Semantic-guided Automatic Natural Image Matting with Trimap Generation Network and Light-weight Non-local Attention

• DOI: 10.48550/arxiv.2103.17020
• Publication Date: 2021-01-01
• Authors: Yuhongze Zhou; Liguang Zhou; Tin Lun Lam; Yangsheng Xu
• Keywords: Upsampling; Convolution (computer science)

Natural image matting aims to precisely separate foreground objects from background using alpha matte. Fully automatic natural without external annotation is challenging. Well-performed methods usually require accurate labor-intensive handcrafted trimap as extra input, while the performance of generation method dilating segmentation fluctuates with quality. Therefore, we argue that how handle trade-off additional information input a major issue in matting. This paper presents semantic-guided pipeline Trimap Generation Network and light-weight non-local attention, which does not need input. Specifically, guided by segmentation, estimates trimap. Then, estimated guidance, our Non-local Matting Refinement produces final matte, whose trimap-guided global aggregation attention block equipped stride downsampling convolution, reducing computation complexity promoting performance. Experimental results show algorithm has competitive state-of-the-art both trimap-free trimap-needed aspects.