Lightweight Image Matting via Efficient Non-Local Guidance

Zhaoxiang Kang, Zonglin Li, Qinglin Liu, Yuhe Zhu, Hongfei Zhou, Shengping Zhang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2884-2900

Natural image matting aims to estimate the opacity of foreground objects. Most existing approaches involve prohibitive parameters, daunting computational complexity, and redundant dependency. In this paper, we propose a lightweight matting method termed LiteMatting, which learns the local smoothness of color space and affinities between neighboring pixels to estimate the alpha mattes. Specifically, a modified mobile block is adopted to construct an encoder-decoder framework, which reduces parameters while retaining sufficient spatial and channel information. In addition, a Long-Short Range Pyramid Pooling Module (LSRPPM) is introduced to extend the reception field by capturing long-range dependency between regions distributed discretely. Finally, an Efficient Non-Local Block (ENB) is presented for guiding high-level semantics propagation from low-level detail features to refine the alpha mattes. Extensive experiments demonstrate that our method achieves a favorable trade-off between accuracy and efficiency. Compared with most state-of-the-art approaches, our method attains an immense descent in parameters and FLOPs with 30% and 13%, respectively, while achieving an improvement of over 15% in SAD metrics. Code and model are available at https://github.com/kzx2018/LiteMatting.