Research On Expression Recognition Based On Deformable Convolution And Adaptive LTP

In the way of human interaction,facial expressions are an important way to convey human emotional information and intentions.The facial expression recognition technology can help computers observe,understand and provide corresponding feedback like humans.At present,facial expression recognition technology faces two problems: one is that feature extraction algorithms based on geometric or texture features are not sensitive to wrinkles,bumps,depressions and other subtle facial changes;the other is that existing algorithms cannot solve the problem in real life.The collected pictures have uneven lighting distribution and noise interference.Aiming at the above problems,this paper designs the facial action unit recognition algorithm of deformable convolutional network and the expression recognition algorithm of expert prior knowledge adaptive LTP respectively,and the feasibility and effectiveness of the designed algorithm are proved through experiments.The main research contents of this paper are as follows:(1)A facial action unit recognition algorithm AU-DCN based on deformable convolutional network is designed.First,the size and direction of the receptive field are adaptively expanded by adding a learnable offset parameter to the convolution kernel;then a modulation mechanism is introduced into the deformable convolution module,so that the deformable convolution model can not only adjust each sample The offset of the point,the weight coefficient of the offset can also be adjusted to distinguish whether the candidate area obtained is the area of interest to us;finally,through the R-CNN feature imitation,the Faster R-CNN feature and R-CNN are calculated The feature cosine similarity constrains the difference between these two features,so that the network always focuses on the region of interest and reduces the influence of irrelevant redundant context information.The results of two sets of experiments show that the F1-score mean value of the designed algorithm on the DISFA and CK+ data sets has increased by 2.6% and 2.4%,respectively.(2)Designed an expression recognition algorithm based on expert prior knowledge adaptive LTP.First,use multi-scale expression for the original input image,divide the image at different scales into several sub-regions,use the LTP algorithm to take the histogram feature of the gray value of each sub-region;then calculate the information entropy of each sub-region to establish According to the corresponding weighting coefficient,the histogram feature of each sub-region is multiplied by the weighting coefficient,and cascaded into a complete texture feature;finally,the facial region is divided into 8 groups according to the prior knowledge of experts,for each The group generates a mask and surrounds it with a bounding box as the bounding box label of the region.The original input picture,the mask and the complete texture feature are sent to the AU-DCN together as the network input.The results of the two sets of experiments show that the recognition rate of the designed algorithm on the Oulu-CASIA NIR&VIS data is increased by 8.4% compared with other algorithms,and F1-score is also significantly better than the other three algorithms.The ablation experiment verifies the effectiveness of each component in the adaptive LTP algorithm based on the prior knowledge of experts.