Semantic Synergy: A Cross-Scale Approach to Lightweight Salient Object Detection

Salient Object Detection (SOD) is a central task in computer vision, aiming to rapidly identify and highlight salient targets in complex scenes. Although deep learning methods have made significant advances in this field, they often require substantial computational resources, limiting their application on lightweight devices. To overcome this bottleneck, this paper proposes novel lightweight Salient Object Detection method called Semantic Synergy. Existing methods tend to overlook the intrinsic connections between multi-scale features, which restricts detection accuracy. The main motivation of this method is to achieve efficient and accurate detection of salient targets through cross-scale semantic information fusion. Semantic Synergy enhances the semantic association between local details and global context through an innovative cross-scale feature fusion strategy, understanding the performance of salient targets at different scales, thereby significantly improving the detection accuracy of salient targets while maintaining computational efficiency. It enables in-depth mining and effective integration of features at different levels. Using this method, we have constructed a new lightweight salient object detection framework named SSNet. Extensive experiments on popular benchmark datasets demonstrate that the proposed SSNet achieves 190 FPS on GPU with only 1.4M model parameters, outperforming advanced lightweight networks and even surpassing some heavyweight models.