A new rotation-invariant and noise-resistant method for texture analysis and classification

One of the basic and important topics in image processing especially for textures is image analysis and classification. In reality, there are some destructive parameters such as "rotation" and "noise" in images. Removing the aforementioned defects has recently become a challenge for many researchers, and consequently various methods have been proposed so far. Among the recently presented methods those based on the multi-resolution transforms have been more popular. This paper presents a new method for texture analysis which is combination of Wavelet, Ridgelet and Fourier transforms. Our approach not only is capable to remove the rotation and noise defects but also in comparison with other approaches its computational cost is less. The method is tested on five datasets, one dataset contains noise free rotated textures and the others are noisy rotated textures. Each dataset includes 2880 textures that produced from 20 main textures. These 20 textures belong to Brodatz album. The results show appropriate performance of the method for texture classification even though the textures are rotated and added with noise.