Automatic image indexing for rapid content-based retrieval

Four models of image data representations are examined for automatic indexing from pixel, nearest neighbourhood, block to full image. For each their invariant properties (translation, reflection and connection) and complexities are assessed. The nearest neighbourhood approach is found to be the best under these criteria. Using the nearest neighbourhood approach, a new automatic feature extraction and indexing algorithm for images on rectangular grid is presented. The algorithm enumerates the number of entire feature points in the ten clusters to form ten integers, which correspond to specific strengths of the ten feature clusters in the image. A probability model is then used to generate a quantitative feature index for supporting the rapid retrieval of images based on their contents. Some sample images and their indexes are also illustrated