Occlusion Face Recognition Based On KPCA Feature Fusion Model

The problem of face occlusion is a great challenge faced by the current face recognition system when it is applied to reality.The problem of face occlusion refers to the problem of human face information being lost because the human face is blocked by shades such as sunglasses,masks,helmets,etc.Once the face occlusion problem appears in places with high security levels such as banks,shopping malls and gateways,it may cause A certain degree of security risk,so to solve the problem of face occlusion in face recognition system has a very important research value.The difficulty in the occlusion discrimination of a face image is mainly reflected in the difficulty in extracting a representative occlusion feature.In the existing feature extraction methods,the features extracted by shallow machine learning methods are not strong enough in generalization ability and stability,and are not suitable for dealing with complex issues such as face occlusion discrimination.In deep learning methods,although the convolutional neural network can extract representative deep features,the training process is too tedious.Although the PCANet model has a simpler network structure than CNN,the overall nonlinear fitting ability of the model is not enough.Therefore,this paper proposes a KPCA feature fusion model based on PCANet to solve the problem of face occlusion.The innovation of this model is reflected in:(1)Introduce kernel transformation at the PCA layer to improve the nonlinear fitting ability of the model.The PCANet of traditional PCANet has weak nonlinear processing capability.Therefore,the original image is partitioned and sampled,then it is mapped into a linear separable high-dimensional space by using a kernel function,and the kernel principal components of the feature space are used to learn the convolution kernel.(2)The concept of fusion between layers is used to enhance the expression ability of the extracted occlusion features.At the same time,the structural detail information in the shallow features and the high-level semantic information in the deep features are retained.The output of the two-layer PCA mapping layer is hash-coded.Then,the histogram information is calculated in blocks,and finally all the histogram features are cascaded.As a final classification feature.(3)Use the random forest to replace the SVM in the traditional PCANet to classify the occlusion category.Experiments show that random forest has higher recognition rate and speed advantage than SVM.