Research And Implementation Of Face Selection And Recognition Algorithm Under Non-cooperative Conditions

The rapid development of science and technology ushered in the tide of artificial intelligence development.In order to apply artificial intelligence to every corner of life,people put forward more and more strict requirements for artificial intelligence,among which face recognition technology is facing such challenges.The existing face recognition technology has a good performance in recognizing face pictures with uniform illumination,correct face posture and no large expression under ideal controlled conditions,but the accuracy of face recognition will be seriously lowered due to the unsatisfactory acquisition equipment and complex acquisition environment under uncooperative conditions.In order to improve the stability of face recognition and implement the algorithm on the embedded development board,the thesis mainly does the following work.Firstly,face recognition technology is divided into three parts,namely: face detection,face selection and face recognition.Considering that in practical application,the processing starts from the image collected by the equipment,the face detection algorithm is to detect whether there is a face in the image and get the position of the face.The face detection MTCNN?O algorithm in this thesis is improved based on MTCNN face detection algorithm.the algorithm MTCNN?O not only ensures the accuracy of face detection,but also greatly reduces the operation time of detection.moreover,the detection algorithm can classify the detected faces,which provides better conditions for the later face selection algorithms.In the part of face selection,it is proved that the low-quality face pictures have a great influence on the performance of recognition network,which shows the necessity of face screening algorithm.Then,the face screening algorithm SER in this thesis is compared with the traditional image quality evaluation algorithm NRQ?LBP algorithm and the Best-Rowden method based on data set quality calibration on two data sets with diverse face samples,which proves that the screening algorithm for recognition network in this thesis is more conducive to the stable performance of recognition algorithm.In addition,in order to balance the screening effect and operation time of the algorithm,the parameters of the face screening algorithm are adjusted.The face recognition algorithm is based on the improvement of Mobilefacenet network.firstly,the redundancy in the network structure is deleted,which improves the network operation time with less precision reduction.secondly,a small convolution network Mask structure is added to the simplified network,which improves the face recognition accuracy of the face recognition network with incomplete face information.finally,a loss function combining Softmax Loss and Center Loss is choosed to train.Finally,the recognition accuracy of the trained recognition network on LFW data set is98.74%.At last,the above algorithms are transplanted to the embedded development board TMDSEVM5728,and the network computing is optimized and accelerated by OpenCL and Linear Algebra library according to the hardware characteristics of the development board.Finally,the algorithm in the thesis is compared with the conventional face technology algorithm,and the feasibility of the algorithm in the thesis is finally proved.