Volume 13 Number 5
October 2016
Article Contents
Santosh Kumar Vipparthi and ShyamKrishna Nagar. Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System. International Journal of Automation and Computing, vol. 13, no. 5, pp. 457-467, 2016. doi: 10.1007/s11633-016-0978-2
Cite as: Santosh Kumar Vipparthi and ShyamKrishna Nagar. Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System. International Journal of Automation and Computing, vol. 13, no. 5, pp. 457-467, 2016. doi: 10.1007/s11633-016-0978-2

Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System

Author Biography:
  • Shyam Krishna Nagar received the Ph. D. degree from Department of Electrical Engineering, Indian Institute of Technology Roorkee, India in 1991. He is currently working as a professor at Department of Electrical Engineering, Indian Institute of Technology, Banaras Hindu University, India. His research interests include image processing, content-based image retrieval, digital control systems and model order reduction. E-mail: sknagar.eee@iitbhu.ac.in

  • Corresponding author: Santosh Kumar Vipparthi received the B. Eng. degree in electrical and electronics engineering from Andhra University, India in 2007, the M. Eng. degree in systems engineering from the Indian Institute of Technology, India in 2010, where he is currently a Ph. D. degree candidate at Department of Electrical Engineering. Currently, he is working as an assistant professor at Department of Computer Science and Engineering, Malaviya National Institute of Technology, India. His research interests include image processing, content-based image retrieval and object tracking. E-mail: santu155@gmail.com (Corresponding author) ORCID iD: 0000-0002-5672-3537
  • Received: 2014-04-26
  • Accepted: 2014-12-04
  • Published Online: 2016-07-25
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Figures (14)

Metrics

Abstract Views (453) PDF downloads (10) Citations (0)

Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System

  • Corresponding author: Santosh Kumar Vipparthi received the B. Eng. degree in electrical and electronics engineering from Andhra University, India in 2007, the M. Eng. degree in systems engineering from the Indian Institute of Technology, India in 2010, where he is currently a Ph. D. degree candidate at Department of Electrical Engineering. Currently, he is working as an assistant professor at Department of Computer Science and Engineering, Malaviya National Institute of Technology, India. His research interests include image processing, content-based image retrieval and object tracking. E-mail: santu155@gmail.com (Corresponding author) ORCID iD: 0000-0002-5672-3537

Abstract: This paper presents a new feature descriptor, namely local extreme complete trio pattern (LECTP) for image retrieval application. The LECTP extracts complete extreme to minimal edge information in all possible directions using trio values. The LECTP integrates the local extreme sign trio patterns (LESTP) with magnitude local operator (MLOP) for image retrieval. The performance of the LECTP is tested by conducting three experiments on Corel-5 000, Corel-10 000 and MIT-VisTex color databases, respectively. The results after investigation show a significant improvement in terms of average retrieval precision (ARP) and average retrieval rate (ARR) as compared to the other state-of-the art techniques in content based image retrieval (CBIR).

Santosh Kumar Vipparthi and ShyamKrishna Nagar. Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System. International Journal of Automation and Computing, vol. 13, no. 5, pp. 457-467, 2016. doi: 10.1007/s11633-016-0978-2
Citation: Santosh Kumar Vipparthi and ShyamKrishna Nagar. Local Extreme Complete Trio Pattern for Multimedia Image Retrieval System. International Journal of Automation and Computing, vol. 13, no. 5, pp. 457-467, 2016. doi: 10.1007/s11633-016-0978-2
  • It is observed that, the multimedia libraries expanded radically in the world wide web owing to easy use of digital cameras. The handling of these immense libraries are extremely annoying rather impractical task. Hence, there is a dire need of some expert search technique to retrieve the relevant images. The text-based image retrieval can be traced back to the late 1970s. Hence, a very popular framework of image retrieval is, first annotating the images by text and then using text-based database management systems (DBMS) to perform image retrieval. Further, many advanced techniques came into the existence, such as data modeling, multidimensional indexing, and query measures. However, there exist two major difficulties, especially when the size of image collections is huge (tens or hundreds of thousands). To overcome these difficulties, content based image retrieval (CBIR) was proposed. In CBIR, the feature extraction is a prominent step and the effectiveness of a CBIR system depends typically on the method of feature extraction from raw images. The CBIR utilizes the visual contents of an image such as color, texture, shape, faces, spatial layout, in order to represent and index the image. There is no single best representation of an image for all perceptual subjectivity, because the user may take the photographs in different conditions (view angle, illumination changes, etc.). However, comprehensive and extensive literature survey on CBIR is available in [1-4]

