Rotation Invariant Texture Retrieval Using Dual Tree Complex Wavelet Transform

Abstract

Texture automatic identification continues to be a challenge in the attempt of object detection and content-based image recognition. While the human eye easily detects various textures, edges, shapes, and objects, the situation is much more complicated for computer-based systems. Images are rarely taken from identical positions, thus, there is an obvious variance in the texture image content. One of the important issues to be solved is rotation invariant texture analysis, actually studied under various methods. Our attempt is based on a new rotated texture analysis using Dual Tree Complex Wavelet Transform (DTCWT). We apply a geometrical transform to the image before extracting the texture characteristics. The feature vectors consist of the mean of the coefficients computed with the DTCWT. Sorting the feature vectors before comparing them is an important stage toward rotation invariance. Our tests proved to return high-accuracy rotated texture retrieval rates (100 % for a dataset with 13 texture images and up to 98.5 % for 112 textures), better than other reported results.