Based On The Sparse Representation Of Facial Expression Recognition, Residuals Fusion

Facial expression recognition (FER) is about using computer to extract facial features of all expression images, and then classify the images to different expression categories according to those facial features. FER makes computer know the expression states from the classification result and achieve Human-Computer Natural Interaction. Although much progress has been made in FER in the past years, many methods need to be researched for purpose of increasing the recognition rate on with the influence of illumination, posture, noise, masking, and so on.In this paper, we give a survey of the common steps of FER and the methods applied in every step. Compressed sensing (CS) is introduced in this paper as well as the Sparse Representation Classification (SRC) algorithm. SRC is used in FER and experiments are carried out on facial expression database. The research work in this paper mainly includes the following.1. SRC is used in FER, and compared with 2DPCA+SVM in JAFFE facial expression database in FER depends on person. The result shows that SRC is efficient. Compared with 2DPCA+SRC, Curvelet+SRC and Random Projection+SRC, the result shows that the role of feature extraction is not so important as in common approach.2. SRC is used in facial expression recognition of the image added with noise. The variance of the white noise form 0.01 to 0.1. The experiment is carried out in JAFFE and compared with 2DPCA+SVM and Curvelet+SVM. The result shows that SRC is robust with noise and performs best in these methods.3. SRC is used in facial expression recognition of the image with blocks. The experiment is carried out in JAFFE and some part of an image of facial expression is replaced by an unrelated image. The replaced area is from 10% to 50%. SRC is compared with 2DPCA+SVM and Curvelet+SVM. The result shows that SRC is robust with block.4. A new approach for facial expression recognition based on fusion of sparse representation is proposed. The gray information is used in the method of SRC, and the texture information is used by LBP. These two methods could fusion by analyze the residuals. The result of the experiment in JAFFE shows that the recognition rate of the new fusion approach is 69.52 which is much high than 62.43% of SRC and 60.52% of LBP+SRC. This new approach also used in CK facial expression database. The result of the experiment shows that the recognition rate is 5% higher than SRC or LBP+SRC.