128, 261–318. [CrossRef]
10. Srisuk, S.; Suwannapong, C.; Kitisriworapan, S.; Kaewsong, A.; Ongkittikul, S. Performance Evaluation of Real-Time Object
Detection Algorithms. In Proceedings of the 2019 7th International Electrical Engineering Congress (IEECON), Hua Hin, Thailand,
6–8 March 2019; pp. 1–4. [CrossRef]
Electronics 2023, 12, 909 24 of 25
11. Kim, J.-A.; Sung, J.-Y.; Park, S.-H. Comparison of Faster-RCNN, YOLO, and SSD for Real-Time Vehicle Type Recognition. In
Proceedings of the 2020 IEEE International Conference on Consumer Electronics—Asia (ICCE-Asia), Seoul, Republic of Korea,
1–3 November 2020; pp. 1–4. [CrossRef]
12. Liu, Z.-Y.; Ding, F.; Xu, Y.; Han, X. Background dominant colors extraction method based on color image quick fuzzy c-means
clustering algorithm. Def. Technol. 2020, 17, 1782–1790. [CrossRef]
13. Elavarasi, S.A.; Jayanthi, J.; Basker, N. Trajectory Object Detection using Deep Learning Algorithms. Int. J. Recent Technol. Eng.
2019, 8, C6564098319. [CrossRef]
14. Canny, J. A Computational Approach to Edge Detection. IEEE Trans. Pattern Anal. Mach. Intell. 1986, 8, 679–698. [CrossRef]
15. Kaggle. Ideas for Image Features and Image Quality. Available online: https://www.kaggle.com/code/shivamb/ideas-for-
image-features-and-image-quality (accessed on 1 January 2022).
16. Vadivel, A.; Sural, S.; Majumdar, A.K. Human color perception in the HSV space and its application in histogram generation for
image retrieval. In Color Imaging X: Processing, Hardcopy, and Applications; SPIE: Bellingham, WA, USA, 2005; Volume 5667.
17. Ray, S.A. Color gamut transform pairs. ACM Sig-Graph Comput. Graph. 1978, 12, 12–19.
18. Atram, P.; Chawan, P. Finding Dominant Color in the Artistic Painting using Data Mining Technique. Int. Res. J. Eng. Technol.
2020, 6, 235–237.
19. Data Engineering and Communication Technology; Raju, K.S., Senkerik, R., Lanka, S.P., Rajagopal, V., Eds.; Advances in Intelligent
Systems and Computing; Springer: Berlin/Heidelberg, Germany, 2020; Volume 1079. [CrossRef]
20. Guyeux, C.; Chrétien, S.; BouTayeh, G.; Demerjian, J.; Bahi, J. Introducing and Comparing Recent Clustering Methods for Massive
Data Management on the Internet of Things. J. Sens. Actuator Netw. 2019, 8, 56. [CrossRef]
21. Sai Satyanarayana Reddy, S.; Kumar, A. Edge Detection and Enhancement of Color Images Based on Bilateral Filtering Method
Using K-Means Clustering Algorithm. In ICT Systems and Sustainability; Advances in Intelligent Systems and, Computing; Tuba,
M., Akashe, S., Joshi, A., Eds.; Springer: Singapore, 2020; Volume 1077. [CrossRef]
22. Peng, K.; Leung, V.C.M.; Huang, Q. Clustering Approach Based on Mini Batch K-means for Intrusion Detection System Over Big
Data. IEEE Access 2018, 6, 11897–11906. [CrossRef]
23. Liu, N.; Zheng, X. Color recognition of clothes based on k-means and mean shift Automatic Detection and High-End Equipment.
In Proceedings of the 2012 IEEE International Conference on Intelligent Control, Beijing, China, 27–29 July 2012; pp. 49–53.
[CrossRef]
24. Mohit, N.A.; Sharma, M.; Kumari, C. A novel approach to text clustering using shift k-medoid. Int. J. Soc. Comput. Cyber-Phys.
Syst. Int. J. Soc. Comput. Cyber-Phys. Syst 2019, 2, 106–118. [CrossRef]
25. Balasubramaniam, P.; Ananthi, V.P. Segmentation of nutrient deficiency in incomplete crop images using intuitionistic fuzzy
C-means clustering algorithm. Nonlinear Dyn. 2016, 83, 849–866. [CrossRef]
26. Yin, S.; Zhang, Y.; Karim, S. Large Scale Remote Sensing Image Segmentation Based on Fuzzy Region Competition and Gaussian
Mixture Model. IEEE Access 2018, 6, 26069–26080. [CrossRef]
27. Liu, Y.; Xie, Z.; Liu, H. An Adaptive and Robust Edge Detection Method Based on Edge Proportion Statistics. IEEE Trans. Image
Process. 2020, 29, 5206–5215. [CrossRef]
28. Latha, N.S.A.; Megalingam, R.K. Exemplar-based Learning for Recognition & Annotation of Human Actions. In Proceedings of
