Chapter 4 Results and Discussion - INFLIBNETshodhganga.inflibnet.ac.in/bitstream/10603/45492/4/c4.pdf · Chapter 4 Results and Discussion The purpose of computing is insight, not

� �Chapter 4

Results and Discussion

The purpose of computing is insight, not numbers.

Richard W. Hamming

In this chapter we discuss the results obtained by the proposed CBIR

system. The results obtained are compared with the existing methods in

terms of precision, time and space complexities. The results obtained by the

Haar wavelet, multiresolution with the statistical feature method and

wavelet based orthogonal polynomial are presented the proceeding sections.

4.1 Haar Wavelet Based Image Feature

In order to implement the proposed framework CBIR system, the image

database discussed in the previous chapter is considered. The Haar wavelet

method decomposes the image into a pyramid like multiresolution structure

is discussed in section 2.3 of Chapter 2. Subsequently, the proposed

system determines the optimum level of the pyramid structure which is

presented in section 2.2.2. The features such as spectrum of energy, and

spatial relationship between the pixels extracted using the expression in

Eqs. (2.3.3) and (2.3.4). Based on these features, a feature vector space is

constructed. The Bhattacharyya distance method and orthogonal based

Cosine distance method are employed to find the distance between the

query and the target images, and then obtained distance values are

arranged in ascending order, which is discussed in section 2.3.3. The

empirical distance values between the query image and target image in the

� �database are presented in Table 2.3.1. Due to the space constraints, for

sample, a few of the images such as structure and texture image returned

by the proposed Haar wavelet system is presented in Fig. 4.1.1. In Fig.

4.1.1, the first column is considered as query image and the subsequent

columns are considered as top retrieved relevant images according to the

distance values and their distance values are presented in Table 4.1.1.

query image

(a) (b) (c) (d) (e) (f) (g) (h) (i)

Fig. 4.1.1 Examples of some retrieval results from both VisTex and Corel databases. The first column represents query images and the neighboring columns represent relevant images in indexed order.

4.1.1 Retrieval Performance

In order to evaluate the efficiency of the Haar wavelet based image

retrieval system, the performance is compared with the orthogonal

polynomial model [KRI12], color Histogram [SWA91], color autocorrelogram

� �[HUA97], Wavelet moment[SMI94], BDIP+BVLC [YOU08] and CSD [MAN02].

Of the above mentioned methods, the orthogonal polynomial model extracts

41 texture features, and the color Histogram, color autocorrelogram, and

CSD models extract 128 features, while the Wavelet method and

BDIP+BVLC method extract as low as 96 features, and Multiresolution with

BDIP+BVLC extract 92 features. But the proposed method extracts only 12

features such as mean vector, and variance and covariance matrix. The

mean vector represents a spectrum of the energy of each color and

covariance represents the relationship among the color pixels. A

comparative study is performed between the proposed system and the

existing systems, and the obtained results are tabulated in Table 4.1.2. The

proposed method requires only 192 bits to store a vector feature, which is

less when compared to that of the existing systems. The obtained results

are presented in Table 4.1.3.

Table 4.1.1 Distance values of top relevant images against query image using Vistex database.

relevant images query Image

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

Terrain

0.0881 0.0949 0.0951 0.4388 0.6458 0.6615 0.6676 0.7628 0.7526

Title 0.0437 0.3004 0.3338 0.5457 0.5966 0.6302 0.7916 0.8168 0.9163

Fabric 0.1488 0.3284 0.3956 0.4103 0.4781 0.5085 0.5512 0.7878 1.2610

prison window

0.1797 0.4397 0.4651 0.5234 0.6425 0.8800 0.9256 0.9395 1.0398

Food 0.1923 0.3049 0.4672 0.8213 0.8878 1.5657 2.2153 2.2477 2.3576

Wood 0.2613 0.3822 0.3973 0.4597 0.4973 0.5232 0.5924 0.6913 0.7334

Building 0.1891 0.2599 0.2649 0.2738 0.3245 0.4850 0.6573 0.9698 1.4163

� �Table 4.1.2 Feature vector dimension and its space of the retrieval methods

Table 4.1.3 Target Feature Vector Dimension of the Retrieval Methods

Specification

Methods Color space

Dimension Remarks

Color Histogram RGB 128 number of quantization

levels: 8 in R,4 in G,4 in B

Color

autocorrelogram

HSV 128 number of quantization

levels: 8 in R,4 in G,4 in B

Wavelet

moment

RGB 96 number of decomposition levels: 8

BDIP-BVLC RGB 96 -

Multi-resolution with BDIP+BVLC

HSV 92(60+32) number of decomposition levels: 2

CSD

HMMD 128 -

Orthogonal polynomial mode

- 41 number of decomposition

levels: 3

Proposed method(Image pyramid)

