Research On Facial Expression Recognition Algorithm Based On Deep Learning

Emotions are the most authentic responses of humans to external stimuli in natural environments.With the advancement of deep learning and neural network algorithms,facial expression recognition technology has made significant progress and has become a popular research direction in the field of AI.Facial expression recognition faces two main challenges.Firstly,inadequate feature extraction directly affects the accuracy of facial expression recognition.Secondly,facial expressions can be ambiguous,where similar expressions may convey different emotions in different backgrounds.To address problem(1),this thesis proposes a fusion attention with multi-scale residual network(AFMS-Res Net).Firstly,the structure of residual blocks in Res Net is optimized,and an efficient channel attention mechanism called ECA attention is introduced.The ECA attention mechanism selectively focuses on important facial features,accelerating the convergence speed of the network.Then,the output feature maps of the residual optimization module are fused together through concatenation after each Basic Block in the Res Net network,forming a multi-scale fusion.Finally,global average pooling is applied before the fully connected layer to reduce the number of parameters in the model and prevent overfitting.Experimental results on the CK+ and FER2013 test datasets demonstrate an accuracy of 97% and 71.97%,respectively,indicating superior performance of this model.To address problem(2),this thesis proposes a feature learning method that utilizes uncertainty information for comparison: mixture of attention mechanism and relatively uncertainty network(MARNet).By introducing the concept of relatively uncertain learning(RUL),which learns facial features and uncertainty,RUL designs a weighted cross-entropy loss function that uses the learned uncertainty as weights to mix facial features,effectively recognizing low-quality expressions in mixed features.This network can effectively mitigate the impact of blurry images and inaccurate photo labels.Experimental results on the CK+ and FER2013 test datasets demonstrate an accuracy of98.9 and 73.36%,respectively,indicating superior performance of this model.