VOL. 2, NO. 1, January 2012 ISSN ARPN Journal of Systems and Software AJSS Journal. All rights reserved

Transcription

1 Color and Texture Based Image Retrieval Sridhar, Gowri Department of Information Science and Technology College of Engineering, Guindy Anna University Chennai ABSTRACT Due to the repaid development of internet technology, image documents have become an important information source. It is hard to retrieve certain images from all available ones. An interactive image recommendation system, which firstly uses color histogram feature and GCLM texture feature to express image contents, then a kernel based K-means is utilized to cluster images into multiple classes by their visual features, finally based on a feature vectors stored in the database the similar images are retrieved. The HSV color histogram is calculated and the joint histogram is derived based on the combination of the hue and saturation in the hue and saturation histogram. The color feature is extracted from the joint histogram. The chi-square is used to find the similarity between the two images. Thus global feature is calculated using the joint histogram. The regional feature is extracted using the GCLM technique in which the neighbor pixels is considered into account. The evaluation results demonstrate the accuracy of the retrieval based on the precision and recall false positive and negative ratio. The ROC curve is used to compare the efficiency of the color, texture and the combination of both the color and the texture. Keywords: Image Recommendation, similarity-preserving Image retrieval, CBIR 1. INTRODUCTION A key component of the Content Based Image Retrieval system is feature extraction. A feature is a characteristic that can capture a certain visual property of an image either globally for the whole image, or locally for objects or regions. Some key issues related to CBIR systems are the following. First, how the extracted features can present image contents. Second, how to determine the similarity between images based on their extracted features. One technique for these issues is using a vector model. This model represents an image as a vector of features and the difference between two images is measured via the distance between their feature vectors. There exist two approaches to search, to browse, and to retrieve images. The first one is based on textual information attributed to the images manually by a human. This is called concept-based or text-based image indexing. A human describes the images according to the image content, the caption, or the background information. However, the representation of an image with text requires significant effort and can be expensive, tedious, time consuming, subjective, incomplete, and inconsistent. To overcome the limitations of the text-based approach, the second approach, Content-Based Image Retrieval (CBIR) techniques are used. In a CBIR system, images are automatically indexed by summarizing their visual features such as color, texture, and shape. These features are automatically extracted from the images. In this paper, we present an interactive similar image retrieval system and evaluate which color or texture features are the most efficient to represent similarity of color images. Our initial results show that the color histogram descriptors are not effective features because they do not consider spatial information of image pixels. Therefore, different images may have similar color distributions. In addition, our results show that the co-occurrence matrix features retrieve much more relevant images than other color and texture features. Additionally, in order to increase precision, the combination of color and texture features should be used in CBIR systems. Qirong Bo,Jinye Peng[1] denotes the Image Recommendation System by extracting the color or texture feature based on the image. Yuli Gao, Jianping Fan et al [2] explains that image search engines, such as Google and Yahoo, have done a good job in which after users input a few keywords, image search engines are able to display the images related to the keywords. Jianping Fan et al [3] explains that a topic network is automatically generated to summarize large-scale collections of manually annotated. Yuli Gao et al [4] explains the way of filtering out Junk images from the Google search image. In content based image retrieval system the color and texture feature is extracted and clustering is done in order to group the similar feature vector and the sample images are extracted from the each group of the image. The functionality of content based image retrieval is shown in the below Fig: 1.1 Block Diagram. 1

