The Research Of Occlusion Face Recognition System Based On ZYNQ

Face recognition as an identification technique,which often necessary to confirm the identity of a person in a non-contact situation,such as visitor access,access control and many other application scenarios.Over the years,face recognition research has focused on the situation where users actively provide data.In these cases,the face data usually has the advantages of light balance,normal posture,normal expression,no occlusion and moderate distance.However,in the case of passively acquiring face data,those conditions are not all available.Those unsatisfactory conditions can cause the face recognition performance to drop rapidly,which makes the actual application of face recognition in passive situations hindered.In these passive face recognition applications,occlusion is the most typical scenario,for example,wanted criminals in order to hide their identities,a person with a broken mouth will actively cover his mouth,a person who get red eye disease actively block the eyes.This thesis focuses on the face recognition application under occlusion,and develops an occlusion face-recognition system with high accuracy,good real-time performance and good offline performance.Based on the general process of face recognition,the author analyzes the difficulty of face recognition under partial occlusion,and then comprehensively considers the general process and difficulties,finally,determines the system algorithm of this thesis.These algorithms are difficult to achieve the expected performance metrics if they are only implemented with a general-purpose processor,and it is difficult to implement the algorithm if only use FPGA.Therefore,the system of this thesis uses ZYNQ that integrates FPGA and dual-core Cortex-A9 processor on one chip is released by is released as the system platform,and using the design idea of software and hardware coordination,an occlusion face recognition system based on ZYNQ is designed.The main work of this thesis has five parts:In the first part,the author described the general process of face recognition,determined the workflow of image acquisition,face detection,feature extraction and face recognition,and explained the typical methods.In the second part,the author analysed the difference between general face recognition and face recognition under the condition of partial occlusion of the face and the difficulty in occlusion.Then,the author designed the algorithm of occlusion facerecognition system.Use improved psychology formulas for RGB to grayscale conversion,use histogram equalization and bilateral filtering to eliminate the effects of illumination and noise,and use the nose,mouth,eyes,and full face features for parallel Ababoost face detection to improve the detection rate of the occlusion face.The 2DGabor features that are robust to illumination and occlusion are extracted from the detected faces,and reduce the dimension of Gabor feature by using PCA algorithm,and the reduced-dimensional Gabor features are sparsely reconstructed using SRC algorithm,and the residual of the original face and the component is calculated to realize face recognition.In the third part,this thesis introduced Hardware and Software Co-design Methods.According to the design flow,firstly,the author analysed the algorithm of occlusion face-recognition system.and then the author divided the hardware function and software functions of the system,the image preprocessing algorithm is implemented in the PL part,other algorithms are put on the PS for implementation,and determine the system communication method that using the AXI bus for PL and PS communication and using off-chip DDR3 for caching.In the fourth part,the author designed the hardware part of the system.Designed the image acquisition module by verilog and packaged into an IP core.Using Vivado HLS implement the image-preprocessing algorithm and packaged into an IP core.The project is built by using Vivado 2015.4,and call a variety of IP cores to to build hardware platforms,such as image pre-processing IP core,AXI Interconnect IP core,VDMA IP core,HDMI driver IP core,etc.to achieve image pre-processing algorithm acceleration,image acquisition and HDMI display drive.In the five part,the author designed the system software part and conducted a system test.Firstly,the author designed the human Machine Interface by Qt of PC,and accomplished the software implementation of face detection algorithm,feature extraction and dimension reduction algorithm and face recognition algorithm based on OpenCV.Then,the author builded embedded software environment on ZYNQ,and carried out system software and hardware joint debugging.Then,the author verified the correctness,convenience and offline of system by using dynamic video image test method.Then,the author using 400 static pictures to accomplished static picture test in oeder to verify the performance of the system.Finally,the author analyzed the test result.The static test results show that the system can correctly cover the occlusion face recognition,the recognition speed is about 1.28 s,the accuracy under non-occlusion is 90%,and the accuracy under occlusion is 75.5%.