基于YOLO v10s‑SEAM‑D2S的工厂化循环水养殖大西洋鲑检测模型

Atlantic Salmon Detection Model for Industrial Recirculating Aquaculture Based on YOLO v10s‑SEAM‑D2S

摘要:

在工厂化循环水养殖环境中，水下目标检测面临捕获图像质量较低及鱼群相互遮挡的挑战。为此，本文提出一种基于改进YOLO v10s的水下大西洋鲑检测模型YOLO v10s?SEAM?D2S，在颈部网络中引入分离增强注意力模块（Separated and enhancement attention module, SEAM），以增强图像未遮挡区域的特征响应，补偿被遮挡区域的响应损失，提高模型对遮挡目标的检测能力。提出并引入深度空间变换（Depth?to?space, D2S）卷积结构，以增加中间特征图像有效信息，提升模型对模糊图像的目标检测性能。在颈部网络中加入分组空间卷积（Grouped spatial convolution, GSConv），以降低模型计算成本并减少内存占用量。试验结果表明，在工厂化循环水养殖环境下改进YOLO v10s?SEAM?D2S模型对水下大西洋鲑目标检测数据集上显著优于原始YOLO v10s，平均精度均值和精确率分别达到91.6%和84.9%，较原始YOLO v10s 分别提升1.3、1.1个百分点。此外，改进模型在目标遮挡检测方面表现出更强的鲁棒性，与现有方法相比，在检测精度上更具优势，更适用于工厂化循环水养殖环境下水下目标检测任务。

Abstract:

Underwater object detection in industrial recirculating aquaculture systems presents significant challenges due to low?quality captured images and fish occlusions. To address these limitations, an enhanced YOLO v10s?based underwater Atlantic salmon detection method, termed YOLO v10s?SEAM?D2S was proposed. Firstly, a separated and enhancement attention module (SEAM) was incorporated into the neck network to amplify feature responses in non?occluded regions, thereby mitigating response loss in occluded areas and improving the model?s robustness in detecting occluded targets. Secondly, a novel depth?to?space (D2S) convolution was introduced to enrich the effective information within intermediate feature maps, enhancing the model?s capability to detect objects in blurred images. Finally, grouped spatial convolution (GSConv) was integrated into the neck network to reduce computational complexity, specifically by decreasing floating point operations per second (FLOPs) and memory footprint. Experimental evaluations demonstrated that the proposed YOLO v10s?SEAM?D2S model significantly outperformed the baseline YOLO v10s on an underwater Atlantic salmon detection dataset collected in an industrial recirculating aquaculture environment. The proposed model achieved a mean average precision (mAP) of 91.6% and a precision of 84.9%, yielding respective exhibited superior robustness in detecting occluded targets. Compared with existing state?of?the?art methods, the proposed approach achieved higher detection accuracy and was particularly well?suited for underwater object detection in industrial recirculating aquaculture systems.