Precise detection and classification of multi-scale floating debris in complex irrigation water sources constitute an important approach for achieving automation in water quality management. However, the existing object detection algorithms have low detection and classification accuracy for densely distributed floating debris of different scales in complex water surface scenes. These challenges are primarily attributed to tightly clustered point targets, substantial scale variation among objects, and water-surface specular reflections. To address the above problems, we propose MSP-RTDETR, a novel real-time detector that improves the RT-DETR architecture by introducing three novel modules. First, MADR is proposed, which incorporates inverted residual and dilated re-parameterized convolutions for multi-scale feature perception. Second, the Spatial Information Reorganization and Multi-Scale Omnidirectional Convolution Kernel Mechanism (SIPR-MSFE) is proposed to realize the fusion of contextual and detailed information. Third, a novel encoder layer (PATE) is introduced to improve the model’s robustness to complex backgrounds through polar coordinate linear attention. Experimental results demonstrate that MSP-RTDETR achieves a precision of 86.8%, a recall of 83.3%, and an mAP50 of 85.1%, surpassing RT-DETR-r18 by 7.0%, 3.2%, and 3.2% points, respectively. The effectiveness of these improvements is further validated through Grad-CAM heatmap analysis and visualization experiments, with GFLOPs and parameter count reduced by 6.5 G and 4.5 M, respectively. Comparisons with several mainstream detection algorithms demonstrate that MSP-RTDETR achieves competitive detection accuracy and strong generalization capability. The proposed algorithm also provides robust support for practical applications such as UAV-based and vessel-mounted surveillance, facilitating intelligent management of irrigation water resources.
Cai et al. (Mon,) studied this question.