    Similarly, the color is one of the significant features in the field of CBIR. The color structure in visual scenery changes in size, resolution and orientation. Stricker and Oreng[5] used the first three central moments called mean, standard deviation (STD) and skewness of each color for image retrieval. Further, Pass et al.[6] introduced color coherence vector (CCV) which partitions each histogram bin into two types, i.e., coherent, if it belongs to a large uniformly colored region, or incoherent, if it does not. Swain and Ballar[7] proposed the concept of color histogram as well as introduced the concept of histogram intersection distance metric to measure the distance between the histograms of images. Similarly, Huang et al.[8] proposed a new color feature called color correlogram. The color correlogram characterizes not only the color distributions of pixels, but also spatial correlation of pair of colors.

    Likewise, texture is another important feature for CBIR. Texture analysis and retrieval has gained wide attention in the field of medical, industrial analysis and many more. Various algorithms have been proposed for texture analysis. Smith and Chang[9] used the mean and variance of the wavelet coefficients as texture features for CBIR. Similarly, Ahmadian and Mostafa[10] used the wavelet transform for texture classification. Moghaddam et al.[11] introduced new algorithm called wavelet correlogram (WC). Gonde et al.[12] proposed the multiscale ridgelet transform for image retrieval. Kokare et al.[13, 14] proposed rotated and complex rotated wavelet filters to be used for CBIR. Subrahmanyam et al.[15] proposed correlogram algorithm for image retrieval using wavelet and rotated wavelets.

    In addition to the texture features, the local image feature extraction is attracting increasing attention in recent years. A visual content descriptor can either be local or global. A local descriptor uses the visual features of regions or objects to describe the image, whereas the global descriptor uses the visual features of the whole image. Several local descriptors have been described. Ojala et al.[16, 17] proposed the local binary patterns (LBP) for texture description and these LBPs are converted to rotational invariant for texture classification. Pietikäinen et al.[18] proposed the rotational invariant texture classification using feature distributions. Guo et al.[19] proposed the rotational invariant texture classification using LBP variance with global matching. Subrahmanyam et al.[20-25] proposed various usages of local patterns. Similarly, Vipparthi et al.[26-29] proposed various local texture features for CBIR. Takala et al.[30] proposed block-based texture feature (BLK-LBP), which uses the LBP texture feature as the source of image description for CBIR. Marko[31] proposed the center-symmetric local binary pattern (CS-LBP) which is a modified version of the well-known LBP. Further, Tan and Triggs[32] enhanced local texture feature sets for face recognition. Few other feature extraction methods are available in [33-35].

    The above discussed features are extracted individually and then integrated for final feature extraction. The concepts of local maximum edge binary pattern (LMEBP)[23] and local ternary pattern (LTP)[32] motivated us to propose a new feature descriptor, called local extreme complete trio patterns (LECTP) for image retrieval.

    The major contributions of the proposed work are as follows. 1) The new feature descriptor (LECTP) integrates the local extreme sign trio patterns (LESTP) with magnitude local operator (MLOP) for image retrieval. 2) The performance of the proposed method is also analyzed individually with and without MLOP for image retrieval. 3) The proposed method collects the complete information by considering sign and magnitude of trio patterns which is absent in other existing methods. 4) The retrieval performance of the proposed descriptor is tested on three benchmark Corel-5K, Corel-10K and MIT-VisTex color databases respectively. The retrieval results show significant improvement after adding MLOP with LESTP in terms of their evaluation measures as compared to the state-of-the-art feature for image retrieval.

