Face Recognition Based On Fusing Global And Local Histograms Of Oriented Gradients Features

At present, face recognition is still a hot spot of research in the field of computer vision,pattern recognition and machine learning, and is one of the most challenging tasks. Despite the current face recognition has made great progress in its algorithms and practical applications, but there are still considerable difficulties under complex environment, such as pose, expression, illumination. So face recognition both in theory and practical application still have great prospects to be achieved.In recent years, because of the pretty good robustness to illumination, expressions and pose change, face representation methods based on local feature extraction are widely used in face recognition. Meanwhile, today a lot of researches have point out that global and local features are both crucial for face representation and recognition. With the widely used in object detection and excellent performance of Histograms of Oriented Gradients (HOG) , there has been some papers apply HOG for face recognition and achieved good results.In this paper we adopts HOG for face representation, study the effects of the various properties in detail,and puts forward new improvement methods.The main work and contribution of the research can be summarized as follows:(1)In connection with the method of face recognition based on regular grid HOG feature extraction, we systematically analysis the effects of various parameters Settings on the performance of face recognition. And to compare it with the two famous face local features representation method:Gabor wavelet and LBP for the pros and cons. Experimental results show that the optimistic HOG parameters, less dimension of HOG features outperform the LBP method in large illumination variation and time interval .in LBP time environment of face complex changes for better performance in the collection, meanwhile HOG descriptor has more advantage than Gabor wavelet in the time of feature extraction and the number of feature vector dimension. (2) To address the problems of the variation of illumination, acquisition device etc,Gamma correction, difference of gaussian(DoG) filter and contrast equalization are adopted to preprocess the images and our experiments prove its effectiveness. In order to solve the problems of computational complexity and small sample size, and further improve the face recognition rate, this paper presents a method of using PCA to obtain dimensionality reduction of HOG features and then LDA for global feature selection.On the other hand, the HOG features are extracted in the face image and then will be partitioned to patches, and sub-classifiers will be trained. Due to the small dimension of each local patches,FLD will be directly applied for training the sub-classifiers which can reserve details and discriminant information maximally. Finally, global features and local features will be combined by the strategy of weighted sum rules. The experimental results in large-scale FERET face database show that the face HOG local features can further restrain illumination variation, improve identification accuracy; And HOG global features can maintain global information, reduce the influence of local variation produced by time interval or facial expression change. Global features and local features complement each other, in our experiments ,we got the best known results currently in FERET face database based on using HOG features for face recognition.