Research On Face Recognition Based On Convolutional Neural Network

With the rapid development of deep learning,face recognition technology has been applied widely,especially in the field of transportation,its application prospect is very broad.Convolutional neural network is a new artificial neural network model that combines artificial neural network technology with deep learning method,its excellent performance in image processing makes it a research hotspot in the field of face recognition,therefore,the face recognition technology based on convolutional neural network has important research significance.This paper mainly studies the core technologies such as face detection and expression recognition in face recognition.The traditional face detection technology is affected by unrestricted factors and there are problems such as inaccurate positioning,missing detection,false detection,and long detection time,which makes it difficult to obtain satisfying results.This paper first studies the structure of convolutional neural network,back propagation algorithm and loss function,comparative researches the characteristics of face detection algorithms based on multi-task cascaded convolutional network,proposes to adopt the YOLO detection algorithm to construct convolutional neural network and improves the existing YOLO framework.The paper utilizes k-means algorithm to assemble the target bounding box of the training dataset which makes it more suitable for human face detection,and then achieves feature fusion by feature pyramid network which makes the detection results more accurate.The algorithm is tested by using the authoritative face dataset FDDB.The results show that the detection result is superior to the multi-task cascade neural network,and achieves good detection effect under different lighting,different angles,dense crowd and wearing glasses,what's more,the detection speed can meet the real-time detection requirements.In addition,this paper also studies facial expression recognition technology based on convolutional neural network.Firstly,the contrastive loss,triple loss and Softmax loss function are studied.In view of the problem that intra-classes difference is larger than inter-classes difference when expression recognition based on convolutional neural network adopted Softmax loss function as the training supervision signal,this paper proposes to adopt multiple pairs function with Softmax loss function formed joint supervision function for algorithm optimization to maximize inter-classes difference and minimize intra-classes differences.The algorithm performs expression recognition test in a constructed deep convolutional neural network.The results show that the performance of joint supervision function is better than Softmax loss function alone and its discrimination is better.