【摘要】 本研究针对人工养殖环境中海马细菌性肠炎体征难以精准检测的问题,构建了专用水下海马图像数据集,评估了多种图像增强算法对检测性能的影响,并提出了一种基于YOLO11的目标检测模型YOLO11-SH。该模型在YOLO11的基础上,通过针对性的网络架构优化来增强目标检测模型性能:引入多尺度空洞注意力与通道注意力机制,增强复杂水下环境中对微小目标的细粒度特征提取能力;采用动态上采样策略自适应调整采样权重,有效恢复浅层空间细节;结合Shape-IoU损失函数,提升海马细长形态的边界框回归定位精度。结果表明,直接应用图像增强算法并未实现预期的性能增益;网络架构优化后的YOLO11-SH在参数量仅为2.9M的情况下,mAP@50和mAP@50-95分别达到84.3%和63.6%,召回率达到77.6%,较基线模型分别提升1.2%、1.7%和2.0%。此外,鲁棒性测试证明该模型在运动模糊、高斯噪声等非理想水下退化条件下仍能维持稳定的检测性能。本研究为养殖过程中海马健康监测及其细菌性肠炎的早期预警提供了有效的技术支撑。更多还原

【Abstract】 To address the challenge of accurately detecting signs of bacterial enteritis in cultured seahorses, this study constructs a dedicated underwater seahorse image dataset and proposes YOLO11-SH, a lightweight detection model based on an improved YOLO11. The study first evaluates the impact of various image enhancement algorithms on detection performance and finds that directly applying enhancement preprocessing does not universally improve performance for object detection. Therefore, this study focuses on targeted optimizations of the YOLO11 network architecture. Specifically, it introduces multi-scale dilated attention(MSDA) and multi-scale channel attention(MSCA)mechanisms to enhance fine-grained feature extraction capabilities for small targets in complex underwater environments. A dynamic upsampling(DySample) strategy was adopted to adaptively adjust sampling weights and effectively recover shallow spatial details. Additionally, the Shape-IoU loss function is integrated to improve bounding box regression and localization accuracy for the elongated morphology of seahorses. Experimental results demonstrate that with a parameter size of only 2.9M, YOLO11-SH achieves an mAP@50 of 84.3%, mAP@50-95 of 63.6%, and a recall rate of 77.6%, representing improvements of 1.2%, 1.7%, and 2.0% over the baseline model, respectively.Furthermore, robustness tests prove that the model maintains stable detection performance under non-ideal underwater degradation conditions, such as motion blur and Gaussian noise. This study provides efficient and feasible technical support for intelligent health monitoring and early disease warning in seahorse aquaculture.更多还原