2 Image Database Color Feature Extraction Texture Feature CLUSTERING ALGORITHM 1.K-Means Figure 1.1 Block Diagram The rest of the paper is organized in the following ways. In Section II color & texture feature is outlined. In Section III the detailed feature extraction is outlined. In Section IV outline about the clustering and Section V is concluding with the results and conclusion. 2. COLOR AND TEXTURE FEATURE In Content Based Image Retrieval [5] the feature extraction plays the major role. The color and texture feature are extracted for the image. These extracted features are used to find the similar images. This section discuss about the function that are used to extract the feature and the feature that are extracted. 2.1 Color Feature Color is an important feature for image representation which is widely used in image retrieval. This is due to the fact that color is invariance with respect to image scaling, translation, and rotation. The key items in color feature extraction consist of color space, color quantization, and the kind of similarity measurements. Color Feature can be extracted using color moment, color histogram, and Color Coherence Vector (CCV) are used. Color Histogram is commonly based on the intensity of three channels. It represent represents the number of pixels that have colors in each of a fixed list of color ranges. Color Moment is based used to overcome quantization effect in color histogram. It represents to calculate the color similarity by weighted Euclidean distance. Color set is used for fast search over large collection of image. It is based on the selection of color from quantized color space. A histogram is the distribution of the number of pixels for an image. The color histogram represents the color content of an image. It is robust to translation and rotation. Color histogram is a global property of an image. The number of elements in a histogram depends on the number similar images use r of bits in each pixel in an image. For example, if we suppose a pixel depth of n bit, the pixel values will be between 0 and 2 n -1, and the histogram will have 2 n elements. The HSV space color histogram is calculated and the joint histogram is calculated by using Hue and Saturation Histogram by calculating the total number of pixel in both the Hue and Saturation Histogram. The joint histogram is calculated using p (h i, s j )=N(h i, s j )/ N total where, N (h i, s j ) is the total number of pixel in both the hue and saturation histogram, N total is the total number of pixel in the image. The joint histogram can be used to efficiently calculate the mean, standard deviation, entropy, skewness and kurtosis of very large data sets. This is especially important for images, which can contain millions of pixels. The sum of all elements in the histogram must be equal to the number of pixels in the image. For evaluation some sample images in order to evaluate different extracted features. In evaluation, a retrieved image is considered a match if and only if it is in the same category as the query image. In addition, the effectiveness of the extracted features has been measured by precision and recall parameters. Precision is the ratio of relevant retrieved images to the total number of retrieved images. Recall is the ratio of retrieved relevant images to the total number of relevant images in the database. 2.2 Texture Feature Texture [6] refers to visual patterns with properties of homogeneity that do not result from the presence of only a single color such as clouds and water. Texture features typically consist of contrast, uniformity, coarseness, and density. There are two basic classes of texture descriptors, namely, statistical model-based and transform-based. The former one explores the grey-level spatial dependence of textures and then extracts some statistical features as texture representation. One example of this group is co-occurrence matrix representation. The latter approach is based on some transform such as DWT. 2D Discrete Wavelet Transform is the wavelet representation of a discrete signal X consisting of N samples can be computed by convolving X with the low pass and high pass filters and down sampling the output signal by 2, so that the two frequency bands each contains N=2 samples. With the correct choice of filters, this operation is reversible. This process decomposes the original image into two subbands: the lower and the higher band. This transform can be 2

3 extended to multiple dimensions by using separable filters. A 2D DWT can be performed by first performing a 1D DWT on each row (horizontal filtering) of the image followed by a 1D DWT on each column (vertical filtering). The first decomposition level (d = 1). In this level the original image is decomposed into four sub-bands that carry the frequency information in both the horizontal and vertical directions. In order to form multiple decomposition levels, the algorithm is applied recursively to the LL sub band. The second (d = 2) and third (d = 3) decomposition levels as well as the layout of the different bands. The 2D DWT has been applied three times on all images. In other words, third decomposition level has been computed. In that level, there are 10 sub bands. The mean and standard deviation of each sub band has been computed as texture features. This means that each image has 60 texture features, which have been obtained using wavelet co-efficient. Gray-level co-occurrence approach uses Gray-Level Co-occurrence Matrices (GLCM) whose elements are the relative frequencies of occurrence of grey level combinations among pairs of image pixels. The GLCM can consider the relationship of image pixels in different directions such as horizontal, vertical, diagonal, and antidiagonal. The co-occurrence matrix includes secondorder grey-level information, which is mostly related to human perception and the discrimination of textures. Four statistical features of the GLCMs are computed. The features are energy, entropy, contrast, and homogeneity. G *G GLCM Pd for a displacement vector d = (dx; dy) is defined as follows. The (i; j) of Pd is the number of occurrences of the pair of gray-level i and j which are a distanced apart. orientation of the image can be calculated using the Gabor texture in which the orientation and the frequency are taken into considerations. A set of 113 images is considered for extracting the feature. The images are JPEG images with standard resolution of 640x480. The images and feature are stored in MySQL database. The features are extracted using the MATLAB. 3. K-MEANS CLUSTERING K-means is one of the simplest unsupervised learning algorithms that solve the well known clustering problem. The procedure follows a simple and easy way to classify a given data set through a certain number of clusters (assume k clusters) fixed a priori. The main idea is to define k centroids, one for each cluster. These centroids should be placed in a cunning way because of different location causes different result. So, the better choice is to place them as much as possible far away from each other. The next step is to take each point belonging to a given data set and associate it to the nearest centroid. When no point is pending, the first step is completed and an early group age is done. At this point we need to re-calculate k new centroids as bar centers of the clusters resulting from the previous step. After we have these k new centroids, a new binding has to be done between the same data set points and the nearest new centroid. A loop has been generated. As a result of this loop we may notice that the k centroids change their location step by step until no more changes are done. In other words centroids do not move any more. In Fig 3.1 the K-Means Clustering is described in detailed. Image Feature Vector Reshape the feature vector Reshape the feature vector Find the center poi nt and calculate no of cluster Figure 2.1 describes the Co-Occurrence Matrix Functionality of the gray level co-occurrence matrix in detail. Display the image cluster Thus the co-occurrence matrix is used to find the pixel level similarity in the image. So, the extracted feature is based on the neighbor pixel value. The different angle images are not retrieved using the GCLM technique. So, the 3