    The organization of the paper is as follows: In Section 1, a brief review of CBIR and related work is given. Section 2 presents a concise review of feature extraction strategies. The proposed descriptor and its equivalent framework is given in Section 3. The experimental results of various methods using different similarity measures are analyzed in Section 4. In Section 5, we present the summery of the work.

  • LBP is derived from general definition of texture in a local neighborhood. The calculation of LBP is given as follows:

    For a given a center pixel in the 3 × 3 pattern, the LBP value is computed by comparing its gray scale value with its neighborhood as

    $LB{{P}_{N,R}}=\sum\limits_{n=0}^{N-1}{{{2}^{n}}\times {{f}_{1}}({{p}_{n}}-{{p}_{r}})}$

    (1)

    ${{f}_{(x)}}=\left\{ \begin{array}{*{35}{l}} 1,\ \text{if}\ \ x\ge 0 \\ 0,\ \text{if}\ \ x<0 \\ \end{array} \right.$

    (2)

    where $N$ stands for the number of neighbors, and $R$ stands for the radius of the neighborhood, $p_r $ denotes the gray scale value of the center (referenced) pixel and $p_n $ is the gray scale value of its neighbors. The LBP encoding procedure from a given 3 × 3 pattern is illustrated in Fig. 1.

    Figure 1.  Example to calculate LBP and LTP for 3 × 3 pattern

  • The simple LBP in the literature is not enough to deal with the range of visual aspects that commonly occur in the natural images due to different illumination, pose, partial occlusion, facial expression, etc. Hence, LTP[32] was introduced for face recognition under different lighting conditions. Tan and Triggs[32] proposed the local ternary pattern operator. The LTP eliminates most of the effects of changing illumination while still preserving the essential appearance details that are needed for identification. The LTP extends LBP to ternary values called LTP, here the gray scale values in the range of width $\tau $ around $p_{r}$ are quantized to zero, the upper pattern ($p_{r}+\tau $) are quantized to +1 and lower pattern ($p_{r}-\tau $) are quantized to -1, i.e., the indicator $f(x$) is replaced with 3-valued function explained in (3) and the binary LBP code is replaced by a ternary code as shown in Fig. 1.

    $f(x,{{p}_{r}},\tau )={{\left. \left\{ \begin{array}{*{35}{l}} +1,& \text{if}x\ge {{p}_{r}}+\tau \\ 0,& \text{if}|x-{{p}_{r}}|<\tau \\ -1,& \text{if}x\le {{p}_{r}}-\tau \\ \end{array} \right. \right|}_{x=({{p}_{n}}-{{p}_{r}})}}$

    (3)

    where $\tau $ is user-specified threshold.

    After computing local pattern LP (LBP or LTP) for each pixel $(x,y$), the whole image is symbolized by building a histogram as

    ${{H}_{LP}}(l)=\sum\limits_{x=1}^{{{N}_{1}}}{\sum\limits_{y=1}^{{{N}_{2}}}{{{f}_{3}}(}}LP(x,y),l),l\in [0,({{2}^{P}}-1)]$

    (4)

    ${{f}_{3}}(a,b)=\left\{ \begin{array}{*{35}{l}} 1,\text{if}a=b \\ 0,\text{otherwise} \\ \end{array} \right.$

    (5)

    where the size of input image is $N_{1} \times $ $N_{2}$.

  • Subrahmanyam et al.[23] proposed LMEBP. The LMEBP is evaluated by considering the magnitude of local difference between the center pixel and its neighbors. Here, it assigns "1" to this particular center pixel if the edge is positive, otherwise "0" to it. Fig. 2 illustrates the calculation of LMEBP.

    Figure 2.  Extreme to minimal edge calculation using binary and trio values for a given referenced pixel (r)

  • The concepts of LBP, LTP and LMEBP motivate us to propose local extreme sign trio patterns for an image retrieval (IR) system. The LESTP extracts the extreme sign information from an image using trio values. The first extreme edge is obtained by considering sign of local difference between the reference pixel and its eight surrounding neighbors at radius one as shown below.

