Research On Face Recognition Based On Improved Maximum Margin Criterion

Facial recognition is a very important research subject in pattern recognition. One of the basic issues in pattern recognition involves feature extraction. In facial recognition, the key point is whether the distinguished feature of face image can be extracted. Based on linear subspace, feature extraction becomes a popular way. This paper will do a deep study on this, mainly focusing on the following aspects:(1)The pretreatment of face image is a prerequisite to feature extraction and identification. Image enhancement on original image is an essential part in pretreatment. As no consideration on image fuzzy enhancement and a simple image changing on contrast or noise control, the traditional image enhancement ---- airspace law(including neighborhood averaging and mean filter method, etc) often cannot reach a better enhancement effect because it weakens some image details when controlling noises. Thus this paper adopts an image fuzzy enhancement law, of which the results of emulation research shows that this law works out well in both noise control and image details.(2) This paper gives an introduction to the linear subspace in facial feature recognition, especially emphasizing on the advantages and disadvantages of the linear discriminant analysis (LAD) and its improved law-----maximum margin criterion (MMC). On this basis, statistical uncorrelated and weighted maximum margin criterion (UWMMC) is put forward and applied to facial feature extraction. Experimented on ORL and Yale face databases, UWMMC proves itself superior to other laws.(3) As a feature extraction law, symmetrical maximum margin criterion (SMMC), is offered to the sample scarcity problem in face recognition. Face itself has mirror symmetry, so according to odd-even separation theorem it can be separated into odd symmetrical image and even symmetrical image. We can use maximum margin criterion law to extract odd and even symmetrical image separately and then put them into fushion for identification. Experiments on ORL and Yale face databases shows that SMMC which fuses odd and even extraction avoids the disadvantages of MMC and exceeds other laws.