Comparative studies on the state of the art feature extraction and classification techniques for human face recognition under low resolution problem, are proposed in this work. Also, the effect of applying resolution enhancement, using interpolation techniques, is evaluated. A gradient-based illumination insensitive preprocessing technique is proposed using the ratio between the gradient magnitude and the current intensity level of image which is insensitive against severe level of lighting effect. Also, a combination of multi-scale Weber analysis and enhanced DD-DT-CWT is demonstrated to have a noticeable stability versus illumination variation. Moreover, utilization of the illumination insensitive image descriptors on the preprocessed image leads to further robustness against lighting effect. The proposed block-based face analysis decreases the effect of occlusion by devoting different weights to the image subblocks, according to their discrimination power, in the score or decision level fusion. In addition, a hierarchical structure of global and block-based techniques is proposed to improve the recognition accuracy when different image degraded conditions occur. Complementary performance of global and local techniques leads to considerable improvement in the face recognition accuracy. Effectiveness of the proposed algorithms are evaluated on Extended Yale B, AR, CMU Multi-PIE, LFW, FERET and FRGC databases with large number of images under different degradation conditions. The experimental results show an improved performance under poor illumination, facial expression and, occluded images.
展开▼