    ${I}'({{p}_{n}})=I({{p}_{n}})-I({{p}_{r}}),~~~n=1,2,\cdots ,8$

    (6)

    ${{n}_{1}}=\arg (\underset{n}{\mathop{\max }}\,(\left| {I}'({{p}_{1}}) \right|,\left| {I}'({{p}_{2}}) \right|,\cdots ,\left| {I}'({{p}_{8}}) \right|))$

    (7)

    where max($x$) calculates the extreme value in an array "$x$".

    If the pixel edge value is greater than or equal to threshold value $\tau _2$, then it assigns "1" to it. Similarly, if the edge value is lesser than or equal to threshold value $\tau _1 $, then it assigns "-1" to it, if in between the threshold values assigns "0" to it.

    ${{I}^{new}}({{p}_{r}})={{f}_{({I}'(}}{{p}_{{{n}_{1}}}}))$

    (8)

    $f(x,{{p}_{r}},{{\tau }_{1}},{{\tau }_{2}})={{\left. \left\{ \begin{array}{*{35}{l}} +1,\text{if}{{p}_{r}}\ge {{\tau }_{1}} \\ 0,\quad \text{if}{{\tau }_{1}}>|{{p}_{r}}|<{{\tau }_{2}} \\ -1,\text{if}{{p}_{r}}\le {{\tau }_{2}} \\ \end{array} \right. \right|}_{x=({{p}_{n}}-{{p}_{r}})}}.$

    (9)

    The detailed representation of the proposed descriptor is shows in Figs. 2 and 3, respectively. Finally, the image is converted into one upper and one lower pattern having values ranging from 0 to 511.

    $LEST{{P}_{up}}(I({{p}_{r}}))=\{(I_{UP}^{new}({{p}_{r}});I_{UP}^{new}({{p}_{1}});\cdots I_{up}^{new}({{p}_{8}}))\}$

    (10)

    $LEST{{P}_{LP}}(I({{p}_{r}}))=\{(I_{LP}^{new}({{p}_{r}});I_{LP}^{new}({{p}_{1}});\cdots I_{LP}^{new}({{p}_{8}}))\}.$

    (11)

    Figure 3.  Calculation of LESTP for a given 3×3 pattern

    After calculating LESTP, the whole image is represented by building a histogram supported by (12).

    $\begin{align} & {{H}_{{{\left. LESTP \right|}_{\alpha }}}}(l)=\sum\limits_{x=1}^{{{N}_{1}}}{\sum\limits_{y=1}^{{{N}_{2}}}{f(}}{{\left. LESTP(x,y) \right|}_{\alpha }},l),\\ & l\in [0,511] \\ \end{align}$

    (12)

    $f(a,b)=\left\{ \begin{matrix} 1,\quad \text{if}a=b \\ 0,\quad \text{ifotherwise}. \\ \end{matrix} \right.$

    (13)

    The size of input image is $N_1 \times N_2 $.

  • The magnitude local operator is measured by calculating the mean and variances of a given reference pixel at a radius one.

    The mean of LOP ($MLOP_{P,R}^{mn}$) is accounted by using (14).

    $MLOP_{P,R}^{mn}=\text{mean}~{{[mn(n),mn(n+1),\cdots mn(P)]}_{n=1,\cdots ,P}}$

    (14)

    $mn(n)=\text{abs}~({{g}_{n}}-{{g}_{r}}),~~n=1,2,\cdots ,P.$

    (15)

    Similarly, the variance of LOP ($MLOP_{P,R}^{vc}$) is accounted as

    $MLOP_{P,R}^{vc}=\text{mean}~{{[vc(n),vc(n+1),\ldots vc(P)]}_{n=1,\cdots ,P}}$

    (16)

    $vc(n)=\text{abs}~({{g}_{n}}-{{g}_{r}}),~~~n=1,2,\cdots ,P.$

    (17)

    After calculating the mean and variance values of each reference pixel $(x,y$), the MLOP operator is quantized into $L$ levels. Finally, the whole image is represented by constructing histogram as shown in (18)-(20).