4 3.1 Architecture of K-Means Clustering The images and their feature are stored in the MySQL database. The feature of the image is clustered using the K-Means Clustering algorithm. The image, number of cluster to be formed and the method of choosing the midpoint is given as the input function. 4. RESULTS AND CONCLUSION When the user give the query image the color and the texture feature is extracted and compared with the feature of the images in the database. The six color feature is compared with the six color feature in the color table. The four texture feature is compared with the four texture feature in the texture table. Both color and texture feature are compared with the color and texture feature in the database. As shown above the color and texture feature are extracted and stored in the database the feature of the query images is also extracted and the feature of the query is compared with the database image. Fig 4.1 and Fig 4.2 gives the detail of the feature extraction. In the figure the color feature for image 5 and 6 are same but it varies in the texture so the both color and texture are important while retrieving the similar images. Table 4.1 Color Feature Table Figure 3.1 Original Image I m ag e Color Feature mean Varia -nce Std deviati on Entropy Skewness Kurtosis Figure 3.2 Image where K (no of cluster) = Figure 3.3 Image where K (no of cluster) =4 4

5 I m a g e Texture Feature Table 4.2 Texture Feature Table contrast correlation energy homogeneity within the threshold value is retrieved from the color, texture and color texture table as show below. Query image In color feature database the output is As show below the query image feature is extracted and compared with the feature table and the images which is In Texture Database the output is In Both Color and texture Database Table 4.3 Evaluation of color and texture feature For evaluation some sample images in order to evaluate different extracted features. In the evaluation, a retrieved image is considered a match if and only if it is in the same category as the query image. In addition, the effectiveness of the extracted features has been measured by precision and recall parameters. Precision is the ratio of relevant retrieved images to the total number of retrieved images. Recall is the ratio of retrieved relevant images to the total number of relevant images in the database. S # Color Feature Texture Feature Color & Texture Feature Precision Recall Precision Recall Precision Rec all

6 The yellow banana picture was selected as a query image. Color histogram features can retrieve 7 images 3 of which are banana images, while the other 4 are not. Therefore, its precision and recall are 3/7 = and 3/6 = 0.5, respectively. Texture features can retrieve 5 images 2 of which are banana images, while the other 3 are not. Therefore, its precision and recall are 2/5=0.4 and 2/6=0.3333, respectively. Both the combination of color and texture image can retrieve 2 images where both 2 images are banana images. Therefore, its precision and recall are 2/2=1 and 2/6=0.3333, respectively. Table 5.1, 5.2, 5.3 describes various values of the color, texture and both color texture database. Thus ROC curve is drawn for the set of 5 value of precision and recall for color, texture and both color and texture feature is shown. Positive Ratio and True Positive Ratio. The ROC curve demonstrates that color feature does not give the uniform curve for every query image as the moves up and down. The color feature curve is shown in green color. The texture curve is also not stable as the false positive and true negative is varying for various query images. The color texture feature gives the increasing curve for the query image. REFERENCES [1] Kun Yan; Xiaoyi Feng; He Huang; Shaochong Fan; A novel algorithm for filtering out junk images interactively from web search results, Issue Date : 9-11 July 2010, On page(s): [2] Yuli Gao; Jinye Peng; Hangzai Luo; Keim, D.A.; Jianping Fan; An Interactive Approach for Filtering Out Junk Images from Keyword-based Google Search Results, Issue Date : Dec. 2009, Volume : 19, On page(s): [3] Yuli Gao,Jianping Fan,Hangzai Luo,Shin ichi Satoh A Novel Approach for Filtering Junk Images from Google Search Results,MMM (2008). [4] Jianping Fan, Daniel A. Keim, Yuli Gao, Hangzai Luo, Zongmin Li, JustClick: Personalized Image Recommendation via Exploratory Search from Large-Scale Flickr Image Collections, IEEE transactions on circuits and systems for video technology, vol. 18, no. 8, August Figure 4.1 ROC curve The ROC curve is used to visualize which feature has more information. From the graph it is clear that the combination of both color and texture curve has more accuracy when compared with the only color feature or only the texture feature. In both the color and the texture the accuracy varies frequently compared to that of combing both the color and the texture feature. [5] A.W.M. Smeulders, M. Worring, S. Santini, A. Gupta, R. Jain, Content-Based Image Retrieval at the End of the Early Years, IEEE Transactionson Pattern Analysis and Machine Intelligence, vol. 22, no. 12, pp , [6] C. Gotlieb, H. Kreyszig. Texture description based on co occurrence matrices. CVGIP, 51:70 86, CONCLUSION In this paper, the color feature is extracted from the joint histogram and the texture feature is extracted using the GCLM feature. The k-means clustering is used to cluster the feature of the image. The ROC curve is drawn in order to evaluate the performance of the feature extraction. ROC curve is drawn in order to evaluate the performance of the feature extraction. The ROC curve is drawn for False 6