Towards Hard and Soft Shadow Removal via Dual-Branch Separation Network and Vision Transformer

Shadow removal is an important task in computer vision, as shadows in real-world scenes affect image color and brightness, making perception and texture recognition more challenging. Many existing methods may not adequately distinguish between hard and soft shadows, resulting in a lack of targeted processing for each type. To order to tackle this problem, we propose a dual-path model that separates the processing of hard and soft shadows using specifically designed loss functions. The model includes a shadow classification module to guide each shadow type through the appropriate path. By integrating a Vision Transformer with UNet++, our approach improves edge detail and feature fusion. On the ISTD dataset, our model achieves an RMSE of 2.905, outperforming state-of-the-art single-path methods.