An Exploring Study of Multi-Scale Complexity Texture Descriptors for Medical Image Retrieval

Since texture describes the local information of pixels´ intensity variation, which can be regarded as the non-linear signals, non-linear signal analysis methods may be applied to texture analysis. Complexity analysis, as a popular non-linear signal analysis approach, is widely used for biological and clinical data analysis. In this paper, for exploring study purpose, a two-dimensional structure complexity measure called 2D-C0 was used for texture analysis and feature extraction. Two texture features, called multi-scale complexity texture descriptors, based on 2D-C0 and multi-resolution image analysis are presented for medical image retrieval. One is multi-resolution complexity histogram and the other is wavelet-based multi-scale complexity feature. In order to compute the multi-resolution complexity histogram, a two-dimensional complexity map with the same size as the original image that encodes complexity at every location in the image should be computed and quantized. Detail algorithm about it was discussed. Preliminary experiments showed that the proposed Db2 wavelet-based multi-scale complexity feature can achieve comparable results to Gabor feature.