Research On Application Of Deep Convolutional Neural Network Models For Feature Extraction And Classification

In the era of Internet,the development of computer vision has ushered in a huge explosion,which has spawned a batch of deep learning models such as deep convolutional neural network and greatly improved the performance of visual tasks such as face recognition,object detection and semantic segmentation.Benefiting from the end-to-end stack design of convolutional modules,deep convolutional neural network can extract the pattern information of samples layer by layer from local to global.Users no longer need to choose hand craft features carefully.The model itself can adaptively fit the target function,which is very suitable for more abstract classification tasks such as facial expression recognition.Focusing on the classification of face images,this paper deeply studies the improvement of loss function,ensemble network design and training optimization based on deep convolutional neural network,proposes some improved algorithms designed for face recognition and expression recognition,and has achieved good results.The main research contents of this paper are as follows:(1)Sphere Margins Softmax Loss For Face RecognitionIn order to further improve the ability of the model to extract discriminative features and reduce the complexity of training,we proposed the Sphere Margins Softmax Loss function based on the L-Softmax and A-Softmax which are modified from traditional Softmax Loss,to force the CNN model to extract well-distributed face features.The algorithm will map the facial feature onto a unit hypersphere,and the discriminative features will be learned from the intra-class margin m1 and inter-class margin m2 by setting stricter decision boundaries.Through extensive open-set face recognition experiments on CASIA-Webface and LFW datasets,we find that the proposed SM-Softmax effectively accelerates the initial convergence and outperforms the proposed SATA algorithms in classification accuracy.(2)Weighted Ensemble With Angular Feature Learning For Facial Expression RecognitionAiming at the problem of uneven training status of different branches in the training phase of ensemble network,we proposed a new ensemble network which can automatically adjust the weight of branches with the update of parameters in back propagation.By normalizing the feature vectors and weight vectors,we map the face image into the angle space for feature extraction.The ablation experiment on LFW face database proves that the angular Softmax loss function has excellent facial feature representation ability.In order to evaluate whether the proposed ensemble can improve the training effect of the diverse branches,we used Grad-CAM to generate a saliency map to visually feel the informative region of the face image.The improved training situation of the weak classifier was reflected in the improvement of the final recognition accuracy of the whole ensemble through the majority voting strategy.Comparative experiments both on CK+and AffectNet show that our WDEA is superior to ESR and other excellent models in facial expression recognition task.(3)Design and implementation of face and expression recognition system based on PyTorch frameworkAccording to the research results proposed in this paper,a real-time face and expression recognition system is designed and realized,which proves the practicability of our research content.Based on the optimal model saved in the previous experiments which were trained from scratch,this system adds the modules of image acquisition and face detection and alignment in the face recognition task and finally realizes the complete local face recognition and classification.The system is mainly composed of face recognition and expression recognition,in which sample acquisition module,face detection module and result output module can be shared.Batch samples experiment ed from the subset of AffectNet also prove that the system has good reliability and practicability.