An Inter-layer Multi-scale Feature Interconstruction Network for Small Target Detection in Underwater Sonar Images

Feature fusion networks provide a more comprehensive data representation by fusing features from different layers, improving the robustness of underwater small target detection. However, fusion can introduce over-smoothing and redundancy that mask weak target patterns and lead to feature loss. To address these challenges, a novel inter-layer multi-scale feature interconstruction network (IMFI-Net) is introduced. This network uses the upsampling feature mixer (UFM) to restore the resolution of deep feature maps, designs the dual-path complementary fusion (DCF) to fuse feature maps from different layers. Finally, the fused feature maps are input into the focused feature refinement (FFR) module for extracting detail features. By combining UFM, DCF and FFR to fuse features from different layers, this network avoids the problem of missing important feature information and insufficient feature fusion. By integrating the outputs of multiple modules, and comparing with existing state-of-the-art detection algorithms, the proposed IMFI-Net outperforms Faster R-CNN, YOLOv5, YOLOv8, ,YOLOv10, RT-DETR, CAGSDL-Net and MFO-Net by 23.7%, 3.4%, 5.1%, 6.2%, 7.7%, 4.7% and 3.4% in mAP 50-95. In terms of various evaluation metrics, the overall performance of the IMFI-Net method surpasses other underwater small target detection methods.