    $\begin{align} & MLO{{P}_{mn}}(l)=\sum\limits_{x=1}^{{{N}_{1}}}{\sum\limits_{y=1}^{{{N}_{2}}}{f(MLOP_{P,R}^{mn}}}(x,y),l),\\ & l\in [0,L] \\ \end{align}$

    (18)

    $\begin{align} & MLO{{P}_{vc}}(l)=\sum\limits_{x=1}^{{{N}_{1}}}{\sum\limits_{y=1}^{{{N}_{2}}}{f(MLOP_{P,R}^{vc}}}(x,y),l),\\ & l\in [0,L] \\ \end{align}$

    (19)

    $f(a,b)=\left\{ \begin{matrix} 1,\quad \text{if}a=b \\ 0,\quad \text{ifotherwise}. \\ \end{matrix} \right.$

    (20)
  • The calculation of LESTP is shown in Fig. 2 for highlighted reference pixel. The local difference between the center pixel and its eight neighbor pixels are "5, -8, 0, 8, -1, -7, 1, -2" as shown in Fig. 2. Further, these values are arranged in descending order with respect to magnitudes, i.e., "-8, 8, -7, 5, -2, -1, 1, 0". The first value refers to extreme edge of higher pixel. These eight edge values are converted into trio, (-1, 0, 1), values using two threshold limits $\tau _1 $ and $\tau _2 $. In this paper, we consider thresholds as $\tau _1 $=-2 and $\tau _2 $=2. According to (9), the conversion "-1, 1, -1, 1, -1, 0, 0, 0" is incurred for LESTP. Similar procedure is espoused to enduring eight neighboring pixels. Finally, the LESTP is converted into one upper and one lower pattern. The histogram calculation for these two patterns is shown in Fig. 3. Further, the magnitude of center pixel is calculated in terms of mean and variance as shown in (14)-(17). After concatenating MLOP features with LESTP, it shows a significant improvement than other state-of-the-art techniques. The significance of LECTP is, it extracts complete edge information as compared to well known LBP.

  • The flowchart of the proposed descriptor's framework is shown in Fig. 4 and its algorithm is illustrated below.

    Figure 4.  Block diagram for proposed system framework

    Algorithm 1. The proposed algorithm involves following steps.

    Input: Image

    Output: Retrieval result

    1) Load the image and convert it into grey scale image (if it is colored).

    2) Calculate extreme sign edges trio patterns.

    3) Construct the LESTP histogram for each pixel.

    4) Calculate MLOP of each reference pixel using mean and variance.

    5) Construct MLOP histogram for each pixel.

    6) Construct feature vector by concatenating LESTP with MLOP.

    7) Compare the query image with images in the database using (28).

    8) Retrieve the images based on the best matches.

    Given below are the acronyms used in the analysis of the results.

    LBP: Local binary pattern features.

    CS-LBP: Center-symmetric LBP.

    BLK-LBP: Blocked based LBP.

    LTP: Local ternary pattern.

    DLExP : Directional local extreme patterns.

    LTrPs : Local tettra patterns.

    LESTP: Local extreme sign trio patterns.

    MLOP: Magnitude local operator.

    LECTP: Local extreme complete trio patterns.

  • In these experiments, each image in the database is used as the query image. The performance of the proposed method is measured in terms of average recall, average precision, average retrieval rate (ARR) and average retrieval precision (ARP) as given in (21)-(25)[23].

    For a given query image $I_{q}$, the precision and recall are defined as follows.

    $\text{Precision}\left( P({{I}_{q}},n) \right)=\frac{1}{n}\sum\limits_{i=1}^{\left| DB \right|}{\left| \delta \left( f\left( {{I}_{i}} \right),f\left( {{I}_{q}} \right) \right)|\text{rank}({{I}_{i}},{{I}_{q}})\le n \right|}$

    (21)

    $\text{Recall}\left( R({{I}_{q}},n) \right)=\frac{1}{{{N}_{G}}}\sum\limits_{i=1}^{\left| DB \right|}{\left| \delta \left( f\left( {{I}_{i}} \right),f\left( {{I}_{q}} \right) \right)|\text{rank}({{I}_{i}},{{I}_{q}})\le n \right|}$

    (22)

    where $N_{G}$ is the number of relevant images in the database, $n$ represents the number of top matches considered, $\delta \left( x \right)$ represents the category of $x$, rank $(I_i ,I_q $) returns the rank of image $I_i $ (for the query image $I_q $) among all images in the database ($|DB|$).

