Research On Robust Facial Expression Recognition Method Based On Computer Vision Under Complex Conditions

Facial expressions contain rich emotional information,and expression recognition has broad application prospects in the fields of human-computer interaction,smart medical care,and psychoanalysis.At present,facial expression recognition is mainly based on images collected in a laboratory standard environment.Actually,it is inevitable to be interfered by complicated factors such as illumination and occlusion,resulting in a decrease in the correctness and robustness of expression recognition.Therefore,the research on expression recognition under complex conditions has become a research hotspot in the field of computer vision applications.In order to overcome the effects of illumination and occlusion on facial expression recognition under complex conditions,this paper has conducted in-depth research from the following three aspects,and the work is summarized as follows:Firstly,under the simple conditions,the basic research of expression recognition is carried out,and a strategy based on multi-feature fusion of ULBP and HOG is proposed.Based on the preprocessing of face detection,image graying and histogram equalization,the three different single LBP operators are analyzed and tested,and the ULBP operator with better recognition accuracy and recognition time is used.Aiming at the shortcomings of local texture and edge information of face expression area in a single LBP operator,a multi-feature fusion strategy based on ULBP and HOG is proposed by combining the local and edge feature extraction with superior HOG features.Different benchmarks are used on the JAFFE expression database for comparison verification.Secondly,under the illumination condition,the influence of illumination unevenness on expression recognition is studied and an adaptive threshold local Gabor binary mode LGBP composite feature algorithm is proposed.By adopting the combination of frequency domain and spatial domain,the Gabor wavelet is used to extract the four-scale and five-direction features of the face illumination image,which weakens the influence of illumination on the face image and reduces the noise of the texture image.Then,the feature extraction of the Gabor processed image is performed by the LBP operator.Subsequently,for the uneven illumination,the fixed threshold LBP operator can not accurately reflect the brightness of the pixel.The adaptive threshold LBP operator and its threshold selection algorithm are proposed.Further,the facial region was divided into three parts,and the entropy weight was used to highlight the mouth,eyes and other areas that reflected expressions.It has been verified on the CAS-PEAL image library of the Chinese Academy of Sciences,which improves the illumination robustness of expression recognition.Finally,under the occlusion condition,the influence of local occlusion on expression recognition is researched,and an improved cross-connected multi-layer LeNet-5 convolutional neural network model is proposed.In order to reduce the robustness of the traditional machine learning method and the poor recognition rate due to the lack of image information and noise interference under occlusion conditions,according to the advantages of deep learning in feature extraction,based on the LeNet-5 model,a convolutional layer and a pooling layer are added,and the low-level features extracted from the network structure are combined with the highlevel features to construct a classifier.By using the trainable convolution kernel to extract implicit features,the pooled layer is used to reduce the extracted implicit features.At last,the Softmax classifier is used for classification and recognition.The contrast experiment between the fixed occlusion and the random occlusion is carried out to verify the occlusion robustness of the improved method.