Deep disentangling learning for real-world image enlightening and restoration

Abstract

Shadow removal is a vital image processing operation that can enlighten the illumination of shadow regions in an image. This application can elevate the accuracy and robustness of high-level computer vision tasks especially those which are heavily deep learning (DL) based (e.g., object detection, person surveillance, and vehicle tracking). Implementation of shadow removal on image or video data can mitigate the risk of unexpected fallouts in computer vision algorithms.

In this Final Year Project (FYP), the prime goal is to devise a potent DL-based shadow removal algorithm that can effectively remove shadows in images without leaving any boundary trace. A comprehensive ablation study is done to investigate the effectiveness of loss functions and network modules in our proposed architecture. Quantitative and qualitative analysis show that not only our proposed model achieving comparable performance in the removal of shadow, but the final output images also have the best reconstruction image quality among the other existing shadow removal methods.