| With the rapid development of artificial intelligence and computer vision technology,human behavior analysis,as an important research direction,has been widely used in many scenes.Among them,Dynamic face recognition,as a research field of face recognition without alerting the detected person,has attracted extensive attention in the industry.Therefore,this thesis proposes two dynamic face recognition algorithms for non-cooperative scenarios and the main contents are as follows:(1)Basic technologies of face recognition are researched: Firstly,deep learning is summarized,and the training process of neural network is introduced.Secondly,the traditional face recognition algorithm and face recognition algorithm based on deep learning are introduced.Thirdly,the process of skin color detection and face preprocessing is briefly described.(2)Multi-angle face recognition algorithm based on 3D intelligence is proposed: Firstly,based on the input image to identify the basic feature points,according to the number of detected feature points,an innovative algorithm for face deflection Angle calculation by scene is proposed,and the frontal face feature points reconstruction is combined with ASM algorithm.Secondly,based on the deformation model,3d face reconstruction is carried out,and the deformation coefficient and texture coefficient are predicted by using the cascaded residual network and reconstructed frontal face feature points.Thirdly,according to the deflection Angle of the face,the 3d face is rotated,and the input picture is compared with the similarity.According to the comparison results,the deformation model is self-feedback punishment,and the prediction accuracy of deformation coefficient and texture coefficient is verified,and the face recognition module is added for multi-angle face recognition.Finally,face recognition algorithms such as 3D intelligence,Open CV and YOLOV3 are combined to simulate and compare multi-angle face recognition.(3)Dynamic face recognition algorithm based on new confrontational learning defuzz theory is proposed: Firstly,based on Deblur GAN＿V2,a novel training cut-off method is proposed to deblur the input image by using the new generative adversance network.Secondly,a face blur evaluation model based on human visual characteristics is proposed.According to the change rate of face blur,the images generated by the generator are operated in stages.For images with too high or too low change rate,the Deblur GAN＿V2 algorithm is abandoned for face recognition directly.For the change rate of the picture that is not high or low,adding the fuzzy post-processing module before face recognition.Finally,the new adversarial learning algorithm,Deblur GAN＿V2 algorithm and generalgeneration adversarial network algorithm are combined to compare the simulation and face images. |
