摘要： 针对鲫鱼病害形态复杂、尺度差异大及病灶边界模糊所导致的检测精度低、误检率高等问题， 提出了一种基于YOLOv11框架的鲫鱼病害识别模型CAI-YOLO。首先，主干网络采用ConvNeXt V2(Convolutional Neural Network with NeXt Units Version 2)模块，该模块采用基于掩码自编码器(Masked Auto Encoders，MAE)的自监督预训练策略，并引入全局响应归一化(Global Response Normalization, GRN)层，有效缓解了特征崩溃问题，增强了特征多样性。其次，在颈部网络集成AKConv(Alterable Kernel Convolution)，通过自适应采样机制提升模型对不规则病斑的多尺度建模能力。最后，损失函数采用IF-IOU(Inner and Focaler Intersection Over Union)，该函数结合了Inner-IOU的内部约束与Focaler-IOU重加权机制，从而加快了模型的收敛并提升了定位精度。在自建鲫鱼病害数据集上进行试验，结果显示：CAI-YOLO模型的准确率、召回率、mAP@0.5和mAP@0.5:0.95分别为85.6%、87.8%、86.7%和58.6%，与基准YOLOv11n相比，mAP@0.5和mAP@0.5:0.95分别提高0.9和1.1个百分点；模型参数量、计算复杂度和模型尺寸分别降低10.89%、8.19%和7.84%。研究表明，CAI-YOLO模型在有效提升综合检测能力的同时也降低了计算资源的需求，为鲫鱼病害检测的轻量化和实际应用提供了参考。

关键词: 鲫鱼, 病害检测, 目标检测, YOLO, 深度学习

Abstract: To address the challenges of low detection accuracy and high false positive rates caused by the complex morphology, significant scale variations, and blurred boundaries of lesion areas in Carassius auratus diseases, this paper proposes a novel recognition model named CAI-YOLO based on the YOLOv11 framework. First, the backbone network incorporates the ConvNeXt V2 module. This module utilizes a self-supervised pre-training strategy based on Masked Auto Encoders and introduces a Global Response Normalization layer, effectively mitigating feature collapse and enhancing feature diversity. Second, the neck network integrates AKConv, which leverages an adaptive sampling mechanism to improve the model's multi-scale modeling capability for irregular disease spots. Finally, the loss function employs IF-IOU, which combines the internal constraints of Inner-IOU with the re-weighting mechanism of Focaler-IOU, thereby accelerating model convergence and improving localization accuracy. Experiments conducted on a self-built Carassius auratus disease dataset show that the CAI-YOLO model achieves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 of 85.6%, 87.8%, 86.7%, and 58.6%, respectively. Compared to the baseline YOLOv11n, the mAP@0.5 and mAP@0.5:0.95 are increased by 0.9 and 1.1 percentage points, respectively. Furthermore, the number of parameters, computational complexity, and model size are reduced by 10.89%, 8.19%, and 7.84%, respectively. The research demonstrates that the CAI-YOLO model effectively enhances overall detection performance while simultaneously reducing computational resource requirements, providing a valuable reference for the lightweight design and practical application of Carassius auratus disease detection systems.

Key words: Carassius auratus, Disease Detection, Object Detection, YOLO, Deep Learning