RGB 12 number of decomposition

levels: 5

Methods

Target feature vector size in bits

Remarks

Color Histogram 1024 number of quantization level: 8 in R, 4 in G, 4 in B. number of features: 128

Wavelet moment 768 number of decomposition levels: 8; number of features: 96

Multi-resolution with BDIP+BVLC

736 decomposition level 2; number features: 92

Orthogonal polynomial mode

328 decomposition level 3 features:41

The proposed method

192 optimum level:5; number features:12

� �Further, we have evaluated the performance of the proposed system with

existing methods in terms of average of the precision and recall values, and

the experimental results are presented in Table 4.1.4. The proposed system

retrieves the images with the precision and recall values 0.934 and 0.755

for Corel database and 0.918 and 0.753 for VisTex database respectively.

It is observed from the results presented in Table 4.1.4 that the

performance of the proposed system is well when compared to that of

existing methods. A graphical representation of the results presented in

Table 4.1.4 is given in Fig. 4.1.2.

Table 4.1.4 Performance measure of the proposed system with existing methods.

DB

Proposed system

Orthogonal polynomial

Multiresolution with BDP+BVLC

Wavelet moment method

precision recall precision recall precision recall precision recall

Corel 0.934 0.755 0.926 0.740 0.826 0.712 0.812 0.797

Vistex 0.918 0.753 0.908 0.752 0.794 0.721 0.765 0.724

Fig. 4.1.2(a) Graphical representation of the performance comparison of the proposed system with existing CBIR systems for Corel database.

�� �� �� � � �� �� � �� �� � �� � � � � � � �� � � � � � � � � � � ���� �� ��� !��" #�$�%&�'()

* + , - . / 0 1 2 3 4� � � 5 6 � 6 � 7� � 5 8 9 9

M1: Orthogonal

polynomial

M2:

Multiresolution

with

BDIP+BVLC

M3: Wavelet

moment method

: ; ;

Fig. 4.1.2(b) Graphical representation of the performance comparison of the proposed method with existing CBIR systems for Vistex database.

4.2 Multiresolution with Statistical Features Based Method

Multiresolution feature based image retrieval method is experimented

with standard databases taken from Vistex and Corel databases, which is

described in chapter 3. During the experimentation, the proposed method

decomposed the image into 5 level multiresolution structures to form the

pyramidal image. The color autocorrelogram as color feature and a set of

texture features such as contrast, coarseness and directionality, are

extracted at optimum level of the pyramid structure, which is formed as

feature vector, and are discussed in chapter 3.

In order to implement the proposed method, we have conducted the

experiment using Vistex and Corel databases which are contains natural

and texture images. In the first set of experiments, Vistex database is

considered. Young, et al. [YOU08] have used Vistex database and reported

multiresolution wavelet based texture image retrieval using four scale

frequency and two orientations. They constructed features vector based on

a combination of color autocorrelogram and BDIP-BVLC moments and

�� �� �� � � �� �� � �� �� � �� � � � � � � �� � � � � � � � � � � ���� �� ��� !��" #�$�%&�'()

* + , - . / 0 1 2 3 4� � � 5 6 � 6 � 7� � 5 8 9 9

M1: Orthogonal

polynomial

M2:

Multiresolution

with

BDIP+BVLC

M3: Wavelet

moment method

: ; :Minkowski-form distance method is used to find the distance between

query and target images. As a result, the dimension of the color and texture

feature vector size is 92. Kavitha, et al., [KAV11] have also used VisTex

database to propose a texture image retrieval method based on color and

texture features. They have divided the images into sub-blocks of equal

sizes and then the color features are extracted using color histogram and

texture features using GLCM in each sub-block to form the feature vector.

The results of the proposed method are compared with the above said

methods and the results are tabulated in Table 4.2.1.

In the second set of experiments, Corel database is used.