the 2020 9th International Conference System Modeling and Advancement in Research Trends (SMART), Moradabad, India, 4–5
December 2020; pp. 91–93. [CrossRef]
29. Wang, Y.; Luo, J.; Wang, Q.; Zhai, R.; Peng, H.; Wu, L.; Zong, Y. Automatic Color Detection of Grape Based on Vision Computing
Method. In Recent Developments in Intelligent Systems and Interactive Applications IISA 2016; Advances in Intelligent Systems and,
Computing; Xhafa, F., Patnaik, S., Yu, Z., Eds.; Springer: Cham, Switzerland, 2017; Volume 541. [CrossRef]
30. Zhou, S.; Wang, J.; Wang, L.; Zhang, J.; Wang, F.; Huang, D.; Zheng, N. Hierarchical and Interactive Refinement Network for
Edge-Preserving Salient Object Detection. IEEE Trans. Image Process. 2021, 30, 1–14. [CrossRef] [PubMed]
31. Wang, L.; Tang, D.; Guo, Y.; Do, M.N. Common Visual Pattern Discovery via Nonlinear Mean Shift Clustering. IEEE Trans. Image
Process. 2015, 24, 5442–5454. [CrossRef] [PubMed]
32. Liu, X.; Zhao, D.; Jia, W.; Ji, W.; Ruan, C.; Sun, Y. Cucumber Fruits Detection in Greenhouses Based on Instance Segmentation.
IEEE Access 2019, 7, 139635–139642. [CrossRef]
33. Megalingam, R.K.; Sree, G.S.; Reddy, G.M.; Krishna, I.R.S.; Suriya, L.U. Food Spoilage Detection Using Convolutional Neural
Networks and K Means Clustering. In Proceedings of the 2019 3rd International Conference on Recent Developments in Control,
Automation & Power Engineering (RDCAPE), Noida, India, 10–11 October 2019; pp. 488–493. [CrossRef]
34. Megalingam, R.K.; Karath, M.; Prajitha, P.; Pocklassery, G. Computational Analysis between Software and Hardware Implementa-
tion of Sobel Edge Detection Algorithm. In Proceedings of the 2019 International Conference on Communication and Signal
Processing (ICCSP), Chennai, India, 4–6 April 2019; pp. 529–533. [CrossRef]
35. Megalingam, R.K.; Manoharan, S.; Reddy, R.; Sriteja, G.; Kashyap, A. Color and Contour Based Identification of Stem of Coconut
Bunch. IOP Conf. Ser. Mater. Sci. Eng. 2017, 225, 012205. [CrossRef]
36. Alexander, A.; Dharmana, M.M. Object detection algorithm for segregating similar colored objects and database formation. In
Proceedings of the 2017 International Conference on Circuit, Power and Computing Technologies (ICCPCT), Kollam, India, 20–21
April 2017; pp. 1–5. [CrossRef]
Electronics 2023, 12, 909 25 of 25
37. Krishna Kumar, P.; Parameswaran, L. A hybrid method for object identification and event detection in video. In Proceedings
of the 2013 4th National Conference on Computer Vision, Pattern Recognition, Image Processing and Graphics (NCVPRIPG),
Jodhpur, India„ 20–21 April 2013; pp. 1–4. [CrossRef]
38. Molada-Tebar, A.; Marqués-Mateu, Á.; Lerma, J.L.; Westland, S. Dominant Color Extraction with K-Means for Camera Characteri-
zation in Cultural Heritage Documentation. Remote Sens. 2020, 12, 520. [CrossRef]
39. Khandare, A.; Alvi, A.S. Efficient Clustering Algorithm with Improved Clusters Quality. IOSR J. Comput. Eng. 2016, 48, 15–19.
[CrossRef]
40. Wu, S.; Chen, H.; Zhao, Z.; Long, H.; Song, C. An Improved Remote Sensing Image Classification Based on K-Means Using HSV
Color Feature. In Proceedings of the 2014 Tenth International Conference on Computational Intelligence and Security, Kunming,
China, 15–16 November 2014; pp. 201–204. [CrossRef]
41. Haraty, R.A.; Dimishkieh, M.; Masud, M. An Enhanced k-Means Clustering Algorithm for Pattern Discovery in Healthcare Data.
Int. J. Distrib. Sens. Netw. 2015, 11, 615740. [CrossRef]
42. Bejar, J. K-Means vs. Mini Batch K-Means: A Comparison; LSI-13-8-R. 2013. Available online: http://hdl.handle.net/2117/23414
(accessed on 1 January 2020).
43. Cheng, Y. Mean shift, mode seeking, and clustering. IEEE Trans. Pattern Anal. Mach. Intell. 1995, 17, 790–799. [CrossRef]
44. Hung, M.-C.; Yang, D.-L. An efficient Fuzzy C-Means clustering algorithm. In Proceedings of the 2001 IEEE International
Conference on Data Mining, San Jose, CA, USA, 29 November–2 December 2001; pp. 225–232. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

