Research On Multi Pose Face Recognition Algorithm Based On Deep Learning

Under unconstrained scenes,the recognition accuracy of facial images is not high due to the interference of many factors in the collection environment,among which posture change is one of the main factors.This article conducts research on multi pose face recognition through theories such as deep learning image recognition and deep metric learning,and evaluates the recognition performance of the proposed model using both public and self built datasets.The research content of the paper roughly includes the following aspects.First,in view of the limited discriminability of the face feature vectors extracted by the multi pose face recognition algorithm,we improved the face recognition baseline model Res Net50,proposed a spatial and channel dual attention mechanism module to obtain the different weights of each pixel on each layer of the image spatial features and each channel of the feature map respectively,and then optimized the joint loss function based on angle distance to increase the classification interval,Enable the network to automatically learn discriminative facial features that are compact within the class and separated between classes at both spatial and channel scales.By adding attention mechanism modules to different positions in the network to verify the effectiveness of facial discrimination,it was found that a dual attention mechanism was added to the residual position of the network,and learnable weight parameters were introduced for residual correction,which improved the recognition effect.Secondly,in response to the problem of low accuracy in front to side recognition caused by multi pose face deflection angles,a face reconstruction module is added to the network convolutional layer feature extraction to achieve front face reconstruction of multi pose faces.In network training,the original face and the reconstructed face are paired to extract the shallow global and local features through the convolution layer,and feature fusion is carried out through the distribution of learnable weight parameters.The fused eigenface is used as the initial feature representation of the face for further depth feature extraction,to train the model with classification tasks,and to optimize the overall feature mapping learning effect of the model.Finally,in order to make the improved face recognition algorithm suitable for specific application scenarios,the proposed algorithm was tested not only on publicly available datasets such as labeled natural face datasets(LFW)and American celebrity face datasets(CFP),but also on self built multi pose face datasets.Transfer learning is used to verify the effect of face recognition in the actual scene.After verification,the optimal recognition accuracy of the model on the public dataset LFW,CFP-FF,and CFP-FP is 99.56%,96.85%,and 90.54%,respectively.The optimal recognition accuracy on the self built dataset is99.23%.