Research On Face Recognition Based On Local Visual Model

Human face recognition technology is still going to be unfolding as a research hotspot of pattern recognition and machine visual territory, which includes the human face detective, face recognition and expression recognition etc. Because the human face recognition technology has extensive application prospect in many domains such as justice, security and military, it has been generally paid close attention to by many researchers. Therefore, the human face recognition technology has achieved great development in recent years. Also, a series of innovation algorithms had been proposed and some preliminary commercial archetype systems had been put into market. Above all the algorithms, the methods based on local visual feature model has gradually become one of the main methods because of its advantages in the recognition for the figures which have great change of illumination or other adverse factors. Based on the previously researches, we try to aim at the robustness of illumination condition and other adverse factors. Start with the local features, we do some discusses in image preprocess, feature extraction and multi-feature fusion recognition. The main tasks we had done in this paper as follows:(1) Image preprocess is an important previous work of human face recognition. It can decrease all kinds of adverse factors efficiently and seriously influence the face recognition results. In this paper, besides the normalization of orientation, size etc, we principally proposed a preprocess algorithm aims to illumination changes. Our method includes the Gamma regulation, difference of Gaussian and contrast balance. The experiment result shows, our method can decrease the influence of illumination, raise the contrast and strengthen the effective feature areas of human face. Our preprocess methods is effective facilitate the following feature extraction and face recognition.(2) Feature extraction is the key task of human face recognition. In this paper, we decide to extract multifarious local visual features of human face. After the analysis of the advantages and disadvantages of traditional algorithms, we do some improvement in LBP texture features firstly based on previously method. The improved LBP (ILBP) feature means that we propose a Dipole compare operator instead the simple difference operation of middle pixel and the 8 pixels border upon. The ILBP algorithm can extract the local features of bigger area in different orientations and scales. And then, we take the advantages of Gabor feature: multi-orientation and multi-scale. Based on this method, we extract the ILBP features on the figures which had already been extracted the Gabor features. We call it ILGBP feature.In another hand, the human faces we get from the practical use condition always influenced by the changes of illumination, scales and other adverse factors. Yet, the SIFT algorithm has the advantages of translation invariance, rotational invariance and scale transform invariance. Therefore, we use both the ILGBP feature and SIFT feature to describe a human face.(3) We use a method that based on D-S evidence theory and Fisherface algorithm to do the human face recognition based on multi-feature fusion. Because the D-S evidence theory is a method of inosculation on the level of strategic decision, in our method, we use the Fisherface algorithm to do the face recognition on ILGBP feature and SIFT feature respectively firstly. After the Fisherface feature contrast, we will get two similarity values which could be transformed to trust values use the regulations of D-S evidence theory. At last, we can successfully get the recognize results by the means of trust values. Because of the property of Fisherface algorithm, our method can decrease the feature dimension efficiently. What's more, it has stronger robustness.The experiment results in MATLAB have told that our method has strong recognized capacity for the low-quality human faces.