Facial Expression Recognition Based On AR-WLD And Convolutional Neural Network

With the continuous development of pattern recognition and human-computer interaction technology,face expression recognition research has become one of the hot research topics in the field of intelligent human-computer interaction.At present,the research of face expression recognition under lab-controlled environment has made great progress.However,under the real-world environment,there are many uncontrollable factors in the expression image,which makes the existing expression recognition algorithms perform poorly.In order to improve the robustness of the expression recognition system,this paper mainly studies the expression recognition from two aspects:(1)making different occlusion simulations on the existing public facial expression databases which are collected under controlled environment,and using traditional method to conduct relevant research;(2)combining the above traditional method with deep learning to further study the expression recognition based on the unconstrained spontaneous expression database,improve the classification performance and robustness of the algorithm.The main works are as follows:(1)Local occlusion will cause identification interference to facial expression recognition,to solve this problem,we propose a novel expression recognition algorithm:from the aspect of feature description,we propose an improved WLD feature extraction algorithm,i.e.asymmetric region Weber local descriptor(AR-WLD).Compared with WLD,AR-WLD can make more fully consideration on the contribution of each pixel within the neighborhood,and the sub-neighborhood partitioning method can help to extract local information better,and weaken the influence of noise,which enhance the robustness of the feature to local occlusion,noise,etc.In the classification process,dividing the expression image into non-overlapping blocks,and then extracting the AR-WLD histograms of each block.We adopt the information entropy to adaptively adjust the contribution of each block,and then use the weighted similarity summation as the criterion of final discrimination.The experimental results on JAFFE and CK databases show the effectiveness of the algorithm proposed in this paper.(2)In the real-world environment,in addition to partial occlusion,there are usually some other uncontrollable factors,such as uneven lighting,posture deflection,individual expression differences,poor image quality,age differences,and facial image stretch deformation,and so on.To solve the above problems,an expression recognition algorithm that combines AR-WLD differential excitation feature with convolutional neural networks is proposed in this paper.On the one hand,the differential excitation of AR-WLD can effectively extract the local salient patterns in an input image,which can powerfully represent the feature for texture,and it is robust to light and noise variations;on the other hand,convolutional neural network can effectively learn the high-level features of the image based on the distribution characteristics of the data itself;in this paper,we take the extracted feature image,which filtered by AR-WLD differential excitation from the input image,as the input of convolutional neural network,guide the convolutional neural network to focus on the local salient patterns of the image,avoiding the effects of some noise and light changes.The experiment results on the RAF database show the effectiveness of the algorithm.