Krishnamoorthy, et al. [KRI12] have used Corel database and have

proposed multiresolution approach for rotation invariant texture image

retrieval, based on orthogonal polynomial model. The orthogonal polynomial

model coefficients are reordered into a multiresolution subband like

structures. Simple statistical and perceptual properties are derived from

each subband of 3-level multiresolution structure to represent the texture

features which is turned as feature vectors. In this experiment, the results

of the proposed method are compared with the methods proposed by

Krishnamoorthy, et al. [KRI12] and Kavitha, et al. [KAV11]. Former method

describes the directional features are combined with mean, standard

deviation, energy and perceptual property at all decomposition levels for

computing the average recognition rate. But the method proposed in this

thesis characterizes both color and texture feature at optimum level only,

and has a short span of feature vectors. Since the proposed system reduces

: ; <time and space complexities, it increases the retrieval recognition rate when

compared to the orthogonal polynomial model [KRI12]. The method

proposed in [KAV11], characterize both color and texture feature in order to

combine the feature vector space, and partition the image into 6 equal parts

and characterize color feature using color histogram and texture features

described in terms of GLCM. The result of the normalized feature vector size

is 74, whereas the proposed method yields 67 features at optimum level to

form the feature vector and also gives a better retrieval performance. In

addition, our method may be considered for online image retrieval and

medical image retrieval oriented applications, since online image retrieval

requires compact feature vector dimension and less computation overhead.

4.2.1 Retrieval Performance

The retrieval performance of the proposed method is evaluated based

on the precision and recall methods. We obtain the average of the precision

and recall values with 0.940 and 0.759 for Vistex database, and 0.921 and

0.762 for Corel database respectively, and the result obtained are compared

with other existing methods, which are presented in Table 4.2.1. A

graphical representation of the results presented in Table 4.2.1 is given in

Fig. 4.2.1.

: ; =Table 4.2.1 Performance measure of the proposed system with other existing

methods

DB

Proposed system

Orthogonal polynomial

model

Multiresolution with BDIP+BVLC

GLCM based method

precision recall precision recall precision recall precision recall

VisTex 0.940 0.759 0.926 0.740 0.826 0.712 0.812 0.717

Corel

0.921 0.762 0.908 0.752 0.794 0.721 0.765 0.681

Fig. 4.2.1(a) Performance comparison of the proposed method with existing CBIR systems for Corel database

� � � 5 6 � 6 � 7� � 5 8 9 9�� �� �� � � �� �� � �� �� � �!��" #�$�%&���� �� ��� '()

* + , - . / 0 1 2 3 4

� � � 5 6 � 6 � 7� � 5 8 9 9M1: Orthogonal

polynomial

M2:

Multiresolution

with

BDIP+BVLC

M3: GLCM

based method

: ; >

Fig. 4.2.1(b) Performance comparison of the proposed method with existing CBIR systems for Vistex database

4.2.2 Retrieval Time

In order to evaluate the time complexity, the proposed system is

implemented with the system specification: Intel Core 2 Duo with a 2.66

GHz processor, 2 GB of memory, and Microsoft Visual C++ 6.0 and Oracle

software. The times consumed by the proposed method to extract the

features, form the feature vector, and retrieve an image are tabulated in

4.2.2. It is observed from the values tabulated that the proposed system

consumes time, which is 2 times lesser than the existing methods.

� � � 5 6 � 6 � 7� � 5 8 9 9�� �� �� � � �� �� � �� �� � �� � � � � � � �� � � � � � � � � � � �!��" #�$�%&���� �� ��� '()

* + , - . / 0 1 2 3 4

� � � 5 6 � 6 � 7� � 5 8 9 9M1: Orthogonal

polynomial

M2:

Multiresolution

with BDP+BVLC

M3: GLCM based

method

: ; �Table 4.2.2 Feature vector dimension, feature extraction, and searching time of the

query image

Proposed system

Multiresolution with

BDIP+BVLC

Orthogonal polynomial model

GLCM based method

Feature vector dimension

12x2 = 24 92x2 = 184 41x2=82 24x2=48

Feature extraction time

0.736s 1.378s 0.975s 1.234s

Searching time

0.06s 0.10s 0.07s 0.09s

Furthermore, we compare the computational complexity of the

proposed method with other three existing schemes. The multiresolution

with BDIP & BVLC method demands 38 additions, 18 multiplications and

15 comparisons, GLCM demand more additions for counting the number of

pixels of 72 bin histogram and orthogonal polynomial methods involves few

additions and more multiplications for polynomial coefficients, whereas the

proposed system involves only a few additions and multiplications after the

number of quantization levels decided in the optimum level subband.

Retrieved images

The retrieved images and their distance values by the multiresolution with statistical feature method are presented in Fig. 4.2.3.

