Research On Lightweight Facial Expression Recognition System

With the rapid development of hardware systems,the performance of computers has greatly improved,which laid the foundation for the development of computer vision in artificial intelligence.Facial expression recognition has also become a current research hotspot and has been widely applied in areas such as human-machine interaction,education,safe driving,and food appreciation.Traditional facial expression recognition technology requires the knowledge of domain experts to manually design suitable feature extractors,and different step results will affect the final classification effect.It is difficult to adapt to the complexity and diversity of tasks.Deep learning-based facial expression recognition methods have become the mainstream research direction,and deepening the neural network structure has continuously improved the recognition accuracy.However,the high computation burden and long training time make it impossible to deploy related algorithms to low-computing edge devices.In order to reduce the parameter and computing burden of neural networks so that they can be deployed on low-computing devices,research has been conducted on lightweight convolutional neural networks and facial expression recognition methods,and a lightweight facial expression recognition system has been implemented.The main work of the paper is as follows:First of all,inspired by Ghost Net and Shuffle Net,the original Ghost Net module was modified by adding channel shuffling operation before output and removing the attention mechanism module that Ghost Net originally had,leading to an improvement in the bottlenecks of Ghost.The input size and model structure of the original Ghost Net model were also modified,and global average pooling was used to replace the fully connected layers in order to reduce the number of parameters and computations,making it more suitable for facial expression recognition tasks.Based on the above two points,an improved model named S-Ghost Net was proposed.Experiments were conducted on the Fer2013 and CK+ datasets,achieving accuracies of 81.238% and 96.447%,respectively.Compared with other mainstream lightweight convolutional neural networks,S-Ghost Net has less model parameters and computations while still maintaining high accuracy,making it more suitable for facial expression recognition tasks.Secondly,in order to study the performance of S-Ghost Net in facial expression intensity analysis,a three-level intensity classification method was proposed to address the problem of the lack of dataset labels in facial expression intensity analysis tasks.This method was used to classify angry expressions in the CK dataset into three levels of intensity and create an angry expression intensity dataset.This dataset was then used to train the S-Ghost Net model,enabling the model to perform facial expression intensity analysis tasks based on different weight parameters.The experiment was carried out on the anger expression intensity data set,and the accuracy rate was 75.401%,and the recall rate of each classification was 0.97,0.57,and 0.94,respectively,which verified that S-Ghost Net is better than the traditional method for three-level intensity classification of expressions.Finally,based on existing research,a facial expression recognition system was designed and implemented using the Python language.The system has the capability to recognize static facial expressions through local images and dynamic facial expressions through the camera.Additionally,it can display the recognition results on the screen in real-time.Testing has shown that the system has a high accuracy for recognizing faces in a positive direction and is fast in recognition speed.