An efficient technique for retrieval of color images in large databases

Traditional image retrieval systems match the input image by searching the whole database repeatedly for various image features. Intermediate results produced for these features are merged using data fusion techniques to produce one common output. In this paper, a new image retrieval technique is presented, which retrieves similar images in three stages. A fixed number of images is first retrieved based on their color feature similarity. The relevance of the retrieved images is further improved by matching their texture and shape features respectively. This eliminates the need of fusion and normalization techniques, which are commonly used to calculate final similarity scores. This reduces the computation time and increases the overall accuracy of the system. Moreover, in this technique, global and region features are combined to obtain better retrieval accuracy. Experimental results on two databases (COREL and CIFAR) have shown that the proposed technique produces better results while consuming less computation time for large image databases.