A multi-scale and multi-orientation image retrieval method based on rotation-invariant texture features

Texture retrieval is a vital branch of content-based image retrieval. Rotation-invariant texture retrieval plays a key role in texture retrieval. This paper addresses three major issues in rotation-invariant texture retrieval: how to select the texture measurement methods, how to alleviate the influence of rotation for texture retrieval and how to apply the proper multi-scale analysis theory for texture images. First, the spectrum influence between a Radon transform and a Log-polar transform was compared after the rotation effect was eliminated. The average retrieval performance of wavelet and NSCT with different retrieval parameters was also evaluated. Based on such analysis, we developed a multi-scale and multi-orientation texture transform spectrum, as well as a rotation-invariant feature vector and its measurement criteria, which can represent the human visual perception sensitive to texture energy. Then a new rotation-invariant texture retrieval algorithm was proposed. The algorithm was developed based on non-parametric statistical features and it applies low band and high frequency directional bands of NSCT coefficients for coarse filtering and fine retrieval, respectively. Experiments on the Brodatz image database show that the constructed rotation-invariant feature vector is appropriate for capturing major orientations and describing detail information. Besides, the combination of the two-step progressive retrieval strategy and multi-scale analysis method can effectively improve retrieval efficiency while ensuring a high precision compared with traditional algorithms.