Shadow Removal Network with Error Map Prediction

The existence of shadows affects many computer vision tasks. Recovering image information in shadowed regions by removing shadows has been proven to improve the performance of various tasks. However, shadow removal is a challenging task. The complex presence pattern makes it difficult for us to com succinct pletely shadows. In this paper, we study the problem of shadow removal, which aims to obtain traceless shadow removal results and preserve the image information under the original shadow area. To more adapt to the shadow removal task in complex backgrounds flexibly, we propose an Error Map mechanism to pre-estimate the feature error generated by the fusion network and use it to guide the refinement work to obtain better shadow removal results. In addition, we redesigned the higher performance network structure. We evaluate the performance of our method on multiple large-scale shadow removal datasets, our method achieves better performance than the original scheme.