基于FAS-YOLOv8n的爬岸上草小龙虾多源图像融合识别方法

Multi-source Image Fusion Recognition of Crawling Bank Grass Crayfish Based on FAS-YOLOv8n

摘要:

针对小龙虾养殖夜间巡塘效率低、劳动成本高的问题，提出了一种基于改进YOLO v8n的轻量化爬岸上草小龙虾识别方法(FAS-YOLOv8n)。首先，针对夜晚自然环境下小龙虾图像质量低的问题，采集RGB和红外图像，融合小龙虾的多源信息。其次，在YOLO v8n的骨干网络中使用Ghost卷积和C2f_Repghost模块，减少模型的参数量。然后，在骨干网络和颈部网络之间添加可变形注意力（Deformable attention，DA）机制，增强模型对小龙虾的关注度，提高模型的特征提取效率。最后，采用VoVGSCSP模块替换C2f模块，提升颈部网络的特征融合速度，进一步降低计算量。实验结果表明，FAS-YOLOv8n模型在融合图像数据集上的识别精确率为90.62%，平均精度均值和召回率分别为92.9%和85%。相较于RGB图像数据集和红外图像数据集，识别精确率、平均精度均值和召回率分别提高6.05、8.46个百分点， 4.78、7.14个百分点， 3.84、3.87个百分点。利用融合数据集进行试验，FAS-YOLOv8n模型较原始模型平均精度均值提高5.1个百分点，参数量和浮点数运算量分别降低13.29%和23.17%，模型内存占用量仅为6.2MB，检测速度为86f/s。识别效果优于其他主流目标检测模型，能够实现模型轻量化部署，为巡塘无人机的应用提供技术支撑。

Abstract:

To address the issues of low efficiency and high labor costs associated with nighttime inspections in crayfish farming, a lightweight method for identifying crayfish on the bank, termed FAS-YOLOv8n, was proposed based on an improved YOLO v8n architecture. Firstly, to tackle the problem of poor image quality of crayfish captured at night, both RGB and infrared images were collected to integrate multi-source information. Secondly, Ghost convolutions and C2f_Repghost modules were employed in the backbone network of YOLO v8n to reduce the model’s parameter count. Additionally, a deformable attention (DA) mechanism was introduced between the backbone and neck networks to enhance the model’s focus on crayfish and improve feature extraction efficiency. Finally, the VoVGSCSP module replaced the C2f module to accelerate feature fusion in the neck network, further reducing computational load. Experimental results indicated that the improved FAS-YOLOv8n model achieved a recognition accuracy of 90.62% on the integrated image dataset, with mean average precision (mAP) of 92.9% and recall rate of 85%. Compared with the RGB and infrared image datasets, the recognition accuracy, mAP, and recall rate was improved by 6.05 and 8.46, 4.78 and 7.14, and 3.84 and 3.87 percentage points, respectively. When tested on the integrated dataset, the improved FAS-YOLOv8n model demonstrated a 5.1 percentage points increase in mAP over the original model, while reducing parameter count and computational load by 13.29% and 23.17%, respectively. The model size was only 6.2MB, with detection speed of 86 frames per second. Its recognition performance surpassed that of other mainstream object detection models, enabling lightweight deployment and providing technical support for the application of drones in pond inspections.