Image Matting Based on Deep Equilibrium Models

Recent works have shown that additional improvement stages can further enhance the performance of image matting models. Inspired by this, we propose deep equilibrium matting (DEQ-Matt), which improves the feature maps for infinite times to achieve optimal performance by using the deep equilibrium (DEQ) models. We further tailor a loss function to train the DEQ models on the image matting task. Besides, we propose to use saliency maps to guide the image matting models, because they can be automatically and reliably predicted. In experiments, our method outperforms state-of-the-art methods and is superior in both semantic estimation and detail processing. Furthermore, we observe an increasing trend in the model’s performance as the number of feature improvement steps approaches infinity, which supports the motivation of this paper. The code is available at https://github.com/XinshuangL/DEQ-Matt .