Research On Face Detection Algorithm Based On Convolution Neural Network

The face detection algorithm provides technical support for identity verification,human-computer interaction,security monitoring,public security investigation and other fields.The traditional face detection algorithm has limited information extraction of face features,especially for small face,blocked face and in complex environment face information extraction problems such as missing detection,false detection.Convolutional Neural Network(CNN)is one of the effective algorithms of Deep Learning(DL)in the field of face detection,but the face detection algorithm based on CNN still has some shortcomings,such as too much internal calculation of the algorithm,unstable convergence of the optimization algorithm for parameter updating,and low accuracy of face detection.Therefore,this paper studies the face detection algorithm based on CNN and improves the activation function and optimization algorithm.The main work of this paper is as follows:(1)Aiming at the problems of low learning efficiency and neuronal necrosis in nonlinear and negative features when CNN is used for face detection,a SE(Softsign-ELU)activation function based on CNN is proposed.Combining the advantages of ELU and Softsign,this activation function not only solves the problem of neuronal necrosis when the existing activation function is negative,but also avoids the deviation phenomenon that the output value of the activation function is infinite greater than 0 when it is positive,thus improving the expression ability of the network and reducing the amount of computation.In this paper,with the Retinaface face detection network,the four functions Softsign,Tanh,Re LU,and SE were compared in three different levels of Hard,Medium,and Easy datasets on the public face dataset widerface.The results show that after adopting the activation function SE,the accuracy values of the three datasets are the highest,respectively 63.3,80.2,and 84.3,which confirms that the improved SE activation function has the most prominent effect.(2)Aiming at the problems of poor generalization ability and non-convergence of Adam optimization algorithm,an improved optimization algorithm S-Adam is proposed.By using the Softsign function to control the change of the second-order momentum,the oscillating amplitude of the network learning rate is reduced,In the theory,the convergence rate of S-Adam for convex functions and non-convex optimization problems is proved by using online learning framework.Finally,the experimental results of S-Adam algorithm are compared with those of common optimization algorithms in complex networks.In summary,this paper studies the face detection algorithms based on convolutional neural networks,and proposes SE activation function and S-Adam optimization algorithm,which play a positive role in improving the accuracy of the face detection algorithm based on convolutional neural networks.Therefore,the research in this paper has certain research significance and application value for the face detection based on CNN.