Face Dectection And Emotion Recognition Based On Convolutional Neural Networks

Facial expression recognition not only represents human emotional information,but alos reflects the mind of human.It's very important in human-computer interaction and emotion analysis for researching and application.The traditional automatic facial expression recognition system should takes two steps to recognize the facial expression,one should detect the face firstly,the other should classify the emotion based on the first step.The two steps have repeated the features operation,so how to share the features of the two steps,and avoid repetitive computation has become a question.Due to the convolutional neural network have a good performace on slef-learning features,this paper designs and implements a multi-task network model based on convolution neural networks to solve this problem.Firstly,to research the performace on the image classification in way of CNN features,we using the Alex Net's features to instead of the traditional features method.And then use the softmax regression instead of training SVM classifier,through comparing the differences of network structure and the experiment results,the network configuration in depth,the characteristics of the convolution kernels filtering.Finally,in view of the traditional automatic face recognition systems will be face detection and face recognition in the practice of the two sub-systems separately,this paper proposes a joint based on convolution neural network method of face detection and face recognition,effectively avoid the traditional method in detection of human face feature extraction,also need to be in the original location of a face image feature extrac tion process.The experimental results show that the depth of the convolutional neural network,the higher the accuracy of the prediction;Increasing the number of filter convolution kernels can be appropriate to improve the accuracy of;Compared to contin uous convolution convolution with single layer,not only increased the nonlinearity,and to reduce the number of arguments.Shallow convolution features is beneficial to the location of the object orientation,and high order convolution characteristics is helpful to the classification of the target.