Efficient rotation- and scale-invariant texture analysis

Abstract

Texture analysis plays an important role in content-based image retrieval and other areas of image processing. It is often desirable for the texture classifier to be rotation and scale invariant. Furthermore, to enable real-time usage, it is desirable to perform the classification efficiently. Toward these goals, we propose several enhancements to the multiresolution Gabor analysis. The first is a new set of kernels called Slit, which can replace Gabor wavelets in applications where high computational speed is desired. Compared to Gabor, feature extraction using Slit requires only 11 to 17% of the numeric operations. The second is to make the features more rotation invariant. We propose a circular sum of the feature elements from the same scale of the feature vector. This has the effect of averaging the feature vector from all orientations. The third is a slide-matching scheme for the final stage of the classifier, which can be applied to different types of distance measures. Distances are calculated at slightly different scales, and the smallest value is used as the actual distance measures. Experimental results using different image databases and distance measures show distinct improvements over existing schemes.