Professional Documents
Culture Documents
Chapter 4
Chapter 4
Segmentation means partitioning the image into various parts. The image is divided
into segments because there are some regions in the image that are not needed or not
contain any information. So, by doing segmentation we can use the essential segments
for processing the image. An image is a collection of various pixels. The pixels which
are having the same intensity are grouped using segmentation [130].
Here, we have introduced a novel segmentation technique called RBIS for lane
detection from the given image sample using an iterative method based on the distance
measures like Bhattacharya, and the neighborhood distance.
The Architecture of the projected method is depicted in figure 4.1. The input image
is provided and segmentation is applied to this image using the sparking method. Then,
the output of the segmentation is used to get multilane detected images by applying
the proposedRBIS segmentation method.
78
.
Figure 4.1: Proposed RBIS method block diagram for lane detection
79
4.1.1. IMAGE SEGMENTATION USING THE SPARKING PROCESS
𝑟2
𝑣1 = 𝑒𝑥𝑝 − [ ]
2𝑇𝑎2
(4.1)
Where,
𝑟is denoted as a random integer, and the size of the random number is 1 × 𝑠. The
threshold value provided by the user is denoted as 𝑇𝑎 .
1
[ 𝑠 ∑𝑠𝑣=1 max(𝑣) − 𝑣] − 𝑣
𝑘=
∑𝑠𝑣=1 max(𝑣) − 𝑣
(4.2)
Now, for the given input 𝐽 the sparking image is obtained. Rotate the input image e
𝐽 atdiverse angles to create 𝑠 images of the identical size. Therefore, the sparking
image gained is expressed below,
𝐿𝑒 + 𝐿𝑚𝑖𝑛 (𝑝, 𝑞)
𝐿𝑛𝑒𝑤 (𝑝, 𝑞) = 255
𝐿𝑚𝑎𝑥 (𝑝, 𝑞)
(4.3)
Where,
80
𝐿𝑚𝑖𝑛 (𝑝, 𝑞) = min [𝑙1 (𝑝, 𝑞), 𝑙2 (𝑝, 𝑞), … . 𝑙𝑏 (𝑝, 𝑞)]
The threshold value is denoted by𝐿𝑒 , 𝑝 represents the row pixel value and 𝑞
represents the column pixel value.
Sparking imge is used to deine the value of optimal threshold. Here, the Neighbor Pixel
(NP) shows that current pixel value is referred by all roalues are indicated by all
coloumns.A count value is usd to specify the NP matrix 𝑁𝑃[𝑝, 𝑞]from the sparking
matrix. Next, pixel are having 0-255 values and these all pixel values estimated with “
X ” to gain the best threshold value. Here,𝐻 it is considered as the threshold. From the
matrix NP, the matrix W and Z are constructed as follows:
𝑊 ← [𝐻 × 𝐻] (4.4)
1
𝑆𝑊 = − 2 [𝑝(𝑊)𝑙𝑜𝑔𝑃(𝑊)] (4.6)
1
𝑆𝑍 = − 2 [𝑝(𝑍)𝑙𝑜𝑔𝑃(𝑍)] (4.7)
Where, 𝑊and 𝑍are the probabilities, and they are denoted as 𝑝(𝑊)and 𝑝(𝑍). The
selected threshold’s cost value is provided by doing the addition of the entropy, and it
is shown in equation (4.8). To perform binarization select the maximum threshold
value as the optimal threshold, so that we will get the final segmented image.
𝑋 = 𝑆𝑊 + 𝑆𝑍
(4.8)
81
(a)
(b)
Figure 4.2: (a) Original input image sample (b) sparking Image
82
Where, X represent final segmented image.
The input sample is demonstrated in figure 4.2(a) and the sparking image is
demonstrated in figure 4.2(b).
