Abstract:
Shadows are the artefacts which are present in nature and if present in images
they might be a source of hindrance for various Computer Vision tasks including object detection, tracking as well as segmentation, scene analysis and
many more. This research presents a novel idea for sake of detecting and removing shadows from a single RGB image. The presented methodology uses
deep-learning based model GANs, which consist of two generators and two
discriminators in order to detect and remove shadows. The shadow free image generated by GANs is further fed into the post processing step, which is
used to refine the shadow region via shadow mask. This post-processing step
leverages the power of traditional image processing techniques combined to gether such as histogram matching, custom filters, shadow boundaries detec tion and estimation. The proposed post-processing step is capable of refining
the shadow free image generated by GANs and hence can produce shadow
free images more efficiently. Moreover, a large-scale benchmark dataset is
also presented which consist of 5352 triplet images i.e., shadow, shadow mask, shadow-free samples and can be used for shadow detection as well
as for shadow removal purpose. The presented dataset “Extended ISTD”
covers a vast variety of dark/hard shadow images as well as multi-color con trast shadow images and is the extended version of already publicly available
“ISTD Dataset”. Deep-learning model GANs once trained on the presented
dataset and undergoing the proposed post processing step, an RMSE of 5.68
is achieved. The proposed methodology is capable of producing the shadow
free images efficiently, even in case of dark, hard shadows as well as multicolor contrast shadow images.
