基于改进YOLO 11n的轻量化金鲳鱼摄食强度检测模型

Lightweight Feeding Intensity Detection Model for Trachinotus ovatus Based on Improved YOLO 11n

摘要:

在循环水养殖系统中,精准投喂对于提升饲料利用率、保障鱼类健康生长具有重要意义。过量投喂会造成资源浪费并诱发水体污染,而投喂不足则会影响鱼类生长效率。针对传统视觉检测模型参数量大、难以在边缘设备中高效部署的问题,以金鲳鱼为研究对象,提出了一种基于改进YOLO 11n的金鲳鱼摄食强度轻量化检测方法,实现金鲳鱼摄食强度的准确检测。提出的模型将传统的YOLO 11n模型中的卷积层替换为可变形卷积核模块(Arbitrary kernel convolution,AKConv),同时,在模型中融入可变形邻域注意力(Deformable neighborhood attention,DN – Attention),在降低模型计算成本的同时,提高了摄食个体数量的检测准确度。将检测模块输出的摄食个体数量与对应的时间帧作为循环神经网络 CNN – LSTM 模型的输入,通过无监督训练,将金鲳鱼的摄食强度分为:强、弱、无3个等级。试验结果证明,改进后的YOLO 11n检测模型,精确率达到87.6%,召回率达到81.5%,平均精度均值达到88.1%,相较于传统YOLO 11n模型,本研究提出的模型在精确率、召回率和平均精度均值上分别提升了3.5、3.6、2.4个百分点。实现了检测精度与轻量化的有效平衡,可为智能投喂装备的边缘部署提供技术支撑。

Abstract:

In recirculating aquaculture systems, precise control of feeding amounts is essential for optimizing feed utilization and promoting the healthy growth of fish. Overfeeding can result in resource waste and water pollution, while underfeeding can diminish fish growth efficiency. To address the challenges posed by large parameter sizes and the difficulty of efficient deployment on edge devices inherent in traditional visual inspection models, focusing on Trachinotus ovatus, a lightweight detection method for assessing feeding intensity was proposed based on an improved YOLO 11n model. The proposed model replaced the convolutional layers in the traditional YOLO 11n with deformable convolution kernel modules ( arbitrary kernel convolution, AKConv) and incorporated deformable neighborhood attention (deformable neighborhood attention, DN – Attention). This integration not only reduced the model's computational cost but also enhanced the detection accuracy of number of individuals ingested being fed. The number of individuals ingested and the corresponding time frames were input into a recurrent neural network ( CNN – LSTM) model. Through unsupervised training, the feeding intensity of Trachinotus ovatus was classified into three levels: strong, weak, and none. Experimental results demonstrated that the improved YOLO 11n detection model achieved an accuracy of 87.6%, a recall of 81.5%, and an average precision of 88.1%. Compared with the traditional YOLO 11n model, the proposed model improved accuracy, recall, and average precision by 3.5, 3.6, and 2.4 percentage points, respectively. This approach effectively balanced detection accuracy and model lightweightness, providing technical support for the edge deployment of intelligent feeding equipment.