Non-bias self-attention learning for weakly supervised semantic segmentation. (January 2023)

• Record Type: Journal Article
• Authors: Sun, Wanchun; Feng, Xin; Liu, Jingyao
• Is Part Of: Computers & electrical engineering. Volume 105(2023)
• Publication Date: 2023-01
• Subjects: Non-bias learning -- Self-attention learning -- Weakly supervised -- Semantic segmentation; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108496
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: · We introduce a novel weakly supervised segmentation method that a non-bias layer is designed to explore the object region in the network training process, overcoming the limitation of the class activation map (CAM) region. · Relying on the powerful graph relevance of the graph convolutional network (GCN), an information enhancement module (IEM) is built to explore semantic relationships between images class. · Compared to other state-of-the-art methods on the most popular benchmark, PASCAL VOC 2012, our method achieves the best performance. Abstract: The weakly supervised semantic segmentation (WSSS) methods for image-level labels are among the most popular research topics. The current WSSS methods mainly generate the class activation map (CAM) regions so as to create the pseudo segmentation label seed. However, the sparsity of seed leads to less accurate local discriminative regions. Therefore, in this paper a non-bias self-attention learning segmentation network (NBSA) is proposed. First, the non-bias layer is designed to guide the network to expand the discriminative field of CAM during the training process. Second, a fine-grained learning strategy—the information enhancement module (IEM) is introduced to construct the graph convolutional network (GCN) for inter-semantic self-attention learning, which can further improve the generalization ability of the model. Experiments were show that on the PASCAL VOC 2012 dataset to compare the performances of our method with others, the results demonstrates that our method achieves the new state-of-the-art performance. Graphical abstract: Image, graphical abstract