Research On Face Image Classification Based On Deep Convolutional Neural Network

In the field of face image classification and recognition,compared with traditional machine learning methods,Convolutional Neural Network(CNN)can capture potential features of images independently,replacing the previous means of manual capture,and has been widely used.Especially after the emergence of Alex Net neural network in 2012,deep convolution neural network has brought profound changes to the image field,which can extract deeper abstract features and greatly improve the classification and recognition accuracy of the model.However,with the deepening of the network layer,the number of network parameters becomes huge,and the training is very time-consuming.Nowadays,the perfect modeling theory has not come out,and the further optimization of network structure needs to be solved,and the author finds that the algorithm used in the network training process has obvious limitations,which will have an impact on the network model.In this paper,the model structure of the deep convolution neural network is studied and improved.On this basis,the training process of the model is further optimized,and an efficient deep convolution neural network model is established.The model is transferred to a new application scenario,the specific research contents are as follows:(1)Firstly,the structure and parameters of the VGGNet network model are analyzed,and the most commonly used and efficient VGG-16 is selected as the prototype of the improved model.Multi-scale sampling layer is added at the end of the VGG-16 convolution part,so that the model can be trained and tested by inputting pictures of any size,while reducing the number of neurons in the whole connection layer,and improving the training speed of the model on the premise of ensuring the accuracy.(2)The stochastic gradient descent algorithm is widely used in the training process of deep convolution neural network.The advantage of this algorithm is that it is simple in thought and can achieve an ideal operation speed.However,this algorithm is not suitable for DCNN parameter training.The main reason is that it will appear local optimum problem and gradient dispersion phenomenon in practical training.In this paper,the implementation principle of Particle Swarm Optimization(PSO)is explored and applied to the parameter initialization of deep convolution neural network,which alleviates the gradient dispersion phenomenon in reverse training to a certain extent and guarantees the global optimal solution.(3)Face multi-attribute classification belongs to the category of face imagclassification,but there are few pictures about face multi-attribute annotation at present,and a good classification model can not be obtained only by using a small amount of training data.In order to train an efficient network model and expand the application range of the deep neural network,this paper applies the VGG-MSL model to the face multi-attribute classification problem by means of transfer learning,and achieves a higher classification accuracy.All the above studies were designed and verified by experiments.The experimental results show that the training speed of the improved VGG-MSL model is 13% ~ 16% higher than that of the VGG16 model,and it is not restricted by the size of the image;Compared with the unoptimized model,the classification accuracy of the VGG-MSL model optimized by the particle swarm algorithm is improved by 2% on average,which is a great improvement for the model with high classification accuracy;When VGG-MSL is transferred to face multi-attribute classification problem,its classification accuracy is 40% higher than that of the model only using data training,and it has better face multi-attribute classification ability.