Research On Face Verification Techniques With Convolution Neural Networks

With the development of social economy,people's requirements for the safety and accessibility of authentication are increasing.In recent decades,face-based authentication technology has made great progress.It is direct,friendly and convenient,which has been widespread concern and research.The goal of face verification is to determine whether the two faces is the same person.It is a two-class pattern recognition problem.After the deep learning proposed,especially the convolutional neural networks has been a great success in the field of image recognition,the researchers began to use CNN to extract face features.Experiments result show that the accuracy of face verification have significantly improved.But,to improve the fitting ability,the number of layers and parameters of the CNN model adopted by the face verification system are significantly increased.One of the major problems with these complex CNN models is that the computational burden is large.So,the speed of the face verification system still need to be improved despite the use of high-speed computing devices such as GPUs.Therefore,in the CNN-based face verification study,in addition to increasing the accuracy of face verification,to speed up face verification is also very necessary.In response to these problems,firstly,this paper choose the open source CNN model to extract the face features,fine-tune the CNNs with the manually screened data sets,compare variety feature similarity measurement algorithms,build a complete face verification system.The accuracy rate of this system is 98.35% in the LFW dataset.Secondly,for the problem that the calculation speed of the CNN model is slow,this paper propose to use convolution theorem to accelerate the convolution layer calculation speed,so as to improve the calculation speed of face verification.Convolution layer is the most time-consuming part of the CNN structure,so the core work of accelerating CNN is to accelerate the calculation of convolution layer.Although the convolution theorem is well known in image processing disciplines.However in the context of deep learning,we still need to analyze which convolution layer in CNN can be accelerated by convolution theorem,how to implement convolution theorem efficiently in parallel on the existing GPU platform and how much the actual CNN acceleration effect.So,researches on these problems still has certain theoretical significance and important practical value.In this paper,we analyze the time complexity of the conventional convolution method and the convolution theorem method,give the acceleration conditions for the convolution theorem.We elaborate the convolution theorem algorithm flow,the flow is as follows.First,the input image and the convolution kernel are expanded to the same size and are transformed into the frequency domain by Fourier transform.Second,the summation of the products in the frequency domain is carried out.In this paper,the summation of the products in the frequency domain is transformed into a matrix product by some techniques,in order to make full use of the parallel optimization of the existing GPU library.Finally,the product sum calculation result inverse Fourier transform and crop the extended boundary.So,we get the convolution layer calculation results.In addition,this paper also introduce implementation details based on the GPU,optimization plan and detailed comparison of the experimental results.The experimental results show that the convolution theorem accelerating method could achieve high speed ratio in the case of satisfying the acceleration condition.For the face verification CNN model used in this paper,the method can accelerate 2.2 times.