    $\delta \left( f\left( {{I}_{i}} \right),f\left( {{I}_{q}} \right) \right)=\left\{ \begin{array}{*{35}{l}} 1,& f\left( {{I}_{i}} \right)=f\left( {{I}_{q}} \right) \\ 0,& \text{otherwise}. \\ \end{array} \right.$

    (23)

    The ARP and ARR are defined using (24) and (25), respectively.

    $ARP=\frac{1}{\left| DB \right|}\sum\limits_{i=1}^{\left| \left. DB \right| \right.}{P({{I}_{i}},n)}$

    (24)

    $ARR=\frac{1}{\left| DB \right|}{{\left. \sum\limits_{i=1}^{\left| DB \right|}{R({{I}_{i}},n)} \right|}_{n\le {{N}_{G}}}}$

    (25)

    where $|DB|$ represents the total number of images in the database.

  • This paper presents four similarity distance measures which are used as given below:

    1) Manhattan or city-block distance or $L_{1}$ distance

    This distance function is computationally less expensive than Euclidean distance because only the absolute differences in each feature are considered. This distance is sometimes called the city-block distance or $L_{1 }$ distance and defined as

    $D(Q,T)=\sum\nolimits_{i}{\left| {{f}_{i}}(Q)-{{f}_{j}}(T) \right|.}$

    (26)

    2) Euclidean or $L_{2}$ distance

    The Euclidean distance is defined as

    $D(Q,T)={{\left( \sum\nolimits_{i}{{{\left| {{f}_{i}}(Q)-{{f}_{j}}(T) \right|}^{2}}} \right)}^{\frac{1}{2}}}.$

    (27)

    The most expensive operation is the computation of square root.

    3) $d_{1}$ distance

    $D(Q,T)=\sum\limits_{i=1}^{Lg}{\left| \frac{{{f}_{T,i}}-{{f}_{Q,i}}}{1+{{f}_{T,i}}+{{f}_{Q,i}}} \right|}.$

    (28)

    4) Canberra distance

    $D(Q,T)=\sum\limits_{i=1}^{{{L}_{g}}}{\frac{\left| {{f}_{T,i}}-{{f}_{Q,i}} \right|}{\left| {{f}_{T,i}}+{{f}_{Q,i}} \right|}}$

    (29)

    where $Q $ is query image, $L_g$ is feature vector length, $T$ is image in database, $f_{I,i} $ is the $i$-th feature of image $I$ in the database, $f_{Q,i} $ is the $i$-th feature of query image $Q.$

  • In this experiment, the performance of the proposed descriptor (LECTP) is tested on Corel-5000 database. The Corel-5000 database consists of 5000 images of 50 different categories. Each category has 100 images of various content ranging from animals, flowers and natural images. The retrieval performance of the proposed descriptor is measured in terms of precision, average retrieval precision, recall and average retrieval rate. The proposed method shows a significant improvement in terms of performance measures as compared to other state-of-art techniques as shown in Fig. 5. The retrieval precision and recall are shown in Figs. 5(a) and 5(b), also group average retrieval precision and average retrieval rate are shown in Figs. 5(c) and 5(d), respectively. From Fig. 5, it is clear that the proposed LECTP shows a significant improvement by adding MLOP with LESTP as compared to other existing methods in terms of their performance measures.

    Figure 5.  Comparison of proposed method with other state-of-art techniques in IR system in terms of: (a) Precision, (b) Recall, (c) ARP, (d) ARR on Corel-5 000 database

    The performance of LECTP is also tested using various distance measures as shown in Fig. 6. From Fig. 6, it is observed that $d_{1}$ distance measure shows a better retrieval performance as compared to other existing distance measures. Fig. 7 illustrates the query results of proposed method on Corel-5K database (top left image is the query image).

