Research On Robust Face Recognition With Occlusion And Complex Illumination

Face recognition is an important research project, which has been widely applied to security and identity authentication fields. Although face recognition has achieved great progress, recognition in unrestricted environment is still a tough problem. Many factors such as occlusion,and vary illumination can result in poor performance of face recognition system. In order to solve these problems, some contributions have been done as follows:To the occlusion problem, a face recognition method with fusion of occlusion detection and collaborative representation based on HOG feature was proposed. Firstly, an occlusion detector based on HOG feature, principal component analysis and support vector machine was designed, which could detect the occlusion on the local patches of the face images that were divided according to common occlusion distribution. Then, based on the results of occlusion detection, the HOG feature of the non-occluded face regions was extracted and collaborative representation was adopted for classification. Finally, experiments were conducted on AR and Yale face databases. The average recognition rate of face images with occlusion in AR face database reached 95.2%. The recognition rates of face images with illumination and expression changes in AR and Yale face databases reached 97.3% and 98.6%respectively. The calculation speed of the proposed method was about 293 times faster than SRC. Experimental results demonstrate that the proposed method yields significant performance improvements compared to other methods.To the illumination problem, a near-infrared(NIR) face recognition method based on local binary pattern(LBP) and collaborative representation was proposed. Active near-infrared combined with optical filter were used to acquire NIR face images that have nothing to do with environmental illumination. So the impact of illumination can be avoided. The lightness of NIR face images changes with distance between face and camera, so LBP feature was adapted to eliminate the single change of lightness. Finally, collaborative representation was adopted for classification. Experiments were conducted on PolyU-NIRFD NIR face databases and the recognition rate reached 96.8%. Experimental results demonstrate that the proposed method yields significant performance improvements compared to other methods on NIR face recognition.