Score level fusion method based on multiple oblique gradient operators for face recognition

The traditional gradient is calculated along y and x axes, and its main object is to obtain vertical and horizontal edges in local image patch. However, there exist a lot of oblique edges in image patch besides vertical and horizontal edges. In order to obtain these oblique edges, we introduce multiple oblique gradient operators. By performing the multiple oblique gradient operators, an image can be transformed into multiple gradient orientation images which display different spatial locality and orientation properties. Each class orientation images are normalized by “ z-score ” method over the corresponding orientation image set. Then, linear discriminant analysis is performed to extract the corresponding low-dimensional features, and the results are used to compute the corresponding distance scores. The different weighted coefficients are calculated for different orientation images based on training samples. A weighted score fusion method is used to combine different distance scores according to their respective salience. Experimental results show that our methods significantly outperform popular methods, and achieve state-of-the-art performance for difficult problems such as illumination and occlusion-robust face recognition.