Research On Image Feature Description Based On Adaptive Fractional Derivative

Images may exist slight blurring in some areas due to moving cameras and mobile device. At the same time, the texture detail especially in smooth areas is usually ignored or improperly described during feature detection. Therefore, these problems bring certain difficulties for image preprocessing, and feature robust representation or detection. In recent years, fractional differential has been widely applied in signal processing, control. Especially, fractional differential has obtained a few achievement in dealing with for texture feature in smooth area or complex texture. In order to better describe complex texture structure, the paper mainly study the establishment principle of fractional differential mask template and its application in image preprocessing, and feature representation. The main work and innovations in the paper are listed as follows,1. An adaptive fractional differential mask template with non-integer step has been developed. Traditional fractional differential operators usually only consider fixed-sized mask template and global fractional derivative order, ignoring the high degree of self-similarity that many images exhibit. Non-integer step and adaptive fractional order are adopted to improve the accuracy of fractional differential computation. Experimental results show that the local texture in smooth areas can be enhanced effectively.2. A novel adaptive fractional differential for orientation histogram establishment has been proposed. Most local descriptors tend to use integral differential and Gaussian weight to construct the orientation histogram to characterize local texture. However, the local texture, especially the blurring pattern in smooth areas, can not be described properly or accurately. Thus, adaptive fractional derivative is used to estimate orientation and magnitude, and joint weight considering both distance and color similarity is also integrated to optimize histogram. Experiment results show that the optimized histogram has an advantage in describing local feature adaptively and accurately.3. A feature description based on adaptive fractional differential has been developed. The blurring in images may bring difficulties for traditional descriptors. Thus, we applied our algorithms into classical descriptors, using the mask template to preprocessing matching images and producing vectors with optimized orientation histogram., experiment results based on the database Graffiti show that the numbers of matching points in two view angles in improved SIFT has been averagely increased 5.94%, and correct matching rate has been averagely increased 9.35%%. In improved SURF, the numbers of matching points has been increased 9.82%, and correct matching rate has been averagely increased 10.85%.