Song Lili, Deng Haixin, Han Jianfeng, Gao Xiongwei
School of Information Engineering, Inner Mongolia University of Technology, Jinchuan Campus, Hohhot 010080, China.
Inner Mongolia Key Laboratory of Intelligent Perception and System Engineering, Hohhot 010080, China.
Sensors (Basel). 2025 Mar 20;25(6):1938. doi: 10.3390/s25061938.
The water surface environment is highly complex, and floating objects in aerial images often occupy a minimal proportion, leading to significantly reduced feature representation. These challenges pose substantial difficulties for current research on the detection and classification of water surface floating objects. To address the aforementioned challenges, we proposed an improved YOLOv8-HSH algorithm based on YOLOv8n. The proposed algorithm introduces several key enhancements: (1) an enhanced HorBlock module to facilitate multi-gradient and multi-scale superposition, thereby intensifying critical floating object characteristics; (2) an optimized CBAM attention mechanism to mitigate background noise interference and substantially elevate detection accuracy; (3) the incorporation of a minor target recognition layer to augment the model's capacity to discern floating objects of differing dimensions across various environments; and (4) the implementation of the WIoU loss function to enhance the model's convergence rate and regression accuracy. Experimental results indicate that the proposed strategy yields a significant enhancement, with mAP50 and mAP50-95 increasing by 11.7% and 12.4%, respectively, while the miss rate decreases by 11%. The F1 score has increased by 11%, and the average accuracy for each category of floating objects has enhanced by a minimum of 5.6%. These improvements not only significantly enhanced the model's detection accuracy and robustness in complex scenarios but also provided new solutions for research in aerial image processing and related environmental monitoring fields.
水面环境高度复杂，航空图像中的漂浮物体通常占比极小，导致特征表示显著减少。这些挑战给当前水面漂浮物体的检测与分类研究带来了巨大困难。为应对上述挑战，我们提出了一种基于YOLOv8n的改进型YOLOv8-HSH算法。该算法引入了多项关键改进：（1）增强的HorBlock模块，以促进多梯度和多尺度叠加，从而强化关键漂浮物体特征；（2）优化的CBAM注意力机制，以减轻背景噪声干扰并大幅提高检测精度；（3）纳入小目标识别层，以增强模型在各种环境中辨别不同尺寸漂浮物体的能力；（4）实施WIoU损失函数，以提高模型的收敛速度和回归精度。实验结果表明，所提出的策略取得了显著提升，mAP50和mAP50-95分别提高了11.7%和12.4%，而漏检率降低了11%。F1分数提高了11%，各类漂浮物体的平均准确率至少提高了5.6%。这些改进不仅显著提高了模型在复杂场景中的检测精度和鲁棒性，还为航空图像处理及相关环境监测领域的研究提供了新的解决方案。
