| With the rapid development of hardware system,the computing performance has been improving greatly in past decades,which lays a solid foundation for the expression recognition through deep neural network.Therefore,the network topology becomes deepen and widen continuously,which improves the recognition accuracy.However,the high computing burden limits their application in edge hardware and mobile devices,such as laptops and embedded hardware,etc.Thus,how to deploy the above algorithms to low computing platforms to realize the edge intelligence is an urgent problem,which would be studied in this thesis as follows.Owing to fewer trained parameters and faster running speed,the lightweight neural network is commonly used on low computing platforms,but cannot fully competent for various classification tasks.In order to improve the model accuracy with less parameter increment,the thesis proposes a weighted fusion neural network model with dual input based on the mini＿Xception,which maintains the original input while add a binary image to strengthen the facial muscle contour features.In fusion layer,the direct addition of tensors is replaced by an adaptive weighted one since the former one is not conductive to network learning due to its fixed feature extraction proportion from each channel,where the weights are adjusted by the back-propagation algorithm automatically.To balance the model accuracy and complexity,each residual module only adds two trainable parameters.Tests on Fer2013 and CK + data sets show that the accuracy is improved by 1.1% and 2%.Most public data sets are only static images,which cannot extract the spatiotemporal features effectively.The long short-term memory(LSTM)network takes into account the spatiotemporal image features but increases the difficulty of data sorting and the complexity of network.In order to solve the above problems,this thesis proposes a network model combination algorithm based on the probability optimization,which combines the outputs of two lightweight neural networks trained by static images via probability optimization to realize the accurate judgment in certain time window.The pain expression experiment verifies the algorithm feasibility in specific cases.As the application of above innovative algorithms,a complete abnormal expression monitoring system is also designed,which uses the TKinter library to design GUI,and realize a real-time tracking of human face.One may use it to realize the real-time monitoring and early warning of abnormal facial expression. |
