摘要： 针对水下目标检测中环境光照不足、目标呈现尺寸较小和目标聚集与相互遮挡导致检测精度下降的问题，提出了一种基于感受野特征的多尺度水下目标检测算法FDM-YOLO。首先，针对水下环境光照不足、水下生物多为小目标且与环境颜色相似，采用RFCADown模块生成大感受野空间特征，强化对水下目标关键信息的提取能力；其次，引入Dysample上采样模块以抑制传统上采样过程中的模糊与失真现象；再次，设计多尺度多维信息协作模块C3K2-IMCA，提升对密集遮挡目标的表征性能；最后，采用WIoU替代CIoU损失函数，缓解小目标因极端形状边界框对模型训练的负面影响。试验结果表明，FDM-YOLO在DUO数据集比基准模型的mAP50与mAP@50-95分别提升了2.1%和2.0%，同时模型参数量和计算量仅为2.35M与6.0GFLOPs。上述结果验证了改进模型在提高水下小目标检测性能方面的高效性。

关键词: 水下目标检测, 感受野特征, 多尺度, 损失函数

Abstract: In order to address the challenges of insufficient ambient lighting, small target size, and target clustering and occlusion leading to decreased detection accuracy in underwater target detection, this study proposes a multi-scale underwater target detection algorithm, FDM-YOLO, based on receptive field features. First, to address the issues of insufficient underwater ambient lighting and the fact that underwater organisms are often small targets with colors similar to their surroundings, the RFCADown module is used to generate large receptive field spatial features, enhancing the extraction of key information about underwater targets. Second, a Dysample upsampling module is introduced to suppress blurring and distortion in traditional upsampling processes. Third, a multi-scale, multi-dimensional information collaboration module, C3K2-IMCA, is designed to improve the representation performance of densely occluded targets. Finally, WIoU is used instead of CIoU loss function to mitigate the negative impact of extreme-shaped bounding boxes on model training for small targets. Experimental results show that FDM-YOLO achieves a 2.1% and 2.0% improvement in mAP50 and mAP@50-95 respectively compared to the benchmark model on the DUO dataset, while the model parameters and computational cost are only 2.35M and 6.0 GFLOPs. The above results verify the efficiency of the improved model in enhancing the detection performance of small underwater targets.

Key words: underwater object detection, receptive field features, multi-scale, loss function