Flexible Texture Based Image Retrieval Of Multi-Parameters

Texture based image retrieval is a hotspot subject in image retrieval. But most of the presented methods don't take geometrical variant and noisy effect into account, so when images are distorted, the retrieval effectiveness becomes low. This paper concerns on flexible texture image retrieval, so as to implement invariant retrieval. The main contributions are as follows:(1)Based on the good directional selectivity of dual-tree complex wavelet transform (DT-CWT), a scale and rotation invariant texture image retrieval algorithm is proposed. It places proper filters between those at dyadic scales in the DT-CWT, converts the scale distortion into coefficients translation in the scale dimension, in this way, scale invariant is achieved.(2)Take advantage of the translation invariant of DT-CWT, log-polar transformed autocorrelation image is introduced to eliminate the geometrical distortions. Influence of white noise is reduced by modifying autocorrelation image. After decomposed by DT-CWT, a feature vector invariant to translation, scale, rotation as well as white noise effect can be extracted from the wavelet subbands.(3)In order to improve the robustness of invariant retrieval, wavelet geometrical features (WGF) for texture characterization is proposed, which takes into account of the spatial information contained in a wavelet repre- sentation. It can be viewed as an extension of statistical geometrical features (SGF) by introducing multi-resolution information. WGF characterize the granulometry and roundness of dark and bright blobs in the texture, as well as the multi-resolution aspect of human vision system, combined with the invariant analysis, retrieval effectiveness is improved.In order to demonstrate the efficiency of aforementioned methods,experiments are carried out on geometrical distorted texture databases and the general texture database respectively. Simulation results show that the proposed approaches overcome the flaw that traditional invariant methods perform worse on general texture database, not only implement invariant texture image retrieval, but also perform well on routine textures.