【摘要】 针对现有绝缘子缺陷检测模型结构复杂、参数量大且在复杂背景下缺陷检测精度低等问题，提出一种改进YOLOv8n的轻量化绝缘子缺陷检测算法——CBL-YOLOv8n(CBL:C2f-ConvFormerCGLU、BiFPN、LDES)。首先，将YOLOv8n中的C2f模块替换为轻量级C2f-ConvFormerCGLU模块，有效减少了模型的参数量；其次，采用加权双向特征金字塔网络(BiFPN)来优化颈部网络，通过可学习的权重进行跨尺度特征融合，在实现多尺度特征融合的同时进一步压缩了模型的大小；最后，提出一种新型的轻量化细节增强共享(LDES)检测头，旨在实现模型精度与效率的平衡，以满足边缘部署需求。在合成雾绝缘子数据集(SFID)上的实验结果表明，与原始YOLOv8n相比，CBL-YOLOv8n模型在检测精度上达到了99.2%,参数量减少了56.7%,计算量降低了39.5%,体积减小了48.3%。结果表明，在确保高检测精度的同时，实现了模型的轻量化。更多还原

【Abstract】 To address the issues of complex structures, large parameter quantities, and low detection accuracy under complex backgrounds associated with existing insulator defect detection models, this paper proposed a lightweight insulator defect detection algorithm based on improved YOLOv8n, namely CBL-YOLOv8n(the acronym CBL represented the three core improvements: C2f-ConvFormerCGLU, BiFPN, and LDES). First, the original C2f module in YOLOv8n was replaced with a lightweight C2f-ConvFormerCGLU module, effectively reducing the number of parameters. Second, a weighted bi-directional feature pyramid network(BiFPN) was adopted to optimize the neck network, performing cross-scale feature fusion through learnable weights, which not only achieved multi-scale feature fusion but also further compressed the model size. Finally, a novel lightweight detail enhancement shared(LDES) detection head was proposed, aiming to achieve a balance between model accuracy and efficiency to meet edge deployment needs. Experimental results on the synthetic foggy insulator dataset(SFID) showed that compared with the original YOLOv8n, the proposed CBL-YOLOv8n model achieved a detection accuracy of 99.2%, with the parameter quantity reduced by 56.7%, computational load decreased by 39.5%, and model size reduced by 48.3%. The results verified that the proposed model achieves excellent lightweight performance while maintaining high detection accuracy.更多还原