Face Detection And Alignment Research Based On Deep Learning

As a typical application of biometrics,face detection,alignment and recognition have received increasing attention from scientific research.Face detection and alignment,as the key preprocessing steps in face recognition,have a major impact on recognition accuracy.However,existing face detection datasets lack decoration and occlusion and images,and existing face detection algorithms present challenges in face these difficult tasks.The existing face alignment algorithm has a large error in processing low-resolution images,and face images reconstructed by the super-resolution algorithm have artifacts and noise or lack of high-ferequency details.Aiming at the above problems,in this paper,our works include:(1)Optimized face detection algorithm of multi-task learningIn view of the high false detection rate of MTCNN in detecting face images with occlusion and decoration,this paper optimizes MTCNN in three aspects.Firstly,we use the orthogonal matrix to initialize the convolution layer weights to improve the performance of the model.Secondly,the angle margin is introduced for the softmax loss function,which increases the difficulty of learning and improves the robustness of the model.Thirdly,we reduce pooling in order to reduce the amount of calculation and improve model efficiency.The accuracy of optimization method proposed in this paper is 0.2%higher than the original algorithm.In addition,this paper also proposes the social network face dataset SNFA,including occlusion and decorative face images.The rate of false positives of the proposed optimization method is in SNFA dataset is 4%lower than MTCNN.(2)Super-resolution and face alignment algorithm optimization and face alignment applicationIn view of the problem of poor face alignment performance of DAN(Deep Alignment Network)on low-resolution images,artifacts and noise problems of reconstruction images using SRGAN(Super-Resolution Generative Adversarial Network)on small size face images,this paper proposes FSR-SEAN-GAN for joint training the face super-resolution reconstruction and face alignment.FSR-SEAN-GAN includes a face super-resolution network based on feature point heatmaps and generative adversarial networks,and a face alignment network using the squeeze excitation module.FSR-SEAN-GAN improves the super-resolution effect of the face,for 24x24 low-resolution images,the PSNR of the proposed method has a 10%improvement on the 300W dataset compared to SRGAN and ESRGAN,and the proposed method has a 3%or more improvement of face alignment performance on low-resolution images.After that,the DSAN face alignment network is designed based on the depth separable structure.Compared with the DAN model,the convolution layer parameters are reduced by 30%and the model size is reduced by 50%.Finally,based on the image transform layer of DAN,the face recognition application based on face alignment is performed on the LFW dataset.Based on the landmark coordinates predicted by DAN,the application of head pose estimation is carried out.