Face Detection And Face Expression Recognition From Natural Scene Images

With the development of AI technologies,people also set a higher request for the human-machine interface design.People are desirous of obtaining a more "human-like"AI assistant,which has ability to recognize the user's emotions and make positive and accurate feedback.Facial expression recognition technology is one of the key links to achieve this goal.Facial expression recognition for practical application is one of the difficulties of current research.The main reason is that the process of face image acquisition is affected by illumination,posture and occlusion,which makes it difficult for designing features by traditional methods to improve the precision.With the support of massive data brought by big data and the maturity of deep learning theory,especially the application of convolution neural network in pattern recognition,it provides new ideas and methods for face detection and facial expression recognition technology.In this paper,we aim at the realization of face detection and expression recognition with the images of facial expression in natural scenes.The convolution neural network,one of the most popular techniques in computer vision field,is used to study the face detection algorithm and facial expression classification algorithm in complex nature scenes.The main work of this paper includes the following two aspects:1.Face detection from natural scenes images based on the deep learning object detection technology.In this paper,three objection detection networks,Faster-RCNN,YOLO and SSD,which are relatively popular in the field of deep learning,are compared.16,106 samples are trained from Wider Face database,and 1000 samples of SFEW database are taken as test sets.The experimental results show that both Faster-RCNN and SSD achieve 100%accuracy in the test set.The detection speed of YOLO is faster,but the detection accuracy is slightly worse.Finally,Faster-RCNN is chose to segment the facial expression area.2.Facial expression recognition from natural scenes images based on convolutional neural network.In this paper,the training and optimization of facial expression recognition network model are realized from the following aspects:In the first place,4 kinds of convolutional neural network structures in common use are studied.Data Augmentation and transfer learning are used to restrain the model over-fitting caused by the shortage of training samples.The experimental results show that the VGGNet-16 model has better recognition accuracy.After that,the problem unbalanced recognition rate of multiple classes and its causes are analyzed.A weighted-loss function is proposed to optimize the network.The experimental results show that the overall recognition accuracy of the model is slightly decreased,while the recognition accuracy of different categories is improved.Then,the network pruning is used to compress the network model,and the over-fitting problem is enhanced by improving the sparsity of network.The recognition accuracy of the model is further improved after the retraining.Finally,a multi-network integration approach of 3 kinds of VGGNet is used to enhance the recognition accuracy of facial expressions.In this study,the highest accuracy rate of a single network model is raised to 54.1%,while the multi-network integration model achieves the highest accuracy rate of 56.2%.Compared with other methods on the database,the accuracy rate of our research is more outstanding.The experimental results show that the algorithm of our research is effective.According to the SFEW database used in the experiment,the accuracy rate of facial expression recognition is improved from the single network model and the multi network integration.The results of the experiment show the effectiveness of the work.