A Novel Similarity Measure for Content Based Image Retrieval in Discrete Cosine Transform Domain

摘要

Content-based image retrieval (CBIR) scheme has gained popularity in the field of information retrieval for retrieving some relevant images from the image database based on the visual descriptors such as color, texture and/or shape of a given query image. In this paper, color features have been exploited from each color component of an RGB color image by using multiresolution approach since most of the information of an image is undetected at one resolution level while some other undetectable information is visualized in other multi-resolution levels. Initially, Gaussian image pyramid is employed on each color component of the color image and subsequent DCT is computed directly on the obtained multi-resolution image planes. Then some significant DCT coefficients are selected according to the zigzag scanning order. For formation of the feature vector, we have derived some statistical values from AC coefficients and all other DC coefficients are included entirely. Finally, a similarity measure is suggested during image retrieval process and it is found that the overall computation overhead is reduced due to consideration of the proposed similarity measure. The proposed CBIR scheme is validated on a two standard Corel-1K and GHIM-10K image databases and satisfactory results are achieved in terms of precision, recall and F-score. The retrieved results show that the proposed scheme outperforms significantly over other related CBIR schemes.