Facial Expression Recognition Based On Deep Learning

Facial expressions are the expression of human inner emotions and an important point of the communication between people.They are widely used in such as intelligent driving,human-machine interaction,and medical assistance.This also makes research on facial expression recognition of great practical significance.Traditional facial expression recognition methods are often affected by objective factors during the feature extraction stage,which can affect the quality of features.The emergence of deep learning has largely solved this problem.Neural networks can obtain higher dimensional feature representations by using nonlinear mapping methods.Therefore,compared to traditional methods,deep learning based methods have relatively stronger capabilities in feature extraction and generalization.However,traditional neural network models often have deep layers,large-scale parameter quantities,and computational complexity,resulting in a huge computational cost for model training.Moreover,feature reuse structures are often added to deep layers networks.If input features are not processed,it will lead to the accumulation of redundant features during the training process,This affects the quality of features and the recognition accuracy of the model.At the same time,the uncertainty and diversity of facial expression features can lead to problems such as feature loss and low feature extraction rate in the feature extraction stage of the model.Therefore,in response to the above issues,this article mainly conducts experiments from two aspects,and the main content is as follows:(1)Traditional deep learning methods often have problems with large parameter quantities and computational complexity.This article intends to use lightweight neural network models to solve this problem,but most lightweight models use depth wise separable convolutions.Using this convolution can indeed reduce parameter quantities,but it can also lead to a decrease in model recognition accuracy,Therefore,this article proposes a lightweight facial expression recognition model based on improved convolution to address the above two issues.The main work is as follows: using the lightweight network Shuffle Net V2 as the baseline network to ensure the lightweight properties of the method;Based on the blueprint separable convolutions and dilated convolutions,an improved convolution form has been designed and introduced to improve feature extraction efficiency and make it capable of multi-scale feature extraction tasks;Designed and introduced a shallow input feature processing layer to simplify the network structure and reduce computational complexity.Finally,this article also conducted experiments on two different datasets,Fer2013 and CK+,and the results showed the effectiveness of the proposed model.(2)In the feature extraction stage,due to the complexity and variability of facial expressions,the completeness of the extracted features is low,and there are many redundant features in deep level networks with feature reuse structures.This article proposes a Residual Multiscale Feature fusion Attentional Network.The main work is as follows: Design and introduce a multi-scale parallel feature extraction path based on the improved convolutional mode,to enrich feature information;Designed and introduced a feature screening module to reduce redundant features generated during model training,while screening out high-quality features to improve feature quality;Introducing channel attention mechanism to highlight local key feature information;Finally,SMU activation function is introduced to improve the nonlinear capability of the model.For the proposed model,this article conducted experiments on the Fer2013 dataset and the CK+dataset,respectively.The experimental results show that the model can achieve better recognition accuracy than other cutting-edge methods on the Fer2013 dataset and the CK+dataset,while ensuring lower parameter quantities and computational costs,and has better robustness.