(IJCSIS) International Journal of Computer Science and Information Security,Vol.
9
, No.
7
, 201
1
Performance Comparison of Image Classifier Using
DCT, Walsh, Haar and Kekre’s Transform
Dr. H. B. Kekre
Senior Professor,
Computer Engineering, MP’STME,SVKM’S NMIMS University,
Mumbai, Indiahbkekre@yahoo.com
Tanuja K. SarodeAsst. Professor,
Meena S. UgaleAsst. Professor
Abstract
—
In recent years, thousands of images are generatedeveryday, which implies the necessity to classify, organize andaccess them by easy and faster way. The need for imageclassification is becoming increasingly important.The paper presents innovative Image Classification techniquebased on feature vectors as fractional coefficients of transformed
images using Discrete Cosine, Walsh, Haar and Kekre’s
transforms. The energy compaction of transforms in highercoefficients is taken to reduce the feature vector size per imageby taking fractional coefficients of transformed image. Thevarious sizes of feature vectors are generated such as 8X8, 16X16,32X32, 64X64 and 128X128.The proposed technique is worked over database of 1000 imagesspread over 10 different classes. The Euclidean distance is usedas similarity measure. A threshold value is set to determine towhich category the query image belongs to.
Keywords
—
Discrete Cosine Transform (DCT), Walsh
Transform, Haar Transform, Kekre’s Transform,
ImageDatabase, Transform Domain, Feature Vector
I.
I
NTRODUCTION
In recent years, many application domains such as biomedical,military, education and web store a big number of images indigital libraries.The need to manage these images and locate target images inresponse to user queries has become a significant problem[26]. Image classification is an important task for manyaspects of global change studies and environmentalapplications.In recent years, the accelerated growth of digital mediacollections and in particular still image collections, bothproprietary and on the Web, has established the need for thedevelopment of human-centered tools for the efficient accessand retrieval of visual information. As the amount of information available in the form of still images continuouslyincreases, the necessity of efficient methods for the retrievalof the visual information becomes evident [30].Image categorization is an important step for efficientlyhandling large image databases and enables theimplementation of efficient retrieval algorithms. Imageclassification aims to find a description that best describe theimages in one class and distinguish these images from all theother classes. It can help users to organize and to browseimages. Although this is usually not a very difficult task forhumans, it has been proved to be an extremely difficultproblem for computer programs.Classification of images involves identifying anareaof knowncovertype and instructing the computer to find all similarareas in the studyregion.The similarities are based onreflectance values in the input images.Digital image processing is a collection of techniques for themanipulation of digital images by computers. Classificationgenerally comprises four steps [27]:1.
Pre-processing: E.g. atmospheric correction, noisesuppression, and finding the band ratio, principalcomponent analysis, etc.2.
Training: Selection of the particular feature which bestdescribes the pattern.3.
Decision: Choice of suitable method for comparing theimage patterns with the target patterns.4.
Assessing the accuracy of the classification.Image classification refers to the labeling of images into oneof predefined semantic categories.Using an image Classification, images can be analysed andindexed automatically by automatic description whichdepends on their objective visual content. The most importantstep in an Image Classification system is the image description.Indeed, features extraction gives a feature vector per imagewhich is a reduced representation of the image visual content,because images are too big to be used directly for indexingand retrieval [30].In this paper the use of Discrete Cosine Transform (DCT),
Walsh Transform, Haar Transform and Kekre’s Transform is
26http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol.
9
, No.
7
, 201
1
investigated for image classification technique A featurevector is extracted for an image of size N X N using DCT or
Walsh or Haar or Kekre’s Transform. The similarity
measurement (SM), where a distance (e.g., Euclidean distance)between the query image and each image in the database usingtheir feature vectors is computed so
that the top closest
images can be retrieved [7, 14, 17].II.
RELATED
WORKMany image classification systems have been developed sincethe early 1990s. Various image representations andclassification techniques are adopted in these systems: theimages are represented by global features, block-basedfeatures, region-based local features, or bag-of-wordsfeatures[8], and various machine learning techniques areadopted for the classification tasks, such as K-nearestneighbor (KNN)[24], Support Vector Machines (SVM)[24],Hidden Markov Model(HMM)[21], Diverse Density(DD)[29],DD-SVM[28] and so on.Recently, a popular technique for representing image contentfor image category recognition is the bag of visual wordmodel [10, 6].In the indexing phase, each image of the database isrepresented using a set of image attribute, such as color [25],shape [9, 1], texture [2] and layout [26]. Extracted features arestored in a visual feature database. In the searching phase,when a user makes a query, a feature vector for the query iscomputed. Using a similarity criterion, this vector is comparedto the vectors in the feature database.A heterogeneous image recognition system based on contentdescription and classification is used in which for imagedatabase several features extraction methods are used andapplied to better describes the images content. The featuresrelevance is tested and improved through Support VectorsMachines (SVMs) classifier of the consequent images indexdatabase [26].In literature there are various Image classification methods.Some of these methods use wavelets transform and supportvector machine [33]; some methods use effective algorithmfor building codebooks for visual recognition [14]; someadvanced image classification techniques use Artificial NeuralNetworks, Support Vector Machines, Fuzzy measures andGenetic Algorithms [23] whereas some methods are proposedfor classifying images, which integrates several sets of Support Vector Machines (SVM) on multiple low level imagefeatures [32].III.
