Facial Expression Recognition Based On Improved Deep Learning

Since facial expression is the most effective non-verbal communication method,and can well expressed emotions,mentality and intentions,its recognition and detection technology has gradually attracted the attention of academia and industry.However,the accuracy and robustness of traditional facial-expression recognition algorithms are low,and the change of facial expressions is a dynamic process,the traditional algorithms are not effective in the field of real-time facial expression recognition.This thesis is based on the recognition and detection method of Convolutional Neural Networks algorithm,which is considered to be applied in facial expression recognition and detection.The research contents of this thesis mainly include:1.The images of facial expression need to be pre-processed by using deep learning algorithms to extract facial expression features.It also investigates the basic process of image data pre-processing and common algorithms for each process as well as the experimental effects of some algorithms on the expression library used in this experiment,then converts into image data with less calculation as network input of deep learning algorithm.2.Although there are many existing Convolutional Network models,there is no model suitable for image recognition in various fields.It needs to create a network model suitable for facial expression recognition in order to meet the needs.This thesis has optimized the structure of internet and internal part based on the Le Net-5.In order to solve the over-fitting problem of network models caused by different features,it has improved the level of batch normalization.It selects maximum and average overlap pooling to reduce the amount of data while fully extracting facial expression features.The accuracy and robustness of recognition are effectively improved.What's more,the tolerance to light,posture and obstructions are increased.3.A new Convolutional Neural Network recognition method based on the Inception module and Drop Block overfitting processing method is proposed.The improved Convolutional Neural Network and Inception_v1 Network module are used to construct a deep convolutional neural network.Combined with L2 regularization technology,Drop Block technology is used to replace previous Dropout overfitting processing technology.Thus,the training speed of the network model is effectively improved,andtherobustness of the model and the training accuracy of the model are enhance.4.Results has shown the practicality of the two improved network models when they are applied in real-time facial expression recognition systems.