Shadow Detection Based On Graph Cuts For A Single Outdoor Image

Shadows always exist in images and videos, and some vision tasks such as imagesegmentation, object recognition are influenced seriously while important visual cues areprovided by shadows in other vision tasks such as light estimation, pose estimation, targetsize estimation, depth estimation, stereo matching. Thus, the shadow detection has becomean important field of computer vision research. In this paper, the research of shadowdetection of single outdoor image is carried out, based on image segmentation, edgedetection, the Support Vector Machine, graph cuts.Firstly, image segmentation, edge detection, Support Vector Machine and otherrelevant theoretical knowledge are summarized, in order to expand more in-depth study ofthe shadow detection.Secondly, according to the idea that taking effective use of image boundary and areainformation for shadow detection, which presented firstly in this paper，the Support VectorMachine is used to get the candidate shadow boundaries and the candidate shadow regionpairs, as the boundary and region information respectively. The boundary and regioninformation about shadows is considered as shadow information, to detect shadowsaccurately and continuously.Thirdly, the graph cuts theory is applied to the shadow detection process, and basedon the idea that taking comprehensive use of image boundary and region information, thedetected shadows by our method are more accurate and continuous. During the shadowdetection process, an energy function is built according to the boundary and regioninformation, and an undirected graph is constructed, which minimum cut corresponds tothe energy function’s minimum, and shadows are detected while the energy function’soptimal solution is obtained at the end.Finally, experiments are carried out to verify the feasibility and effectiveness of theproposed method and the idea that taking effective use of the boundary and regioninformation to detect shadows，and the experimental results are analyzed.