Now, we have got the segmented image after the sparkling process. This segmented
image is now exposed to the proposedRBIS segmentation approach for path
recognition. At first, the image is divided into grids or matrices. Then any three grids
are selected randomly, and these selected three grids are the targets. After selecting the
targets, it is now easy to compute the Bhattacharya distance between each grid and
targets. At that point, the Bhattacharya distance is used to specify the centroid with the
grids of minimum distance. The distance measure is used for separating the image
into two parts that are one part is road and the other is lane. Now, divide the segmented
image X into partitioned into 3 3 grid size, as following,
𝐺 = {𝐺1 , 𝐺2 , … … , 𝐺𝑗 , … … . . 𝐺𝑣 }
(4. 9)
Three grids as a target are selected from the partitioned grids, and is given as,
T = T1 , T2 , T3 (4.10)
The phases associated with the proposed segmentation strategy are talked about
beneath:
Step1: Initialization
In this step, we are randomly selecting grids k as a target among the v grids. An
iterative procedure is used,and the random targets are optimized at the final step. The
table shown in the figure 4.3 depicts that out of 9 grids, we have selected three grids as
a target.
83
Figure 4.3: Constructed matrix
(4. 11)
Where,
84
The Bhattacharya distance is denoted by 𝐵𝐷 (𝐺, 𝑇), variance of the E th distribution is
given by 𝑉2𝐸, and T th distribution variance is given by 𝑉2𝑟 . 𝑀𝐸 is the mean of E th
the distribution, and 𝑟𝐸 and 𝑟𝑇 are the centroid kurtosis .
Therefore, based on the distance calculation, the grid that belongs to a particular target
is shown in the figure4.4 with the sample values. The figure 4.4 shows that the distance
between grid 1 and target 1 is the minimum as compared to the distance calculated
between grid 1 and target 2 as well as grid 1 and target 3, therefore, grid 1 belongs to
the target 1. Similarly, when the distance is calculated with grid 4, the distance between
target 2 and grid 4 is less as compared to the distance between grid 4 and target 1 as
well as grid 4 and target 3. In the figure 4.4, grids are highlighted as belonging to their
particular targets.
85
Step 3: update the Target
In this step, the mean of all the data points assigned to the specific target is taken, and
the targets are recomputed. If in case the sample is not present in the closest grid, then
switch the sample to the corresponding grid and update the target.
Step 4: Grouping the grids
Repeat steps 2 and 3 iteratively till optimal target are determined. The minimum
distance measure is used to select the best grid. Therefore, at the end of the iteration,
the selected grids are grouped based on Bhattacharya distance.
Step 5: Use the neighbor distance measure to merge the grids into two segments
As our objective is to separate lanes from the roads, so, the 3 targets achieved are
merged into two segments. The separation is done by measuring the nearest neighbor
distance of the grids. The grids which are having the same intensities are taken into
account to fit into the same segment. Therefore, the grids which are having minimal
distance with their neighbor are merged so that the grids that are grouped as 3 are
merged into 2 segments: road segment and lane segment.
86
4.2. RESULTS FOR THE PROPOSED RBIS SEGMENTATION METHOD
The proposed region-based segmentation using iterative seed method’s obtained
results are demonstrated in this section. The performance of the projected system is
evaluated concerning the existing methods. The consequences for the various road
scenarios are depicted in figure 4.6. figure 4.6 (a) shows the Original input image
sample, figure 4.6 (b) shows the Sparking image, and figure 4.6 (c) demonstrates the
covered/overlay illustration detected by RBIS segmentation
87
(a) (b) (c)
Figure 4.6: RBIS method experimental results, (a) Originalinput image, (b)
Sparking image, and (c) Overlay/covered image identified by RBIS segmentation
4.2.1EXPERIMENTAL SETUP
4.2.2.DATABASE DESCRIPTION
Here we have used the KITTI vision benchmark dataset. This dataset we have already
discussed in chapter 3 section 3.5.2.
The metrics used for evaluating the enactment of the proposed system are
accuracy,specificity,andsensitivity. These measues are explained below.
88
detection accuracy ought to be high for the best execution. The articulation for the
detection accuracy is given underneath,
𝑇𝑟𝑢𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝑇𝑟𝑢𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
𝐷𝐴 =
𝑇𝑟𝑢𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝑇𝑟𝑢𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 + 𝐹𝑎𝑙𝑠𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝐹𝑎𝑙𝑠𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
(4.12)
Sensitivity: Sensitivity measures the true positives correctly identified, utilizing the
proposed strategy. The articulation for the sensitivity is,
𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑟𝑢𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 (𝑇𝑃)
𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑟𝑢𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒(𝑇𝑃)+𝐹𝑎𝑙𝑠𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒(𝐹𝑁)
(4.13)
𝑇𝑟𝑢𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 (𝑇𝑁)
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑟𝑢𝑒𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒 (𝑇𝑁)) + 𝐹𝑎𝑙𝑠𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 (𝐹𝑃)
(4.14)
4.2.4.COMPETING METHODS
89
4.3. PERFORMANCE ANALYSIS
With the help of the metrics, sensitivity, specificity and detection accuracy, the
performance analysis of the implemented RBIS segmentation method is carried out.
