Deep Learning based Shadow Removal in Single RGB Images

Abstract

This report presents a deep learning based model to remove shadow from single camera images. Since the inception of camera images, the quality of the images remained a huge concern. Shadow in the captured image is one of the crucial factors to decide the threshold of the quality of the image. So, it is inevitable to remove a shadow from an image to enhance quality. It not only improves the visibility of the image but is also necessary to retrieve any information from the image. Researchers initially used Digital Image Processing techniques for shadow removal. But in the 20th century alongwith many other advancements in science and technology, Deep Learning emerged asthe apple’s eye of high-end innovators across the globe. It has facilitated human beingsin many fields. Researchers did an amalgamation of Digital Image processing, Artificial Intelligence, Machine Learning, Deep Learning, and computer vision and proposed many solutions for Shadow Removal. This paper also explores the said techniques to implement shadow removal with the highest possible accuracy and precision to have a state-of-the-art algorithm. It predominantly illustrates usage of spatial attention layer of transformers for shadow removal process. It is one step process which unifies both steps into one. i.e. first is to identify the shadow and the second is to remove it. It is one stage network having capacity to directly eliminate shadow from image. It does not need separate shadow detection. Hence, it is adaptable to RGB images de-shadowing for shadows projected on different semantic regions. It consists of a series of Fourier transform residual blocks as well along with two-wheel joint spatial attention of transformer layer. The Image Shadow Triplets dataset(ISTD) is the dataset for this paper. There is an evaluation metric for all the results of different techniques. This project yields the best possible accuracy to remove the shadow from an input image