Research On Occlusion Face Recognition Based On Image Segmentation

Compared to other biometric identification technologies, face recognition has obvious advantages. Face recognition system has a wide range of applications in identification, self-help service, information security and video surveillance. In ideal conditions, the existing face recognition systems have good robustness. But, if test images are affected by uncontrolled factors, including obvious light and posture changes, sunglass occlusion etc, performance of the recognition system will degrade abruptly. Over the last two decades, the research of face recognition under uncontrolled scene have been a hotspot in the field of pattern recognition.The existence of occlusion area among whole face image greatly reduces the coding ability of classification algorithm. Therefore, this thesis mainly studied face recognition with occlusion based on removing outliers area. Extraction method of face recognition with occlusion based on segmentation algorithm is firstly studied. Then, because illumination imbalance lead to the inefficiency of segmentation algorithm about removing the occlusion, two kinds of new image preprocessing algorithm are proposed to enhance the contour information of occlusion area. Thus, the performance of the detection ability of segmentation algorithm can further improve. The main works can be described as follows:(1) Face recognition algorithms based on the sparse coding and linear regression representation were studied. In order to lay a theoretical basis for the following algorithm, the corresponding classification algorithms are used to simulate on face images in this paper. Firstly, the fundamental theory and specific algorithms about feature extraction methods, sparse coding and linear regression representation are comprehensively introduced. Secondly, we use the global or local feature extraction algorithms based on corresponding classification algorithms to simulate on face images without occlusion, and compare the performance of classification algorithm. Finally, we discuss the recognition performance of whole occlusion face and partition blocks of occlusion image based on using different classifiers, and analyze the recognition performance of adopting different block patterns.(2) Face recognition methods with occlusion based on removing outliers area were discussed. Aiming to the issue of face recognition with partial occlusion in this paper, because the image segmentation theory can remove occlusion area effectively,an improved face recognition method based on removing the outlier area was proposed. The key of the algorithm is to look for the segment image that can effectively distinguish the occlusion areas from the whole face. In this paper, we simultaneously assume that the images in occlusion areas and outside the occlusion have obvious separability. A mean face image is firstly obtained from train images, which is subtracted by the test face to form an error face image. The extraction process of the error image is simple and rapid, and the subsequent classification algorithm is no longer limited. Compared to the similar algorithms, the proposed method has considerable recognition performance improvement with relatively low computational complexity. But when test image are influenced by uneven illumination, the performance of classification algorithm need further improve.(3) Ilumination equalization preprocessing method for face image based on Low rank decomposition was proposed. When the test images are influenced by uneven illumination, only using the error image mentioned above is so sensitive to uneven illumination that degrades the performance of the segmentation methods. Considering that low-rank matrix decomposition algorithm has good eliminating effect to different components of face images, the illumination equalization preprocessing for face image based on low-rank decomposition was discussed in the thesis. In order to remove the interference factors, the data matrix generally is generated by combining the interference image with clean image together. However, the ultimate goal of our research is to remove the uneven illumination and get clear occlusion contour. The data of the low-rank matrix recovery is constituted by a set of extended test images. Thus, an illumination equalization preprocessing for face image is proposed in this paper. Firstly, a set of extended test images are generated by considering the fact that the face image has symmetry. And then, the vectorized test face image set is formed a matrix, from which the illumination equalized test face image is extracted by low rank decomposition technique for the following recognition. Simulation experimental results demonstrate that the proposed method could efficiently improve the performance of detecting the bilateral symmetry occlusion of face, and thus promote considerably recognition rate.