The metrics' sensitivity, the specificity and the detection accuracy must have to
improve as it will enhance the performance of the methods.
90
The detection accuracy based performance analysis for varying image sizes is depicted
in figure 4.7. In the figure, the limit for detection accuracy value given is 0.9 to 0.96
and the value of the iterations varies from 40 to 160. The iteration is selected as 40,
and then the detection accuracy for image size 64 64 is 0.9023, 128128 is 0.9237,
192192 is, 0.9399, and 256 256 is 0.9448.
Table 4.1: Performance analysis for RBIS method based on image size for
Specificity
64 64 0.9023
128128
0.9237
40
192192
0.9399
256 256
0.9448
64 64 0.9102
128128
0.9344
160
192192
0.9469
256 256
0.9547
When the number of iterations is 160, the detection accuracy value for different image
sizes is measured and it is, for image size 64 64 is 0.9102, 128128 is 0.9344,
91
192192is 0.9469, and 256 256 is 0.9547. It is visible from the above figure 4.7
that for the image size, 256 256 the detection accuracy value gained is maximum that
is 0.9547 for the iteration count 160.
The performance analysis for the sensitivity metric is demonstrated in figure 4.8.for
the variation in the number of iterations is from 40 to 160. The number of iteration is
40, and the sensitivity value for the image size 64 64 is 0.8870, 128128 is 0.9306,
192192 is 0.9519, and 256 256 is 0.9616.
92
Now, the number of iteration is 160, and the sensitivity value measured for the image
size 64 64 is 0.8874, 128128 is 0.9313, 192192 is 0.9522, and 256 256 is
0.9616. After observing the figure, we can conclude that for the number of iteration
160, and the image size is 256 256 then the proposed method gained better
enactment as associated to the existing techniques.
Table 4.2: Performance analysis for RBIS method based on image size for sensitivity
64 64
0.8870
128128
0.9306
40
192192
0.9519
256 256
0.9616
64 64
0.8874
128128
0.9313
160
192192
0.9522
256 256
0.9616
93
Figure 4.9: Performance analysis of the proposedRBIS segmentation method based
on image size for Specificity
The performance analysis forthespecificity metric is depicted in figure 4.9. Here, the
variation in the number of iterations is from 40 to 160. The number of iteration is 40,
and the specificity value for the image size 64 64 is 0.8289, 128128 is 0.8202,
192192 is 0.8015, and 256 256 is 0.8029. Now, the number of iteration is 160 and
the specificity value measured for the image size 64 64 is 0.8309, 128128 is
0.8269, 192192is 0. 0.8113, and 256 256 is 0. 0.8047. After observing the figure
4. 9, we can conclude that for the number of iteration 160, and the image size is 64 64
then the proposed method gained better performance as associated to the existing
techniques.
94
Table 4.3: Performance analysis for RBIS method based on image size for
Specificity
64 64
0.8289
128128
0.8202
40
192192
0.8015
256 256
0.8029
64 64
0.8309
128128
0.8269
160
192192
0.8113
256 256
0.8047
95
Figure 4.10: Performance analysis of the proposedRBIS segmentation method based
on the number of grids for Detection accuracy
Table 4.4: Performance analysis for RBIS method on the number of grids for
detection accuracy
3
0.8702
4
0.9338
40
5
0.9463
6
0.9571
96
3
0.9102
4
0.9398
160
5
0.9573
6
0.9687
97
The performance analysis forthesensitivity metric is demonstarted in figure 4.11. for
the variation in the number of iterations is from 40 to 160. The number of iteration is
40, and the sensitivity value measured for the grid size 3 is 0.8711, for grid size 4 is
0.9544, for grid size 5 is 0.9751, and for grid size 6 is 0.9673.
The same way, when the number of iteration is 160, then the sensitivity value measured
for the grid size 3 is 0.8712, for grids size 4 is 0.9573, for grid size 5 is 0.9791, and for
grid size 6 is 0.9721. The figure shows that for the grid size 5, and iteration count 160,
maximum sensitivity measure of 0.9796.
