Face feature matching based on semantic Information

Abstract Face feature matching problems have always been a very important research topic in the field of image processing. Face feature matching is always the most important part of common face recognition algorithms, but how to match features quickly and accurately is a quite essential and urgent problem to be solved. Therefore, we propose a face feature matching method which can effectively improve the matching efficiency and accuracy. For each feature point in the input image, face feature matching based on nearest neighbor ratio method requires global search for the best matching in the reference image. However, we find that such matching method can achieve better matching results, but it costs a lot of computing resources and time. Therefore, we use the feature vectors constructed by CNN convolution feature map to provide the local matching region based on semantic information for handcrafted descriptors after feature matching. Although two feature matching processes are used in this paper, the matching efficiency is improved several times compared with the traditional global search scheme. Experiments show that our proposed can be used as an important module of face recognition system to improve the accuracy of face recognition.