Adaptive tetrolet based color, texture and shape feature extraction for content based image retrieval application

The performance of any content-based image retrieval (CBIR) system depends on the quality and importance of the extracted features. Those extracted features like texture, shape, and color carry the most vital image information, reflecting the image's visual perception. Since a natural image possesses these features, in this paper, we have proposed a novel CBIR system that uses all these primitive image features to realize an efficient CBIR system. It has been observed that a natural image contains entirely overlapping information, so in this approach, we have evaluated concerned image features from their respective component. Hence, we have used YCbCr color space for the feature extraction process because Y, Cb, and Cr color planes are minimally overlapped. Since a natural image carries a significant amount of redundant and dispensable pixel values. Hence, as a pre-processing step, we have employed a mid-rise quantization scheme on an individual component. This step reduces the non- essential information and fastens the image feature extraction process by a significant margin. To extract texture and shape information from the intensity, i.e., Y-plane, we have deployed the difference of inverse probability (BDIP) and block variance of the local correlation coefficient (BVLC). We have subsequently used adaptive tetrolet transform in the output of BDIP and BVLC to extract local textural and geometrical features. Parallelly, we have selected the Cb and Cr component and used adaptive tetrolet transform to analyze the regional local color variations of the image. The use of tetrolet transform will enhance not only the local geometrical and textural features but also emphasis the color distribution on the entire image. Finally, we have combined the non-overlapping extracted shape, texture, and color features to form the final feature vector for the retrieval process. The proposed method has been tested on three color dominated, two shape dominated, and textural image dataset and subsequently, results are drawn from each of them in terms of precision, recall, and f-score. Further, the proposed scheme has also been compared with different state-of-art CBIR methods, and the results are showing satisfactory improvement over other methods for most instances.