Improved texture features for CBIR using response scaling and locally normalised convolution

Independent Component Filters have been shown to be more effective for collection-specific CBIR than generic texture features such as Gabor filter banks. This paper describes changes to a filter selection method and feature extraction process that significantly improve the performance of both ICF- and Gabor filter-based features. We describe, and correct a possible oversight in, a previously published variance-based filter selection method. We also propose the use of locally normalised convolution as a technique to better match texture patterns in images with local intensity differences. We evaluate these changes using a simple CBIR system and our Precision and Recall results are significantly better than those previously published.