Query

image

0.0000 0.7915 0.958 0.9822 1.0073 1.1452 1.2884 1.4075 1.4135 1.4373

Query

image

0.0000 0.7487 0.8704 0.9886 0.9886 0.9956 1.0582 1.2376 1.3278 1.4562

: ; �Query

image

0.0000 0.7003 0.9002 0.9253 1.0506 1.0775 1.192 1.2811 1.3612 1.5759

Query

image

0.0000 0.8001 0.8414 0.8871 1.0649 1.1378 1.2045 1.2123 1.3854 1.4199

Query

image

0.0000 0.3710 0.7958 0.9355 1.7092 1.7436 1.7556 1.7866 1.7989 1.8842

Query

image

0.0000 0.4525 0.8946 0.8998 1.072 1.1789 1.2964 1.3255 1.4462 1.5487

Query

image

0.0000 1.2916 1.3897 1.3915 1.4935 1.5538 1.6906 1.7085 1.7452 1.9318

Query

image

0.0000 1.1500 1.1895 1.2233 1.2305 1.3145 1.3281 1.386 1.4793 1.9572

Query

image

0.0000 0.7266 0.7601 0.8100 0.8277 0.8434 0.9024 0.9039 1.2587 1.3215

Query

image

0.0000 0.8028 1.0149 1.0733 1.1599 1.1772 1.1867 1.5002 1.5778 2.2729

Fig. 4.2.2 The first column represents the query image and neighboring columns are

relevant images.

4.3 Wavelet Based Orthogonal Polynomial Method

During the experimentation, the proposed wavelet based orthogonal

polynomial method decomposes the image into two level multiresolution

: ; �structure using wavelet packet transform (WPT) results in 16 subbands.

The 1LL and 2LLLL subabnds contains an approximation version of original

at half of the resolution, 1/ 1/ 2LH HL HHLL subbands contains edge

information in different directions and 1HH and 2HHHH subbands contain

edges in diagonal direction. The orthogonal polynomial model applied into

each subbands as described in Chapter 2.5. The low-order orthogonal

polynomial applied on the low-frequency subbands and the high-order

polynomial applied on high-frequency subbands for spatially localizing the

subbands coefficients. For color feature, the color autocorrelogram is

derived from low-frequency coefficients such as 1LL and 2LLLL subbands,

which is described in section 2.5.3.1. The color autocorrelogram

discriminate “different” images and identify “similar” images very well. In

addition to that it does not demand computational complexity. A set of

texture features is extracted from GLCM, which is constructed using high-

frequency subbands. The contrast feature 1f is extracted from 1LH

subbabd, correlation 2f features extracted from 1HH subband, variance 3f

derived from 2HHHH subband and the diagonal distribution feature 4f is

considered from 2HHLL , subband, which is described in section 2.5.3.2 of

Chapter 2, The feature vector f is constructed based on the color and

texture elements.

In order to evaluate the performance of the proposed retrieval method,

the two datasets are used: Diabetic Retinopathy Database (DRD), and

Digital database for screening mammography (DDSM). The Manhattan

? @ Adistance method is used to find the distance between the query and the

target images, and the obtained values are stored in indexing order, which

is presented in Fig. 4.3.1 and Fig. 4.3.2. The topmost values in the index

list represent the target image, which is more similar or same to the query

image. We are mainly interested in medical grade of each image on two

medical dataset. In the training phase, the database is divided into training

and testing datasets depends upon the size of the database.

Query image

0.1349 0.2367 0.3113 0.5123 0.6230

Query image

0.2134 0.3145 0.5678 0.6754 0.9886

Query image

0.2345 0.4563 0.6782 0.7543 0.8006

Query image

0.3456 0.4563 0.7654 0.8871 1.0649

Query image

0.5643 0.6432 0.7958 0.8355 0.9092

Fig. 4.3.1 Diabetic Retinopathy Database (DRD)

? @ B

Query image

0.3421 0.4321 0.5643 0.7651 1.0073

Query image

0.1432 0.2341 0.8704 0.9886 0.9886

Query image

0.2314 0.7433 0.8012 0.9253 1.0506

Query image

0.3421 0.4531 0.8414 0.9871 1.0649

Query image

0.5431 0.5467 0.7958 0.8761 1.1510

Fig. 4.3.2 Digital database for screening mammography (DDSM)

