PMDS-YOLO: A lightweight multi-scale detector for efficient aquatic product detection

Abstract

Object detection is critical for smart aquaculture systems, yet current algorithms suffer from high-parameter models, low accuracy, and weak performance of small-target detection. For these challenges, we propose an aquatic products detection model PMDS-YOLO, dedicated to computational complexity reduction coupled with detection accuracy enhancement. First, the PMD (PConv-MDEMA) module is constructed. By cascading partial convolutions (PConv) with our proposed MDEMA (MLP-Drop path-EMA) mechanism, this structure substantially decreases model parameters and computational costs, concurrently attaining a marked enhancement in detection performance. Secondly, the MScat is designed to achieve multi-scale feature map fusion, enhancing detection capability. Meanwhile, the deep scale enhancement (DSE) model is designed to achieve collaborative optimization of cross-level features by establishing an interaction mechanism across hierarchical features. These designs enhance model robustness and make it better adapted to complex underwater scenarios. Finally, the shared fusion head (SFH) is proposed, improving detection accuracy while reducing the redundant computations. Simulation results show that PMDS-YOLO attains 84.8 % [email protected] performance on the URPC dataset, surpassing Faster-RCNN, SSD, EfficientDet-d0, RT-DETR-L YOLOv10n, YOLOv11n, YOLOv12n and YOLOv13n by 11.55 %, 9.4 %, 4.3 %, 4.9 %, 2 %, 1.4 %, 1.1 %, and 1.3 % respectively. Furthermore, experiments on RUOD and WSODD datasets confirm the superior generalization capability of PMDS-YOLO for underwater object detection tasks.