LMFF-Net: Lightweight Multi-Scale Feature Fusion Object Detection for Underwater IoT Networks

With the advancement of Underwater Internet of Things (U-IoT) networks, underwater object detection has become crucial. However, detection performance is severely hindered by underwater image degradation, small object scales and weak textures, as well as the stringent resource limitations of deployment platforms. Meanwhile, existing detectors rely on heavy backbones and standard downsampling, which not only increases computation but also discards fine-grained spatial information, resulting in low precision for small objects and suboptimal deployability. To address these challenges, a lightweight multi-scale feature fusion network named LMFF-Net is proposed. Utilizing RepGhostNet as the feature extraction network and incorporating an enhanced bidirectional feature pyramid structure, the proposed LMFF-Net effectively reduces the number of parameters and computational cost while maintaining detection accuracy. In the feature pyramid, a multi-scale object enhancement (MSOE) module is designed, which synergistically combines spatial multi-scale convolutions with spatial-frequency attention mechanisms. This synergy enhances feature discriminability and improves small object representation. Furthermore, a channelized spatial fusion (CSF) module is constructed to achieve lossless downsampling of feature maps by leveraging the spatial-to-depth transformation principle, thereby maximizing the retention of fine-grained spatial information in deep networks. Experimental results on the DUO dataset show that LMFF-Net achieves an mAP@0.5 of 84.7% with only 1.73 M parameters, significantly outperforming other classic object detection models. Additionally, generalization experiments on the URPC2020 and RUOD datasets also demonstrate the excellent generalization ability of the proposed model.