GILNet: Grouping interaction learning network for lightweight salient object detection

Recently, salient object detection (SOD) has achieved significant progress with the rapid development of convolutional neural networks (CNNs). However, the enhancement of SOD accuracy often comes at the cost of increased network size and computational complexity, hindering the application of existing SOD methods to lightweight devices, particularly robot devices. To address this issue, we propose a lightweight grouping interaction learning network (GILNet) for efficient and effective multi-level feature learning and shared feature aggregation. Specifically, our novel grouping interaction learning module (GIL) enables efficient feature extraction, and based on this module, we construct a lightweight backbone network to extract multi-scale features. Furthermore, we introduce a shared feature aggregation (SFA) module to aggregate these features in a shared manner, and a progressive guidance prediction (PGP) module to gradually refine the saliency predictions. Extensive experiments on five popular benchmarks demonstrate that GILNet yields comparable accuracy with state-of-the-art methods. More importantly, GILNet operates at a GPU speed of 345 frames/s with only 1.21M parameters, representing a significant reduction in computational cost and model size. These results highlight the significance of our method in achieving a better trade-off between accuracy and efficiency.