The Research Of Facial Expression Recognition Under Partial Occlusion Based On Local Features

Facial expression recognition technology is a challenging cross-subject of physiology, psychology, image processing, pattern recognition, computer vision and other fields. With the further study of facial expression recognition, the researchers found that sunglasses, masks, scarves and other decorations can cause facial occlusion. In this paper, the typical occlusion forms (eyes, mouth, left and right occlusion) firstly to expand the research. And then taking full account of the characteristics of randomness occlusion (occlusion can occur anywhere in the human face, the size and shape of the occlusion range are under unknown) situation, the expression recognition of random case is studied; the main contents of this paper are as follows:1. The facial expression recognition of typical occlusion form1) Facial expression image preprocessing and segmentation. Since the facial expression images in the database are slightly tilted head and sizes vary, that need to go through the preprocessing to eliminate these differences. By rotating the horizontal alignment of the eye and the distance between the eyes, a rectangular area containing only positive facial expression can be cut out of the original facial expression image. The facial expression images in two databases are normalized as 128×104, and using histogram equalization to increase the local contrast of images in some regions. For facial expression, the salient features between different expression classes mainly concentrated in the areas of eyes, nose and mouth. In order to better extract the features from the eyes, nose and mouth, and reduce the occlusion effect on facial expression recognition, the facial expression image will be segmented into 2 rows 3 columns altogether 6 regions in this paper.2) In view of the identification information loss problem caused by partial occlusion, proposing the local description based on the Weber (Weber Local Descriptor, WLD) histogram feature extraction method. WLD can simulate human perception to find features with significant change in the image, the significant micro-patterns are calculated by differential excitation and established significant patterns in the pixels along with these statistics on the gradient direction. The WLD descriptor employs the advantages of SIFT in computing the histogram using the gradient and its orientation, and those of LBP in computational efficiency and smaller support regions.3) For the extracted local facial expression feature and by considering the characters of the occlusion effect on the facial image, mainly concentrated in a contiguous area of the image, a greater impact in the region, and other areas affecting the minor, the method based on linear model based on the method of classifier fusion is proposed. Based on block decision strategy is another effective way to resolve occlusion, the main idea of strategy is to divide the facial expression image into several blocks and design a classifier for each block, finally the class of test facial expression will be got by fusing the output of each classifier.2. Facial expression recognition under the random occlusion caseRandom occlusion is characterized by occlusion can occur anywhere on the face, and occlusion size and the shape of the occlusion is unknown, there is no a priori knowledge about it. According to the randomness characteristic of occlusion and sparse representation of the original model assumes that actually said residual is follow the Gaussian distribution, this may not be enough to accurately describe the actual said residuals of the facial expression recognition system. Therefore, this paper proposes a face representation model based on robust regularization of coding to solve random occlusion cases of facial expression recognition.