Multistage content-based image retrieval

Content based image retrieval has found its application in many areas like government, academia and hospitals. A new image retrieval technique is presented in this paper, which retrieve similar images in stages. The images are first retrieved based o n their colour feature similarity. The relevance of the retrieved images is then further improved by matching their texture and shape features respectively. Generally a CBIR compare query image feature vector with all other images in the database. This decreases the accuracy of the system as the search encompasses the whole database which contains a wide variety of images. Moreover success of shape based CBIR depends on accuracy of Segmentation technique employed. Unfortunately it has been shown that accurate segmentation is still an open problem. Present approach eliminates the dependency over precise segmentation technique to some extent by narrowing down the search range at each stage. The proposed system has three layers feed forward architecture where each stage output is the input for next stage. Proposed approach also reduces the problem of high dimensionality of feature vector because at each stage only a part of the feature vector representing the desired feature need to be compared thereby resulting in reduction of computational overhead of the overall system. Retrieval in stages also reduces the semantic gap. Advantages of global and region features are also combined to achieve better retrieval accuracy. Experimental results have shown that the proposed system can improve the retrieval accuracy while consuming less computation time.