GAMF: Collaborative Salient Object Detection Based on Global Attention and Multi‐Scale Fusion

Co‐saliency detection (CoSOD) aims to identify common salient regions from a group of related images. However, existing CoSOD methods often suffer from insufficient cross‐scale feature fusion and limited discriminative power of feature representations. To address these issues, we propose a novel collaborative salient object detection framework termed global attention and multi‐scale fusion (GAMF). The proposed framework adopts a three‐stage encoder–collaborative modeling–decoder architecture and incorporates two key components: the cross‐scale feature fusion (CSFF) module and the convolutional block attention module (CBAM).The CSFF module integrates high‐level semantic features with low‐level detail features through up‐sampling, concatenation and channel compression, while introducing a global context‐aware channel attention mechanism to enhance feature representation. This design improves boundary preservation and the detection of small salient objects while reducing redundancy introduced by naive multi‐level feature aggregation. Meanwhile, the CBAM module exploits both channel and spatial attention to refine feature representations and highlight salient spatial regions. Extensive experiments on three benchmark datasets, including CoCA, CoSal2015 and CoSOD3k, demonstrate that the proposed GAMF framework consistently achieves superior performance across multiple evaluation metrics, verifying its effectiveness, robustness and strong generalization capability.