Image Retrieval by Shape and Color Contents and Relevance Feedback

Abstract

It is quite often that one needs to search a specific image for a particular situation based on visual contents of image while working with the digital images. Content based image retrieval (CBIR) is one of the modern ways to search huge digital image repositories for specific images. With the growing usage of World Wide Web, the use of CBIR has acquired enhanced currency on most of the websites, software and database systems. It is for this reason that CBIR has increasingly attracted the interest of application designers and researchers to enhance its efficiency and make it compatible with its growing demand. The sizes of image databases as well as the applications of image retrieval are increasing. More and more sophisticated algorithms and techniques are being developed to meet the increasing and complex requirements of content based image retrieval applications. Different CBIR systems have different approaches to find images based on their contents and have different performance and accuracy measures. Five algorithms have been proposed in this thesis which is briefly described below. Qualitative analysis for proposed algorithms has been performed with widely used and accepted quantification measures, precision and recall rates and comparison has also been performed with other state of art descriptors which proves the reliability and supremacy of the proposed methods. The first proposed algorithm “An Efficient Content Based Image Retrieval using EI Classification and Color Features” is an effective method for image search and retrieval. It decomposes an image into cells and then extracts its edges to get same number of features as a result of which chances of missing edge features has been reduced. Great number of edge features helps to find more relevant images. Pixels classification is done on the basis of edge pixels and inner pixels. Features are selected from edge pixels for populating the database. Moreover, color differences are used to cluster similar color retrieved results. Level of best performance is to find maximum real edges to retrieve more relevant images. For comparison it is first analyzed with only color features and then with color and shape features combined. Results declare that combined features can obtain higher accuracy. The proposed algorithm is robust against image scaling, rotation and variation but could not provide greater performance for much complicated images. Average precision and recall rates are 90% and 68% respectively. Allowed dimensions of images are between 640x480 pixels and 1024x840 pixels. In the second proposed algorithm “Powerful Descriptor for Image Retrieval Based on Angle Edge and Histograms”, a new angle orientation histogram has been introduced. By applying Pythagorean theory to image, very useful characteristics have been obtained for image matching, search and retrieval. This technique presents satisfied results even for complicated images like one or more complex objects or a natural scene having many objects. It divides the image into angles for more detail and extracts inner and edge ix pixels. Then it computes histogram to carefully analyze changes of edges in the image. The technique is also robust against image scaling, rotation and translation for both simple images and complicated images. Best performance level will be to retrieve the most complicated images like more than one animal in an image. It is compared with existing techniques of AOP, ARP, MPEG-7, SIFT and EPOH and results show that the proposed methodology has achieved greater relevancy. Using a centered point in the process of image decomposition, the method is suitable for different sized images because it always divides the image into equal parts and produces the same number of features that is comparable even in case of images of different size. Precision and recall rates are enhanced by this technique to 94% and 71% respectively. The third proposed algorithm named “Content based Image Retrieval by Combining Multiple Features of Shape and Color” has introduced a new shape features extraction model named symmetry, area, direction angle and arc length features (SADAF) model. SADAF model is generic in nature. In this approach, objects have been located first with image segmentation. SADAF model has then been applied to extract visual shape contents of image. After image matching, searching and retrieval process, the retrieved results have been arranged on the basis of distance of color from query image as compared to the retrieved images, calculated through color histogram. It is compared with the existing techniques of FD, CSS and SRD and best performance results have been obtained for the proposed methodology. SADAF is robust to image alterations like changes in size, scale and orientation because of using symmetry, area, angle direction, and arc length as features from the image. Average precision and recall rates are 79% and 68% respectively. The algorithm has higher processing time and directly proportional to number of images to be retrieved, so it sets initially 50 images to be retrieved against a query image. In this algorithm, when the number of images to be retrieved is low, precision increases and recall decreases. In the fourth proposed algorithm “Content Based Image Retrieval by Shape, Color and Relevance Feedback”, identification of image content is done using combined features of shape, color and relevance feedback which is a key operation for a successful CIR system. By adopting the strategy of combining multiple features of shape, color and relevance feedback for the retrieval of images, very successful results have been obtained. For true image representation, accurate demonstration of shape semantically is essential to achieve correct image matching and retrieval, so shape is used as a primary feature to identify the relevant images whereas color and relevance feedback have been used as supporting features to make the system more efficient and accurate. A good balance between precision and recall is necessary for better system performance and a higher degree of relevancy, so best performance in the proposed algorithm can be achieved if image is segmented more accurately. It presents 0.79 average results against FD, CSS, ART and IM. Average precision for 60 retrieved images is 88% while for 100 it decreases to 68%. Although the technique is good to get more relevant images but if we increase the number of images to be retrieved, its precision gets decreased. The fifth proposed algorithm named „Content Based Image Retrieval Using Combined Features of Shape, Color and Relevance Feedback” presents a unique approach of combined features of shape, color and relevance feedback. A new and effective methodology has been introduced for shape calculation and representation. Shape features are estimated through second derivative, least square polynomial and shape coding methods. Color is estimated through max-min mean of neighborhood intensities. The methodology for relevance feedback is also novel because it is based on key points determination of query image which is an ideal solution to retrieve the best matched images. The index table is automatically populated with features taken by query and database images without bothering the user. Level of best performance will be the highest probability to be counted for an image. The proposed algorithm has been compared with existing color histogram based techniques like CCM, HSI color histogram and CCM, HSI color histogram, CCM and Edge histogram descriptor and HSI color histogram, CCM, Edge histogram descriptor and Relevance feedback. 90% precision and 45% recall has been calculated for the proposed methodology. It is best among all the proposed techniques as its performance is not decreasing while increasing the number of retrieved images and from the results it is also evident that the best matched images have been retrieved through it. Level of best performance for CBIR is higher precision and higher recall and it can be seen from the results that the proposed methods show a nice balance of precision and recall in minimum retrieval time along with robustness against geometric attacks which was a major drawback in the current literature. The achieved results of proposed algorithms comprise of 66%-100% rate for precision and 68%-80% rate for recall. Some recommendations have been given in the end of thesis to overcome various limitations present in the field of CBIR.