Micro-expression Recognition Based On Diffusion Model And Frequency Domain Feature Extraction

Microexpression is known for its subtlety and spontaneity.It often reflects a person’s true emotions,revealing the real feelings that people try to hide.Therefore,they have important applications in fields such as lie detection,criminal investigation,medicine and criminal psychology.However,micro-expression recognition also faces many difficulties such as small data sets,difficult feature extraction and high uncertainty.These properties make the task of micro-expression recognition(MER)a challenge.This thesis focus on innovative research and experiments on micro-expression recognition based on deep learning methods from two perspectives.The details include:(1)This thesis proposes a micro-expression diffusion model(MDM)based on the implicit diffusion model(DDIM)to address the problem of small microexpression data samples.The proposed cross attention U-Net(CAU-Net)is used to replace the original U-Net,making the generated images more stable and clear.By training with existing micro-expression images data,a large amount of microexpression data can be generated.In addition,through the designed fusion sampling process(FSP),the model can be combined more conveniently with other classification models.The problem of small micro-expression samples is solved by adding a large number of micro-expression data generated by MDM.Experimental comparisons were conducted using several existing methods,which proved that the proposed method can effectively improve the accuracy of micro-expression recognition.(2)In response to the difficulty in extracting micro-expression features,this thesis proposes an innovative phase driven Transformer(PDT)network to complete the MER task.Specifically,amplitude and phase information are generated by two generator networks,and then fused together using a fusion module for subsequent network training.By incorporating image features in the frequency domain,the richness and diversity of features are better increased,enabling the model to extract more effective information and solve the problem of unclear micro-expression features.To address the issue of high uncertainty caused by small data sample sizes,this thesis uses dense relative localization loss to increase model perception field and enhance cognitive ability for improved robustness.The experiments were conducted on three public datasets: SMIC,SAMM and CASME II.The unweighted average F1-score(UF1)and unweighted average recall(UAR)is 83.68% and 84.61% on the dataset that combined the three datasets(3DB-combined)and there are also a little improvement on single dataset.The results show that our proposed method outperforms other methods.