MSN-TE: a network for multi-scale object detection in turbid environments

Underwater object detection is challenged by turbidity-induced low contrast, scattering, and small targets that occupy only a limited portion of the image. To address these issues, this study proposes MSN-TE, a multi-scale detection network for turbid underwater environments. MSN-TE contains three task-oriented modules: C3k2_FTEM for dynamic-range feature enhancement, CSP_DPA for lightweight dual-path feature aggregation, and MAEDH for multi-level auxiliary detection. C3k2_FTEM improves the representation of degraded underwater features, CSP_DPA strengthens multi-scale detail aggregation while controlling computational complexity, and MAEDH enhances high-resolution small-target detection. Experiments reported in the study show that MSN-TE achieves 86.8% mAP on URPC2020 and 87.1% mAP on URPC2019, with 13.7 GFLOPs and 2.2M parameters, indicating a favorable balance between detection accuracy and computational cost for underwater engineering applications.