Salient Object Detection in Complex Backgrounds Guided by High-Level Feature Aggregation

Salient object detection is essential for many computer vision tasks and aims to detect the most prominent objects in images. However, existing methods often perform poorly when dealing with complex scenes. To overcome these issues, we propose a novel salient object detection network. A Global Feature Enhancement Module (GEM) is designed to solve the problem of inaccurate object localization in complex scenes. GEM enhances the importance of high-level semantic information in the network, using high-level features to guide lower-level features. Additionally, we introduced a Local Feature Refinement Module (LRM) designed to enhance the capture and optimization of low-level details, aggregating coarse saliency features generated by GEM with finer features from each decoding stage. By progressively refining features, the network effectively suppresses shallow background noise and achieves precise saliency maps. Comparative experiments on five public datasets with twelve advanced methods demonstrate the superior performance and efficiency of our proposed method.