Research About Image Matching Method Based On The Convolution

As the fundamental technology of image processing and computer vision, image matching is commonly used in face matching, face recognition, finger-print recognition, industrial product testing, traffic management, satellite image interpretation and many other fields. The main methods are summarized into two categories: one is based on the image area(also be called template-based matching) and the other one is based on image features. On the basis of the two above methods, a variety of matching algorithms are proposed. Numerous algorithms for specific applications can achieve a satisfactory result. But almost each one has some limitations, such as, the high computation, the slow computing speed, the narrow application range, the poor anti-interference ability. Therefore, to overcome the problems, new technologies and methods with strong robustness, high accuracy, and good matching effect have become pursue goals for researchers and are becoming more and more popular in this field.In order to explore new algorithms which have high matching efficiency, strong applicability, and can quickly locate the target in the image or video sequence, the knowledge of image preprocessing, templates matching, multi-objective optimization, image convolution, and threshold determination are utilized. A novel method for image matching is proposed based on template matching and convolution. In the method, single template and multi-templates are included, the main research contents:Firstly, build image library. In the paper, human face images are used for experiments. Images are from the Internet, cameras, ORL and Yale face database. After the images are collected, operations such as illumination equalization, size normalization, graying, and histogram equalization should be dealt. Ultimately, the experimental image library is built.Secondly, synthesize convolution template. Based on the normal proportion used in face recognition, and combine the features of the experimental image library, the sizes of templates are set. For single template, the size of face convolution template is needed to be set. For multi-templates, three different sizes for face, eyes and mouth-nose convolution templates need to be set. Then, the MATLAB toolbox is implemented to synthesize them.Thirdly, normalize the pixel value. Pixel values in convolution templates and library images should be extracted from different color spaces on different channels. Then they should be normalized. After above operations, the judgment theory for image matching should be got according to the normalized methods.Fourthly, train threshold. The convolution value is calculated by using all pixel values which are normalized. Then train the thresholds based on them. For single template, the threshold should be trained just according to the distribution of them. For multi-templates, the multi-objective optimization function should be constructed based on optimization theory. Then, train three kinds of different weight coefficients and threshold.For single template, experiments are carried out in separate channels from RGB, HSI, YUV and YCr Cb color spaces. For multi-templates, experiments are performed in the gray space. Experimental results show that, on this image library, the proposed image matching method based on convolution is right, and can achieve the result simply, the amount of computation is small and the speed is fast, the correct matching rate is high. It can achieve a satisfactory matching consequence, even though the images have a slight light affection.The main innovations in the paper include:(1) A novel image matching method based on convolution is proposed, corresponding to the template matching.(2)Through to optimization theory, the multi-objective optimization function is constructed. The method for training different kinds of weight coefficients and threshold is proposed.