| 英文摘要 | As an important branch of biometrics, face recognition has broad application prospects in identification, video surveillance, human-computer interaction, cartoon, attendance, access control and so on. The main content of this thesis is composed of researches about some key algorithms of full-automatic face recognition, such as feature point location, illumination processing, feature extraction and dimension reduction. It can be summarized as follows: (1) Feature Point Location. First, since the location accuracy of different feature points by the original Active Shape Model (ASM) is inconsistent, a new two layer Active Shape Model is proposed. It first divides the feature points into inner contour points and outer contour points, and then locates them respectively to prevent the location of the inner contour points be affected by the outer contour points. Second, seeing that the original ASM is very sensitive to the initial position of each feature point, a new method for setting the initial position based on key points location and PCA is proposed. The experimental result shows that this method can improve the accuracy of initialization greatly and make ASM get rid of local optimum and converge faster. (2)Illumination Processing. First, because the Gaussian filter used by Retinex is isotropic, a new horizontal edge filter is proposed to substitute it. On the processed image, this new filter can highlight horizontal edges, which are more important than vertical ones for face recognition. Thus, it can improve the performance of Retinex for processing illumination. Second, a new method fusing different illumination pretreatment algorithms is proposed. It fuses the normalized images gotten by different pretreatments into a new one to preserve the advantages of those pretreatments. (3)Feature Extraction. First, we propose a new probabilistic Local Binary Pattern (LBP) operator. It uses both the amplitude and sign of the difference between the center and its neighborhood to depict the occupancy probability of each pattern, thereby the robustness of LBP to illumination and noise can be further enhanced. Second, we present a new method based on LBP and a single image. In this method, edge image is used for extracting LBP histogram and can obtain better result than gray image. Besides, elastic matching is suggested to calculate the similarity. It can make this algorithm more robust to the variances brought by translation, rotation, expression and so on. (4)Sub... |
修改评论