Research On Rotation Invariant Feature Extract Algorithm Based On Low-rank Textures

Texture analysis as a fundamental problem in the field of image processing and computer vision has gained a large development in recent decades. The rotational invariance texture analysis as an important branch of texture analysis has received the widespread attention in recent years. Invariant texture analysis has a wide range of applications in many scientific research and practical field, such as remote sensing image analysis, biometrics, medical image analysis, content-based image retrieval and target recognition. Although there have made great breakthrough about texture invariance research in recent years, but the recognition accuracy and efficiency are still not well positioned to meet the needs of practical application.This paper investigates the low-rank matrix showed in the texture images, and builds a mathematical model to describe the texture rotation invariant feature by the rank of the matrix. Combined with the fast convex optimization algorithm, we propose a new invariant feature extraction algorithm based on the rotation of the low-rank texture. The main contents are as follows.Firstly, we analyze the integrity of the advantages that using the rank of the matrix descript the texture image information and the feature that the rank of the matrix in a regular texture image is always very low, and then propose a theory that measure the rotation of the texture image in the rank of the matrix invariant feature with a detailed analysis of its feasibility.Secondly, we convert the process to extract low-rank texture into the matrix rank optimization problems, and establish the mathematical model in view of the above. Based on the convex optimization methods related the matrix rank minimization problem, we propose a new invariant feature extraction algorithm based the matrix rank. Analysis and comparison with traditional invariant feature extraction algorithm by feature points, we can see that our algorithm has the great advantage in extracting the whole invariant texture information from images.Finally, in consideration of the problems of accuracy and efficiency required by the algorithm in practical applications, we raise the multi-resolution strategies and multi-step method to optimize the algorithm implementation process, and check up our algorithm from various angles by three different groups of trials. After a comprehensive analysis of the final results of the experiment, we reach the objective evaluation and summary about the actual treatment effect of the algorithm.