Research On Face Recognition Algorithm Based On LBP And Its FPGA Implementation

Face recognition occupies an important position in the field of biometrics with its special advantages. After decades of development, face recognition technology has made a great breakthrough and a growing number of face recognition technologies as theoretical research in laboratories have begun to step into a variety of fields in our lives in form of products. As a result, people have higher requirements for the speed and other aspects of face recognition systems. Field programmable gate array(FPGA)'s unique hardware architecture makes it has great performance in image processing. So, more and more image processing systems, especially the real-time ones, use FPGA as the image processing core. FPGA-based face recognition algorithms also have been widely researched.Facial feature extraction is the key of face recognition and directly relate to the effectiveness of a face recognition system. Local binary pattern(LBP), as a kind of local feature descriptor in gray-scale, has been widely used in the field of face recognition and reached great success as its simple calculation and efficient in feature description. However,the traditional LBP operator is not perfect and still has some problems need to be solved in the face recognition application. To solve the existing problems, this paper introduces the relative gradient and multiscale image analysis method into the traditional LBP method, called RGMLBP and implements the RGMLBP-based face recognition system using FPGA. During the face recognition process, the FPGA first captures a video frame from a camera and uses median filter to eliminate image noise. After that, the gray-level facial image is converted into the relative gradient domain, then scales the relative gradient facial image to different sizes and extracts features from all these scales of images. Finally, chi-square distances are calculated to measure the dissimilarity scores for each face template stored in the database with the captured face image and through them to find the best match. In the process of hardware implementation, make full use of pipeline structure to realize functional modules' execution in parallel, which greatly accelerates the speed of face recognition. The experimental results demonstrate that our system has a satisfactory recognition rate with cooperative subjects and can reach real-time face recognition for small and medium-sized database. Our design is implemented and evaluated on an Altera Cyclone IV platform.