    Figure 6.  Illustration of the performance of proposed method with different distance measures in terms of ARP on Corel-5 000 database

    Figure 7.  Two examples for IR system by proposed method (Top left image is query image) on Corel-5 000 database

  • In this experiment, Corel-10000 database[36] is used to test the performance of the proposed method. Figs. 8(a) to 8(d) summarizes the retrieval performance of various methods in terms of precision, recall, ARP and ARR on Corel-10000 database, respectively. From Figs. 8(a) to 8(d), it is evident that LECTP shows better improvement in terms of ARP and ARR on Corel-10000 database. The retrieval performance of LECTP is also tested using various distance measures is shown in Fig. 9. From Fig. 9, it is observed that d1 distance measure shows a better performance as compared to other existing distance measures. Fig. 10 illustrates the query result of proposed method on Corel-10000 database (top left image is the query image). From above experiments, it is observed that the performance of the proposed method shows low retrieval performance on five categories (3, 10, 30, 33, 38) out of fifty categories as compared to the other existing methods as shown in Fig. 5. Similarly, From Fig. 7, it is observed that, the proposed method shows low performance only on five categories (21, 39, 49, 60, 96) out of hundred categories as compared to the other existing methods. The reason behind this is that these categories are having less discriminative information as compared to others. However, the overall (average) performance of the proposed method shows a significant improvement as compared to the existing methods in terms of precision, recall, average and average retrieval rate on Corel-5000 and Corel-10000 databases.

    Figure 8.  Comparison of proposed method with other state-of-art techniques in IR system in terms of: ARR on MIT-VisTex Color database

    Figure 9.  Illustration of the performance of proposed method with different distance measures in terms of ARR on MIT-VisTex Color database

    Figure 10.  Two examples for IR system by proposed method (Top left image is query image) on MIT-VisTex color database

  • MIT-VisTex color[37] database is used here to evaluate our proposed descriptor. Fig. 11 shows an example where one image from each category of MIT-VisTex color database is used. The MIT-VisTex color consists of 640 images which have 40 different textures. Each texture consists of 16 images.

    Figure 11.  Sample images from MIT-VisTex color database (one image from each category)

    The proposed method shows significant performance than existing methods in terms of its performance measures. Fig. 12 illustrates the retrieval performances of various methods. From Fig. 12, it is evident that, the LECTP shows 6.86% improvement compared with standard LBP in terms of ARR. The retrieval performance of LECTP is also tested using various distance measures as shown in Fig. 13. From Fig. 13, it is observed that $d_{1}$ distance measure shows a better performance as compared to other existing distance measures. From Fig. 12, it is concluded that proposed method yields better performance than other state-of-art techniques. Fig. 14 illustrates the query results of proposed method on MIT-VisTex color database (top left image is the query image).

    Figure 12.  Comparison of proposed method with other state-of-art techniques in IR system in terms of ARR on MIT-VisTex color database

    Figure 13.  Illustration of the performance of proposed method with different distance measures in terms of ARR on MIT-VisTex color database

    Figure 14.  Two examples for IR system by proposed method (Top left image query image) on MIT-VisTex color database

  • In image retrieval, feature extraction and similarity measures are two fields which determine the accuracy of the CBIR system. In this paper, a new feature descriptor named, local extreme complete trio patterns is proposed for CBIR system. The LECTP extracts complete extreme to minimal edge information in all possible directions using trio values. The LECTP integrates the local extreme sign trio patterns with magnitude local operator for image retrieval. The performance of the proposed method is also analyzed individually with and without MLOP for image retrieval.

    In this paper, extensive and comparative experiments have been conducted on three public natural databases namely Corel-5000, Corel 10000 and MIT-VisTex color database. The retrieval performance of LECTP is improved by 8.93 %, 3.8% on Corel-5000, 6.14%, 3.5% on Corel-10000 database in terms of ARP and 6.86%, 1.62% on MIT-VisTex color database in terms of ARR as compared to LBP and LMEBP, respectively.

Reference (37)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return