Abstract | ||
---|---|---|
Automatically learning to focus on salient regions while suppressing irrelevant regions is very useful in some specific image segmentation tasks like medical diagnosis and automatic driving. In this paper, a fast semantic segmentation method based on attention gate and multi-layer fusion is proposed. In our model, an attention gate module with few model parameters is designed as a bridge between downsampling layer and corresponding upsampling one, which can highlight the features of foreground, and improve feature representations in semantic segmentation. In addition, a multi-layer fusion mechanism is proposed to integrate the semantic information from different downsampling layers, which can supplement the lost pixels by utilizing the semantic complementarity among different layers. Considering the real-time of segmentation, we use the lightweight model as the backbone network to extract feature. The proposed architecture makes a right trade-off between segmentation accuracy and efficiency on CamVid, VOC and Cityscapes datasets. Specifically, for a 512×512 input, we achieve 72.9% mean IOU on the CamVid test dataset with the speed of 43 FPS. |
Year | DOI | Venue |
---|---|---|
2022 | 10.1007/s11042-022-12519-6 | Multimedia Tools and Applications |
Keywords | DocType | Volume |
Image semantic segmentation, Multi-layer fusion, Attention gated mechanism | Journal | 81 |
Issue | ISSN | Citations |
15 | 1380-7501 | 0 |
PageRank | References | Authors |
0.34 | 4 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Yanping Tang | 1 | 0 | 0.68 |
Canlong Zhang | 2 | 2 | 2.15 |
Qinghe Cheng | 3 | 0 | 0.68 |
Zhixin Li | 4 | 12 | 19.62 |
Luyang Qian | 5 | 0 | 0.68 |