Research On Robust Feature Representation Of Face Image

Face recognition is a typical application and a challenging research topic in pattern recognition. Now existing face recognition methods may achieve good results under the controlled conditions, while many evaluations and practices demonstrate that the performances of existing face recognition systems decrease tremendously under the uncontrolled conditions (e.g., illumination variation, expression changes, pose and occlusion). Recently, the research in face recognition has focused on developing a face representation that is capable of capturing the relevant information in a manner which is invariant to facial expression, illumination, pose and occlusion.Effective feature representation is the key to improve the recognition result in face recognition. Based on analyzing the state of the art algorithms, this thesis attempts to research the robust face feature representation method under the uncontrolled condition. The main contributions of this thesis are summarized as follows:(1) Kernel principal component analysis based on Laplacian orientations.The traditional dimensionality reduction method based on pixel intensities very often fails to estimate reliably the low-dimensional subspace of a given data population. To address this problem, in this paper, we propose a kernel principal component analysis algorithm based on Laplacian orientations. The usage of the Laplacian orientations results in a more robust dissimilarity measurement between images. In addition, the explicit mapping is not considered in kernel dimension reduction method, so the low-dimensional principal subspace base is not calculated directly. In this paper, we introduce a robust cosine correlation measure, and then obtain an explicit nonlinear mapping through a mathematical derivation. Using the explicit mapping, the low-dimensional principal subspace base can be obtained finally. Experimental results show that the proposed algorithm significantly outperforms popular methods and achieves state-of-the-art performance for difficult problems such as expression, illumination and occlusion-robust face recognition. For a single sample per person, the proposed algorithm can also obtain a higher recognition rate.(2) Face feature representation based on difference local directional pattern.A face feature representation method based on difference local directional pattern (DLDP) is proposed. Firstly, each pixel of every facial image sub-block gains eight edge response values by convolving the local neighborhood with eight Kirsch masks, respectively. Then, the difference of each pair of neighboring edge response values is calculated and forms eight new difference directions. The top k difference response values are selected and the corresponding directional bits are set to1. The remaining (8-k) bits are set to0, thus we get the binary expression of a difference local direction pattern. In addition, high-resolution Kirsch masks only consider directions but ignored the weight values of each pixel location. DLDP propose a design method of weight values. Finally, the sub-histogram is calculated by accumulating the number of different DLDP of image blocks respectively. All sub-histograms of an image are concatenated into a new face descriptor. Experimental results show that DLDP achieves state-of-the-art performance for difficult problems such as expression, illumination and occlusion-robust face recognition in most cases. Especially, DLDP get better results for occlusion problem.(3) Face feature representation based on image decomposition.A face feature representation method based on image decomposition (FRID) is proposed. FRID first decomposes an image into a series of orientation sub-images by executing multiple orientations operator. Then, each orientation sub-image is decomposed into a real part image and an imaginary part image by applying Euler mapping operator. For each real and imaginary part image, FRID divides them into multiple non-overlapping local blocks. The real and imaginary part histograms are calculated by accumulating the number of different values of image blocks respectively. All the real and imaginary part histograms of an image are concatenated into a super-vector. Finally, the dimensionality of the super-vector is reduced by linear discriminant analysis to yield a low-dimensional, compact, and discriminative representation. Experimental results show that FRID achieves better results in comparison with state-of-the-art methods, and is the most stable method.