Global Information Progressive Aggregation Network for Lightweight Salient Object Detection
Junwen Li, Hongying Zhang, Bin Han, Hanyu Liu · 2022 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) · 2022
Most of the research on salient object detection pursues performance but ignores efficiency, resulting in poor practicability. In this paper, we propose a lightweight SOD solution, named GIPANet, which better alleviates the contradiction between model size and performance. Firstly, to solve the problems of shallow depth and insufficient information extraction in lightweight feature extraction networks, a Global Pooling Aggregation Module (GPAM) is constructed. Secondly, a Feature Aggregation Enhancement Module (FAEM) is constructed to fuse features more efficiently in each layer. Then, we use a hybrid loss function that fuses Binary Cross Entropy (BCE), Intersection-over-Union (IoU) and Progressive Self-Guided (PSG) losses to efficiently locate and segment salient objects. Finally, the experimental results on five public benchmark datasets show that the proposed network reaches a running speed of 125fps on a single GTX 1080Ti GPU for 352×352 images and uses only 3.68MB parameters to achieve an equivalent or even better performance than current state-of-the-art methods.