Optimized and Efficient Color Prediction Algorithm

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimized and Efficient Color Prediction Algorithm

Uploaded by

Copyright:

Available Formats

electronics

Department of Electronics and Communication Engineering, Amrita Vishwa Vidyapeetham,

Electronics 2023, 12, 909. https://doi.org/10.3390/electronics12040909 https://www.mdpi.com/journal/electronics

on the results and concludes the work.

3. Dominant Color Prediction Color Map (DCPCM) Model

S1 is the name of the range in which “S” belongs.

V 1 is the name of the range in which “V” belongs.

Table 1. Combinations of “S” and “V” parameters.

Saturation (S1 ) Value (V 1 ) Color

Predictions for the combination S = “medium” and V = “medium”.

if (H (◦ )e[0, 10] or H (◦ )e[331, 360])

Predictions for the combination S = “high” and V = “high”.

4. AVW and PXS Algorithms

Table 2. Symbols used for modeling AVW and PXS.

Saturation (S1 ) Value (V 1 )

4.1. Preprocessing Required for AVW and PXS Algorithms

4.2. All-Pixel Approach

4.3. Average Windowing (AVW) Algorithm

4.3.1. Predicting Window Size (“f ”)

4.3.2. Pseudo Code of AVW

4.3.3. Corner Cases for AVW

4.4. Pixel Skip (PXS) Algorithm

4.4.1. Predicting Skip Size (S)

4.4.2. Pseudo Code for PXS

4.4.3. Corner Cases

Electronics 2023, 12, 909 13 of 25

(a) Processed image. (b) Extracted image.

(a)Processed image. (b)Extracted image.

Electronics 2023, 12, x FOR PEER REVIEW 15 of 26

Figure 7. Standard deviation of time taken in each iteration of AVW.

Figure 7. Standard deviation of time taken in each iteration of AVW.

Figure 10. Time comparison plot of K-means, AVW and PXS.

6.4. Mean-Shift Clustering

6.5. Gaussian Mixture Model

You might also like