MGSD-YOLO: inland vessel target detection algorithm based on multiscale feature fusion

Abstract

To address the challenges in inland waterway vessel detection, such as uneven distribution of multi-scale targets, interference from complex background noise, and truncation of large-scale target images, this study proposes a multi-scale object detection algorithm, MGSD-YOLO (Multi-Grouped Semantic Decoupled YOLO), based on an improved YOLOv8 architecture. First, to enhance the model's feature representation capability for large-scale targets, a P6-level detection layer with a larger receptive field is introduced into the baseline network structure. Second, a feature extraction module based on group convolution, MGSA, is designed. This module employs parallel multi-scale convolution operations to effectively mitigate the degradation of small-target features caused by 64-fold downsampling, enabling the synergistic extraction of global semantic information and local detailed features. Finally, a lightweight decoupled detection head, DCE-Head, is constructed, which integrates classification, regression, and contextual information through a task-specific feature fusion mechanism, while optimizing bounding box regression accuracy using the DEL (Distribution Enhanced Loss) function based on KL (Kullback-Leibler) divergence. Experimental results demonstrate that, compared to the YOLOv8n baseline model, the proposed algorithm achieves a precision of 93.4% (an improvement of 3.9%) and a recall of 87.2% (an improvement of 5.5%) in the inland waterway vessel detection task, demonstrating the high detection accuracy of the improved algorithm.