Edge prior guidance and hybrid receptive field perception for water surface object detection in adverse navigation environments

Objective Robust water surface object detection is essential for collision avoidance and safe autonomous navigation of intelligent ships. However, fog, rain, nighttime low illumination, and backlighting degrade image contrast, texture, and target boundaries; repeated downsampling in deep networks further attenuates edge information. Water surface objects also vary substantially in scale and geometry, from small compact buoys to large horizontally elongated vessels. These coupled factors limit the robustness of existing detectors.

Method To address these challenges, an Edge-Prior and Hybrid-Receptive-Field DETR (EP²H-DETR) is proposed. First, a Difference-of-Gaussian Stem (DoG-Stem) explicitly introduces edge priors at the input stage. DoG filtering enhances intensity-transition regions, Gaussian smoothing suppresses degradation-induced noise, and dual-path downsampling combines depthwise convolution and max pooling to preserve salient boundary responses. Second, a Global Edge Feature Enhancement (GEFE) mechanism establishes a dedicated edge pathway across the backbone and encoder. A multi-scale edge feature extractor derives multi-scale edge representations from shallow features, while edge feature fusion modules inject them into semantic features throughout the network to compensate for progressive boundary information loss in deeper stages. Third, a CSP-based Hybrid Receptive Field Perception Module (CSP-HRFPM) is developed. Its multi-receptive-field dynamic convolution adaptively assigns channel-wise weights to square, horizontal-strip, and vertical-strip depthwise convolution branches, enabling content-dependent adaptation to compact, horizontally extended, and vertically structured targets while improving feature reuse and computational efficiency. An Adverse Conditions Inland Waterway Dataset (ACIWD) is also constructed with 7,270 images and 38,421 annotated instances, with adverse conditions accounting for 88.6%.

Results On ACIWD, EP²H-DETR achieves 73.7% mAP@0.5:0.95 versus 70.6% for the baseline, corresponding to a 4.4% relative improvement. The parameter count decreases from 19.9 M to 15.1 M (24.1%), and the model size from 77.0 MB to 59.0 MB (23.4%). Under fog and nighttime conditions, mAP shows relative gains of 10.0% and 7.7%, respectively. For representative small- and large-scale targets, AP for buoys increases from 52.9% to 70.5%, while AP for container ships rises from 66.3% to 79.7%, demonstrating substantially improved detection performance across different target scales. After TensorRT FP16 optimization, inference speed increases from 48.4 to 102.3 FPS on an RTX 4060Ti without accuracy loss. Cross-dataset experiments on WSODD, SeaShips, and FloW further demonstrate consistent performance improvements, indicating good generalization across different maritime datasets.

Conclusion The proposed method significantly improves the robustness of water surface object perception for intelligent ships operating in complex adverse navigation environments, providing effective technical support for inland waterway safety assurance and intelligent navigation perception systems.