Color, texture, and shape features for content-based image retrieval

We investigate image features useful for comparing images. We are motivated primarily by content-based image retrieval to provide a system for classifying a large set of images based on low-level and intermediate-level criteria, specifically color, texture, and shape.; The main contributions of our work are towards the development of a prototype content-based retrieval system, the Multimedia Analysis and Retrieval System (MARS), a joint effort by several research groups under the Digital Library Initiative. In MARS, we have developed and implemented a shape-matching method that is very fast computationally and invariant to rotation, translation, scale, and spatial quantization. We propose several generalizations of monochrome texture methods to work with color images and discuss their use in segmentation. We also perform a database retrieval accuracy test to quantitatively compare the performance of several texture methods. Based on the test results, we make observations on how color information can improve classification.