Shadow Generation for Composite Image with Multi-level Feature Fusion
Tianyanshi Liu, Yuhang Li, Youdong Ding · 2022
Shadow generation for composite image aims to generate shadows for the foreground with reference to the background information, reducing the floating feeling of the foreground due to the lack of shadows caused by simple overlays. The existing two-stage methods are not effective in generating shadows for composite images with complex foregrounds. First, in the Shadow Mask Prediction stage, foreground information and background lighting information are not fully extracted, resulting in inaccurate prediction range and shape of foreground shadows. Second, in the Shadow Filling Stage, the amount of shadow parameter information will decrease as the residual network level deepens, and the shadow parameter prediction is inaccurate, which greatly affects the foreground shadow filling effect. Therefore, we add a multi-scale feature enhancement module to obtain a wider range of feature information and improve the mask prediction accuracy. Meanwhile, we propose a multi-level feature fusion module to reduce the loss of information in the process of shadow parameter prediction by multiplexing features at different levels. Our experiments on the public dataset DESOBA show that the method generates shadow regions with more accurate range and shape.