In order to evaluate the effectiveness and efficiency of the proposed

wavelet based orthogonal polynomial image retrieval system, performance

comparisons are made with highly adaptive wavelet method [QUE12b],

wavelet optimization method [QUE10] and Effective feature extraction

method [WAN08], considering with top k -matches and precision and recall

measurement. Highly adaptive wavelet method [QUE12b] uses different

wavelet based filter to characterize each query image. They built the image

characterization map based on the image signature, which is computed

using standard deviation and kurtosis at analysis of different scale. The

characterization map is used to rank the images in the increasing order to

? ? @query image. Wavelet optimization method [QUE10] uses two kinds of

generic wavelet based image signature to characterize each subband of the

decomposed image. The wavelet transform with the lifting scheme

framework is used to decompose the image. A distance method is applied to

compute the distance between the two image signatures in order to obtain

relevant images from the image database. Wan Ahmad et al. proposed a

technique [WAN08] for CBIR are identified by comparing different

techniques with various features. They used Gabor and Discrete wavelet

transforms for texture feature extraction, Hu moment invariants and

Fourier descriptor for shape feature, and gray-level histogram and gray-

level coherence vector for intensity. The standalone and combined feature

vectors are generated using feature extraction techniques with CT brain

image data set. These vectors are stored in separate feature vector

databases. Distance metric is used to compute similarity between the

query image and feature vectors of database image. They have reported that

the Discrete Wavelet Frame (DWF) yields good results for texture images;

Gray Level Histogram (GLH) method obtains better results for the Gray-level

images; Fourier Descriptor (FD) method yields good results for shape

feature. The combination of DWF + FD produces remarkable results for

medical images.

The existing methods are considered with different modalities to

capture the relevant image. For fair comparison, we have considered the

important influence factors of CBIR such as color and texture, feature

extraction tool, size of feature vector and distance method. We have

? ? ?evaluated the existing methods with these techniques and features using

the above mentioned database. As a result, the proposed system performs

equally well when compared to that of the highly adaptive wavelet method

[QUE12b], in which the different bases functions with given support is used

to characterize the each query image. Moreover, the proposed system

obtained better performance when compared to that of the optimization

wavelet method [QUE10] and CBIR system proposed in [WAN08]. The Table

4.3.1 shows the performance of the proposed method and other existing

methods in terms of precision and recall measure. The proposed method

obtains the average precision and recall rate as 0.78% and 0.68% on DRD

database and 0.81% and 0.76% on DDSM database. In addition, the

graphical representation of the performance comparisons of the proposed

method and existing methods are plotted and shown in Fig. 4.3.3.

Table 4.3.1 Performance measure of the proposed system with other existing

methods

DB

Proposed system Highly adaptive wavelet method

Wavelet optimization

method

Effective CBIR techniques

precision recall precision recall precision recall precision recall

DRD 0.781 0.679 0.775 0.669 0.695 0.661 0.652 0.641

DDSM 0.813 0.762 0.801 0.752 0.734 0.721 0.714 0.711

? ? C

Fig. 4.3.3(a) Performance comparison of the proposed method with existing CBIR systems for DRD database.

Fig. 4.3.3(b) Performance comparison of the proposed method with existing CBIR systems for DDSM database.

4.3.1 Computation Time

Table 4.3.2 shows the computation time for various components

involved in the proposed system in order to retrieve an image from datasets.

The proposed system takes 1.82s for retrieving an image from DRM

database, 2.65s for DDRM database. This experiment is conducted using

an Intel Core 2 Duo with a 2.66 GHz processor and 2 GB of memory with

VC++ and Oracle software.

D E D F G H G EG I G J G K G EFE FK FH FL FJ FI FG FD FM FE F F

N O P Q P R S TU S V W P T X E X K X HYZ[ Y\ Z]_ ZYa bYcdeZfghi j k l m n o p q r s

Q O S t u R u P vO S t w x x

D E D F G H G EG I G J G K G EFE FK FH FL FJ FI FG FD FM FE F F

N O P Q P R S TU S V W P T X E X K X HYZ[ Y\ Z]_ ZYa bYcdeZfghi j k l m n o p q r s

Q O S t u R u P vO S t w x x

M1: Highly

adaptive

wavelet method

M2: Wavelet

optimization

method

M3: Effective

CBIR

h i

M1: Highly

adaptive

wavelet

method

M2: Wavelet

optimization

method

M3: Effective

CBIR

? ? y Table 4.3.2 Estimation of computation time of the proposed system

Database

Scheme

DRM DDSM

Orthogonal polynomial model

0.23s 0.62s

WPT with Daubechies-4

0.45s 0.58s

Color autocorrelogram

features

0.57s 0.67s

Texture features 0.34s 0.47s

Distance method 0.23s 0.31s

Overall time 1.82s 2.65s

In this study the image feature information are considered from 6

subbands: 1, 1, 1, 1, 2, 2LL LH HH HHLL LLLL HHHH , through the proposed

method generates 16 subbands, because of the sufficient image features are

available in these 6 subbands. This considerably reduces the computational

complexity, and it is the one of the main advantages of the proposed

framework.