Image Retrieval Using Multi Texton Co-occurrence Descriptor

Image retrieval system is one of a challenging topic and is not yet finalized. A number of features extraction methods has been proposed, for example Gray Level CoOccurrence Matrix (GLCM), Texton Co-Occurrence Histogram (TCM), Multi Texton Histogram (MTH), Micro Stucture Descriptor (MSD), Enhanced Micro Structure Descriptor (EMSD) and Color difference Histogram (CDH). However, the precision rate of those methods are relatively low, between 40% and 60%. Therefore, there is a need of a new approach to improve the results. Looking to those methods, in term of computational complexity, MTH is the simplest. The problem is that there is weakness in representing image features. First, MTH using local features to representate the image. Second, The weakness occurs in the proces of detecting pairs of pixel using texton for color quatization and edge orientation quantization. This study proposes a new approach to perform features extraction in image retrieval systems. Contribution of this study is to add new types of Texton to detect pairs of pixels and adding GLCM features. The method in this study is called Multi Texton Co-Occurrence Descriptor (MTCD). MTCD works by extracting color features, texture features and shape features simultaneously using Texton, then calculates the global image representations with GLCM. Texton detects concurrency of pairs of pixels on each RGB component and the edge orientation of image, while GLCM represents the image as global viewpoint by the value of energy, entropy, contrast and correlation. Features that are detected by MTCD are presented as histogram. The data used in this study is a 300 batik data and a 10,000 Corel data. In order to measure image similarity, Canberra Distance is used. For performance measurement, precision and recall are used. Test data randomly selected consists of 50 Batik data, 2,500 for Corel 5.000 and 5.000 for Corel 10.000. Based on the results of the testing that has been done, the addition of 2 new texton and GLCM features can improve the precision 2.86%, 3,40% and 3,06% on Batik, Corel 5.000 and Corel 10.000 respectively. MTCD is superior than MTH for image retrieval.