摘要： 针对鱼类目标检测存在精度低和计算量大的问题，提出了一种基于改进YOLOv8模型的轻量化鱼类目标检测方法YOLOv8-FCW。首先，引入FasterNet中的FasterBlock替换YOLOv8中C2f模块的Bottleneck结构，减少网络模型的冗余计算；其次，引入注意力机制CBAM(Convolutional Block Attention Module)，实现高效提取鱼体特征，提升网络模型检测精度；最后，引入动态非单调聚焦机制WIoU(Wise Intersection over Union)替代CIoU(Complete Intersection over Union)，加快网络模型的收敛速度，提升网络模型的检测性能。结果显示，与原模型相比，改进YOLOv8-FCW模型精确率提升了1.6个百分点，召回率提升了5.1个百分点，平均精确率均值提升了2.4个百分点，权重和计算量分别减少为原模型的80%和79%。该模型具有较高的精确率和较强的鲁棒性，该研究能够帮助养殖者精确计算鱼群数量，提高养殖效率。

关键词: 鱼类, 图像处理, 图像识别, 目标检测, YOLOv8

Abstract: At present, fish culture is moving in the direction of precision culture, and fish target detection is an important part of precision culture. Fortunately, the use of deep learning holds promise for fish target detection. However, the existing fish target detection models have the problems of heavy computation and low accuracy.To address the issues of low accuracy and high computational load in fish target detection,a lightweight fish target detection method based on an improved YOLOv8 model was proposed and named YOLOv8-FCW in this study. Firstly,The experimental comparison of MobileNet, ShuffleNet, GhostNet and C2f-Faster show that C2f-Faster has the best performance.Therefore,the FasterBlock from FasterNet was introduced to replace the Bottleneck module in C2f of YOLOv8, reducing redundant computations in the network model. Secondly, the Convolutional Block Attention Module (CBAM) attention mechanism was incorporated to efficiently extract fish body features and enhance the detection accuracy of the network model. Finally, The experimental results show that the loss value and convergence speed of Wise intersection over union (WIoU) loss function are better than Complete intersection over union (CIoU), Distance intersection over union (DIoU) and Generalized Intersection over Union (GIoU).Therefore,a dynamic non-monotonic focusing mechanism WIoU was introduced to replace CIoU, accelerating the convergence speed of the network model and improving its detection performance.In order to verify the detection effect of YOLOv8-FCW on fish, the original model and YOLOv8-FCW were trained and tested on the fish data set.The fish data set consists of 1000 images, which were divided into training set, verification[] set and test set according to the ratio of 8:1:1. Experimental results show that compared with the original model, the improved YOLOv8-FCW model had increased precision by 1.6 percentage points, recall by 5.1 percentage points, and mean average precision(mAP) mAP0.5 metrics by 2.4 percentage points, while the weight and computational load were reduced to 80% and 79% of the original model, respectively.YOLOv8-FCW achieves high detection accuracy and efficiency with very small model volume and low computational cost. The model shows high accuracy and robustness. The research can help breeders accurately calculate the number of fish and provide technical reference for fish target detection.

Key words:  , image processing, image recognition, fish, target detection, YOLOv8