Research On The Technology Of Face Detection And Recognition Based On Improved CNN

In recent years, face recognition has been widely used in access control. At the meantime,the deep learning network,whose model is hierarchical and has large parameter, has set off a huge wave in face recognition technology, which can display data features better. In depth learning, CNN, with the invariance of rotation,translation and scaling of spatial positions in image processing, can be used to avoid the influence of face translation and other forms of deformation on recognition,thus having a good efficiency in face recognition. However,CNN over-fitting is a difficult problem facing at present. This paper, based on the frontal face recognition in entrance guard, studies the face detection based on CNN neural network and the means to solve network over-fitting and enhance recognition rate under limited samples by using the Caffe deep learning framework. The main work is as follows:?1? Aiming at the problem of CNN over-fitting for face detection, this paper proposes an improved design method of network optimization module in the network.The random and sparse way of the original Dropout training phase then will be improved into three grades ?the former 1/3, the middle 1/3 and the posterior 1/3? in terms of output values, from high to low. Then, the three levels are set with different sparsity ?the sparsity rate is inversely related to the output value? and the whole connection layer is sparse processed in the network test phase. Based on the depth learning Caffe framework, this paper uses the improved dichotomy network to implement face detection, fine tunes the networks with a well-trained model in GitHub and the data, conducts total convolution network transformation in dichotomy CNN, and detects faces with sliding windows.?2? This paper uses improved Alexnet polytomies CNN structure for face recognition. Firstly, different incentive functions are used to analyze the influence of different excitation functions on network performance. In view of the problem that the Alexnet network has many parameters, which is easy to cause the network overfitting and affect the face recognition rate, Alexnet network structure will be improved to make the network parameters lightweight and improve the recognition rate. The design plans are as follows: the maximum number of parameters in the network of a fully connected layer will be firstly removed, then the original 11x11 convolution kernel size Conv1₁ vertical will be split into 7x7 and 5x5 of the size of convolution core, and the Conv1₂ layer of the original 5x5 convolution kernel size will be split horizontally into three layers of 1x1, 3x3 and 5x5 convolution cores. Finally, based on Caffe, an improved Alexnet network is used to realize face recognition. The face images to be recognized are aligned by human eye and randomly divided into 10 sample images with different coordinates starting point, which use the trained multi-class CNN Model to identify.?3? In order to verify the improved CNN algorithm, this paper integrates the face detection and recognition algorithms needed for the experiment,and builds a Cafe-based CNN face recognition system. Based on the experimental comparison, the improved CNN recognition rate has increased by 2.33% and the generalization of the network has been improved. The improved Alexnet network parameters have reduced by 0.7 times, and the recognition rate is enhanced.Finally, this paper summarizes the work and analyzes the shortcomings of it.