LWGSS: Light-Weight Green Spill Suppression for Green Screen Matting

Green spill is a significant challenge in green screen matting and it affects the overall visual quality of the images. Due to the limitations of user interaction and the lack of specific datasets containing green spill information, it is difficult to generate high-quality foreground images automatically in green screen matting. In this paper, we propose a light-weight green spill suppression method for green screen matting, named as LWGSS, to generate high-quality foreground images without green spill in a concise way. Specifically, we adopt an anomaly detector to detect the anomalous pixels with green spill, and normalize these pixels through the spill removal module. With the help of the green spill suppression stage, our method can break the limitations of user interaction and specific datasets. Additionally, we switch the learning objectives from predicting the alpha mattes of foregrounds to predict the alpha mattes of background by using label inversion, which enables us to address all kinds of objects. Furthermore, we present a light-weight network with the smallest model size and parameters to further improve the inference speed. Extensive experiments on Composition-1k and Distinctions-646 demonstrate the superiority of our LWGSS.