DISCRETE
COSINE
TRANSFORM
(DCT)In general, neighbouring pixels within an image tend to behighly correlated. As such, it is desired to use an invertibletransform to concentrate randomness into fewer, decorrelatedparameters [13].The Discrete Cosine Transform (DCT) hasbeen shown to be near optimal for a large class of images inenergy concentration and decorrelating. It has been adopted inthe JPEG and MPEG coding standards [12][3]. The DCTdecomposes the signal into underlying spatial frequencies,which then allow further processing techniques to reduce theprecision of the DCT coefficients consistent with the HumanVisual System (HVS) model. The DCT coefficients of animage tend themselves as a new feature, which have theability to represent the regularity, complexity and sometexture features of an image and it can be directly applied toimage data in the compressed domain [31]. This may be a wayto solve the large storage space problem and thecomputational complexity of the existing methods.The two dimensional DCT can be written in terms of pixelvalues
f(i, j)
for
i,j
= 0,1,…,
N
-1 and the frequency-domaintransform coefficients
F(u,v
):The DCT tends to concentrate information, making it usefulfor image compression applications and also helping inminimizing feature vector size in CBIR [23]. For full 2-Dimensional DCT for an NxN image the number of multiplications required are N
2
(2N) and number of additionsrequired are N
2
(2N-2).
IV.
WALSH
TRANSFORMWalsh transform matrix [18,19,23,26] is defined as a set of Nrows, denoted Wj, for j = 0, 1, .... , N - 1, which have thefollowing properties:Wj takes on the values +1 and -1.Wj[0] = 1 for all j.Wj xW
KT
=0, for j
≠
k and Wj xW
KT
=N, for j=k.Wj has exactly j zero crossings, for j = 0, 1, ...., N-1.Each row Wj is even or odd with respect to itsmidpoint.Walsh transform matrix is defined using a Hadamard matrixof order N. The Walsh transform matrix row is the row of theHadamard matrix specified by the Walsh code index, whichmust be an integer in the range [0... N -1]. For the Walsh codeindex equal to an integer j, the respective Hadamard outputcode has exactly j zero crossings, for j = 0, 1... N - 1.
27http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol.
9
, No.
7
, 201
1
For the full 2-Dimensional Walsh transform applied to imageof size NxN, the number of additions required are 2N
2
(N-1)and absolutely no multiplications are needed in Walshtransform [18].
V.
HAAR
TRANSFORMThis sequence was proposed in 1909 by Alfred Haar. Haarused these functions to give an example of a countableorthonormal system for the space of square-integral functionson the real line. The Haar wavelet is also the simplest possiblewavelet. The technical disadvantage of the Haar wavelet isthat it is not continuous, and therefore not differentiable.The Haar wavelet's mother wavelet function (t) can bedescribed as:(3)And its scaling function can be described as,(4)VI.
KEKRE’S
TRANSFORMKekr
e’s transform matrix can be of any size NxN, which need
not have to be in powers of 2 (as is the case with most of othertransforms). All upper diagonal and diagonal values of
Kekre’s transform matrix are one, while the lower diagonal
part except the values just below diagonal is zero [23].
Generalized NxN Kekre’s transform matrix can be given as:
For taking Kekre’s transform of an NxN image, the number of
required multiplications are 2N(N-2) and number of additionsrequired are N(N
2
+N-2).
TABLE ICOMPUTATIONAL COMPLEXITY FOR APPLYING TRANSFORMS TOIMAGE OF SIZE NXN [18]DCT Walsh Haar
Kekre’s
TransformNumber of Additions2N
2
(N-1) 2N
2
(N-1) 2N
2
log
2
(N) N[N(N+1)-2]Number of MultiplicationsN
2
(2N) 0 0 2N(N-2)TotalAdditions fortransform of 128 x128image37715968 4161536 229376 2113280
[Here one multiplication is considered as eight additions forlast row computations]VII.
PROPOSED
ALGORITHMThe proposed algorithm makes use of well known Discrete
Cosine Transform (DCT), Walsh, Haar and Kekre’s
Transform to generate the feature vectors for the purpose of search and retrieval of database images.We convert an RGB image into gray level image. For spatiallocalization, we then use the DCT or Walsh or Haar or
Kekre’s transformation. Each image is resized to
N*N
size.
DCT or Walsh or Haar or Kekre’s Transform is applied on the
image to generate a feature vector as shown in figure 1.
A.
Algorithm for Image Classification
1.
Feature vector of the query image is generated as shownin figure 1.2.
Feature vector of the query image is compared with thefeature vectors of all the images in the database.Euclidean distance measure is used to check the closenessof the query image and the database images.3.
Euclidean distance values are sorted w.r.t. ascendingorder sequence to find first 50 closest matches with queryimage.4.
The closest matches with query image for all 10categories are calculated.5.
A threshold value is set to determine to which categorythe query image belongs to.
6.
Display the category of the query image
.
28http://sites.google.com/site/ijcsis/ISSN 1947-5500
Search History:
Result 00 of 00
00 results for result for
p.
Performance Comparison of Image Classifier Using DCT, Walsh, Haar and Kekre’s Transform
In recent years, thousands of images are generated everyday, which implies the necessity to classify, organize and access them by easy and faster way. The need for image classification is becoming increasingly important. The paper presents…
(More)
585
Reads
0
Readcasts
0
Embed Views
Recommended
More From This User
Notes
Load more
