Unpaired Shadow Removal: Enhancing Attention to Shadow Areas via Dropkey

초록

Shadows appear in image areas where an object obstructs the light path. These areas, having lower values than non-shadow areas, degrade image quality and lead to issues in object recognition and segmentation. Supervised shadow removal depends on datasets with shadow and shadow-free image pairs, sparking increased interest in unpaired techniques. However, unpaired shadow removal faces challenges in accurately focusing on shadow areas without direct supervisory signals from ground truth. Consequently, shadows are not effectively removed, and non-shadow areas may suffer from unintended distortions. In this paper, we introduce a transformer-based network designed to identify shadow areas accurately, leveraging global context through spatial and channel attention mechanisms. Additionally, in the training phase, the transformer network is trained to precisely concentrate on shadow areas using a domain classifier and a dropkey mechanism, which randomly drops features of the keys to enhance focus. Our method is tested across several benchmark datasets for shadow removal, such as ISTD, ISTD+, SRD and WSRD, demonstrating better performance compared to current unpaired approaches.

키워드