Face recognition in real world environments is mainly affected by critical factors such as illumination variation, occlusion and small sample size. This paper proposes a robust preprocessing chain and robust feature extraction in order to handle these issues simultaneously. The proposed preprocessing chain exploits Difference of Gaussian (DoG) filtering as a bandpass filter to reduce the effects of aliasing, noise and shadows, and then exploits the gradient domain as an illumination insensitive measure. On the other hand, Linear Discriminant Analysis (LDA) is one of the most successful facial feature extraction techniques, but the recognition performance of LDA is dramatically decreased by the presence of occlusion and small sample size (SSS) problem. Therefore, it is necessary to develop a robust LDA algorithm in order to handle these cases. At this point, we propose to combine Robust Sparse Principal Component Analysis (RSPCA) and LDA (RSPCA+LDA). The RSPCA is performed first in order to reduce the dimension and to deal with outliers typically affecting sample images due to pixels that are corrupted by noise or occlusion. Then, LDA in low-dimensional subspaces can operate more effectively. Experimental results on three standard databases, namely, Extended Yale-B, AR and JAFFE confirm the effectiveness of the proposed method and the results are superior to well-known methods in the literature.
The HPE Cognitive Computing Toolkit (CCT) is an opensource modeling platform backed by Hewlett Packard Enterprise. CCT provides a domain-specific language designed for problems like vision modeling and deep learning. The CCT platform compiles programs written in this language to native graphics processor (GPU) code. Developing vision models in CCT is far simpler and more productive than writing GPU code directly, but without sacrificing the performance gains of GPU acceleration. This programming model scales to interesting problems like dense optic flow, anisotropic diffusion, and deep learning. CCT is particularly powerful when combining multiple state-of-the-art techniques in a single algorithm.
Understanding the depth order of surfaces in the natural world is one of the most fundamental operations of the visual systems of many species. Humans reliably perceive the depth order of visually adjacent surfaces when there is relative motion between them such that one surface appears or disappears behind another. We have adapted a computational model of primate vision that fits important classical and recent psychophysical data on ordinal depth from motion in order to develop a fast, robust, and reliable algorithm for determining the depth order of regions in natural scene video. The algorithm uses dense optic flow to delineate moving surfaces and their relative depth order with respect to the parts of the static environment. The algorithm categorizes surfaces according to whether they are emerging, disappearing, unoccluded, or doubly occluded. We have tested this algorithm on real video where pedestrians and cars sometimes go behind and sometimes in front of trees. Because the algorithm extracts surfaces and labels their depth order, it is suitable as a low-level pre-processing step for complex surveillance applications. Our implementation of the algorithm uses the open source HPE Cognitive Computing Toolkit and can be scaled to very large video streams.