Table 4.5: Performance analysis for RBIS method on the number of grids for
sensitivity
3 0.8711
4
0.9544
40
5
0.9751
6
0.9673
3 0.8712
4
0.9573
160
5
0.9791
6
0.9721
98
The performance analysis forthespecificity metric is depicted in figure 4.12Here, the
variation in the number of iterations is from 40 to 160. The number of iteration is 40,
and the specificity value measured for the grid size 3 is 0.7787, for grid size 4 is 0.7255,
for grid size 5 is 0.6117, and for grid size 6 is 0.6137.
In the same way, when the number of iteration is 160, then the specificity value
measured for the grid size 3 is 0.7815, for grids size 4 is 0.7269, for grid size 5 is
0.6124, and for grid size 6 is 0.6378. The figure shows that for the grid size 3, and
iteration count 160, maximum specificity measure of 0.7815.
99
Table 4.6: Performance analysis for RBIS method on the number of grids for
Specificity
3
0.7787
4
0.7255
40
5
0.6117
6
0.6137
3
0.7815
4
0.7269
160
5
0.6124
6
0.6378
The proportional analysis is performed based on the detection accuracy metric for
varying number of iterations is as shown in figure4.13. Here, the number of iterations
100
is 70, and the detection accuracy values for the existing method are, DVPE is 0.9141,
MMBMT is 0.8368, DEEP CNN is 0.9519, and EW-CSA based DEEP CNN 0.9579.
101
method gained a maximum detection accuracy value as compared to existing methods.
After analyzing the above data, we can infer that the projectedRBIS
segmentationtechniquehas accomplished better enactment than the existing
comparative techniques.
The specificity metric is used for comparative method analysis, and it is depicted in
figure 4.14. The number of iterations is selected here is 40, and the specificity value
for DVPE is 0.74, MMBMT is 0.6088, DEEP CNN is 0.5, and EW-CSA based DEEP
102
CNN is 0.6677. The specificity value for the proposedRBIS segmentation method is
0.7817 for 40 iterations. Here, we can observe that the projectedRBIS
segmentationtechnique has given the highest specificity value as compare to other
existing methods. Now, the number of iterations is increased to 13, then the specificity
value for DVPE is 0.74, MMBMT is 0.7085, DEEP CNN is 0.8919, and EW-CSA
based DEEP CNN is 0.7590. The proposed RBISsegmentation method has the
specificity value 0.8402. The above data clearly shows that the projectedechnique has
gained maximum specificity as associated to existing methods.
The sensitivity metric comparative analysis for varying iterations for the projected
method and the existing method is described in figure 4.15. The number of the iteration
103
is 100 and the existing technique DVPE has 0.8712 value , MMBMT has 0.9073 value
, DEEP CNN has 0.9707 value, and EW CSA based DEEP CNN has 0.9718. These
values are less when related with the projectedtechnique. The proposedRBIS
segmentationtechnique has a sensitivity value of 0.9903 for the 100 iterations. Now,
the iterations are increased, it is 130, so the sensitivity value for DVPE is 0.8712,
MMBMT is 0.9373, DEEP CNN is 0.9733, and for EW-CSA based deep CNN is
0.9767. The proposedRBIS segmentation method’s sensitivity value or 130 iterations
is 0.9908. It is visible that the proposedRBIS segmentation method has performed well
and improved sensitivity value 0.9916 when no of iterations is 160.
DVPE
0.9141 0.8712 0.74
DCNN
0.9838 0.9762 0.7041
MMBM
0.9107 0.9417 0.7440
104
The performance measures used for achieving the performance.The specificity is 0.74,
sensitivity is 0.8712 and and detection accuracy is 0.9141 for DVPE. The specificity
is 0.7440, sensitivity is 0.9417, and detection accuracy is 0.910 for MMBMT
TheDCNN has the performance specificity 0.7041, sensitivity is 0.9762 and detection
accuracy is 0.9838, The EW-CSA based deep CNN gives specificity 0.7655,
sensitivity is 9788 and detection accuracy is 0.9785.The proposedRBIS segmentation
method has given an improved performance as compared to existing methods. Its
specificity is 0.8863, sensitivity is 9916 and detection accuracy is 0.9889. The
improved performance of the proposed system shows that it is a useful technique for
lane detection.
Summary
105