A multiscale based approach for automatic shadow detection and removal in natural images

Abstract

Shadow is a natural phenomenon observed in most natural images. It can reveal information about the objects shape as well as the illumination direction. In computer vision algorithms, shadow can affect negatively image segmentation results, feature extraction, or object tracking. For that, it is necessary to detect and eliminate shadow. Texture remains the best feature used to detect the shadow and photometric information can be used to eliminate it. However, in case of an image with a shadow projected on a complex texture, most of the proposed approaches in literature are useless. In this study, we propose an automatic and data-driven approach for shadow detection and elimination based on the Bidimensional Empirical Mode Decomposition (BEMD). The main idea is to decompose the shaded image into intrinsic components (IMF) that contains only texture and a residue with only objects shape. Then, shadow detection is performed on the IMFs by matching the pair of segmented regions using texture features, while elimination is carried out via a Gaussian approximation applied only on the residue. Finally, the shadow-free image is obtained by adding all the IMFs and the shadow-free residue. The proposed approach is evaluated in comparison with recent approaches on images with the different type of shadow.