Deep Learning Based Shadow Detection in Images

Abstract

Various computer vision applications required Shadow detection and removal for example object tracking and recognition, scene interpretation, and video supervision. Since shadows have similar characteristics as that of the objects shadow pixels can be classified as part of object. This may cause problems such as merging or loss of object, alternation, and misinterpretation of object shape. To deal with this problem, we represent a deep learning based framework to automatically detect shadows in images. Our method learns many significant features automatically using supervised approach in Convolutional Neural Networks (ConvNets). The approach also makes use of drop out to improve results. The proposed methodology is tested on SBU dataset and results are promising.