Rotation-invariant Texture Analysis

Texture invariant analysis is very useful in many applications, such as: image search based on its content,remote sensing image analysis,iatrical image analysis,target recognition and other fields. Although the researches about this have made a great development during the past few decades, the precision and efficiency of the identification are still not acceptable with practical needs. So further more work are needed. The main work of this thesis is summarized as follows:1. A new rotation-invariant texture analysis technique using Radon and Fourier transforms is proposed. In this method, the Radon transform of the image is first calculated and then the Fourier transforms and its corresponding module is computed to extract the invariant features. In the Radon process, the rotation quantity in the rotated image is transformed into the motion quantity in the image; and to maintain the uniformity of the Radon transform for different orientations,we calculate the Radon transform for a disk area of the texture image; then after using a Fourier transform and following making a module, as the module has the property of translation invariant, so the translation quantity is also eliminated and the description of rotation invariant features of texture image are gained. The theoretical analysis and experimental results also show the feasibility and good robustness of this method.2. A rotation-invariant texture analysis method based on the Gabor filters and the statistics invariant moments is also performed. In this technique, texture images are divided into sub-domains with different frequency, and then the spectrum in different sub-domains is inverse transformed into time-domain, and finally identified by the statistics of invariant moments calculated in these different frequency domains. And here the Hu moments,Legendre moments,Zernike moments and pseudo-Zernike moments are calculated respectively for the identification. The experimental results show that Zernike moment is more effective.And in the process of the classification, two different classifier including the k-nearest neighbor's classifier and minimum distance classifier are employed and compared to classify the texture patterns, which euclidean distance is used to caculated the distance between the being-classified set and the training set. For each classifier, experiments are performed with Brodatz and MIT texture images. The results show that the k-nearest neighbor's classifier is better. Moreover, the classification accuracy of the method using Radon and Fourier transforms is better than the method based on the Gabor filters and statistics of invariant moments.In a word, texture invariant analysis is one of the challenging topics in texture analysis, and rotation-invariant analysis is one of the difficult problems. A new rotation invariant texture analysis technique is proposed in this paper, and this method has good classification efficiency and robustness. And those works in this paper are hoped to do some promotion effection with the future researches about the texture invariant analysis.