Deep Face Recognition Method In Natural Scene

The face recognition in the natural scene always is confronted with many disturbances,such as environment illumination,expression,background clutter,camera quality,occlusion and age,etc.In the thesis,the deep learning algorithm was adopted to construct a variety of deep convolution neural network models for face recognition in the natural scene.Unlike most of the traditional face recognition algorithm which the performance is sharply dropped in the natural scene,the algorithms based on deep learning make a good and stable performance on the LFW dataset.The main works of the thesis as follows:1.Based on the forward algorithm,stochastic gradient descent and back propagation,the deep convolution neural network structures were designed for face recognition in the natural scene.With the mini-batch training technology,the CNN model training can convergence in CASIA-WebFace,and in the training process,the data augmentation and pre-processing were used in face database.In each convolution-pooling layer module,the cascade cross channel parametric pooling structure was introduced,then a deeper convolution neural network was constructed with the stronger recognition ability.Based on the traditional sigmoid function,the PRelu activation function was presented by improving the Relu nonlinear activation function.2.Based on the multi-branch convolution neural network,the maximum layer was designed for the disturbances of the face image in the natural scene.With keeping the output dimension of the single convolution-pool module unchanged,the channel width of the hidden convolution layers were increased.The recognition performance of the proposed algorithmwas obviously improved on the face datsets in the natural scene.3.For the situation of the large number of the person under the natural scene,the metric learing was introduced to improve the Softmax loss function by adding the intraclass and interclass distance in the fully connected layers' output.Based on the trait of the gradient descent,the update method was designed for the class center in the model training,and the metric learning was combined into different fully connected layers.With the data-dirven mode in deep learning,the training strategy of deep learning was combined with the face data distribution in the natural scene to improve the loss function with intraclass and interclass distance metric learning,which makes the recognition performance better.