无线电工程

光伏组件表面缺陷的高效检测是保障光伏系统可靠性的关键，现有方法在微小缺陷检测、复杂背景干扰抑制等方面存在显著挑战。为此，提出一种基于YOLOv8n改进的光伏缺陷检测模型MBL-YOLOv8n。该模型通过两模块优化策略提升性能。在Backbone中引入了C2f-MBConv模块，结合深度可分离卷积与挤压-激励(Squeeze-and-Excitation, SE)注意力机制，增强了对微小缺陷(如油墨点、微裂纹)的细粒度特征提取能力；引入大尺度选择性卷积核注意力(Large-scale Selective Convolutional Kernel Attention, LSK-Attention)机制，通过动态融合多尺度卷积核的特征响应，自适应调整感受野范围，显著提升了模型对低对比度缺陷的敏感性与复杂背景的鲁棒性。在自制的光伏组件数据集上的实验结果表明，与原始的YOLOv8n相比，所提出的MBL-YOLOv8n模型在各项指标上取得了良好的提升，将mAP50提高了2%,召回率和精度分别提升了1.8%和2.5%。MBL-YOLOv8n模型在光伏组件缺陷检测中展现出较好的精度，为光伏产业智能化质检提供了有效的解决方案。

Efficient detection of surface defects on PV modules is crucial for ensuring the reliability of PV systems, however, existing methods still face significant challenges in the detection of tiny defects and the suppression of complex background interference. To address this, an improved PV defect detection model MBL-YOLOv8n based on YOLOv8n is proposed, which improves the performance through a two-module optimization strategy. Firstly, a C2f-MBConv module is introduced in Backbone, which combines depth-separable convolution with the Squeeze-and-Excitation(SE) attention mechanism to enhance the fine-grained feature extraction capability for tiny defects(e.g. ink dots, micro-cracks). Then, the Large-Scale Selective Convolutional Kernel Attention(LSK-Attention) mechanism is introduced, which dynamically fuses the feature responses of multi-scale convolution kernels to adaptively adjust the range of receptive field. This significantly improves the model's sensitivity to low-contrast defects and its robustness to complex backgrounds. The experimental results on the self-made PV module dataset show that, compared with the original YOLOv8n, the proposed MBL-YOLOv8n model achieves good improvement across all indexes. The mAP50 is improved by 2%, and the recall and precision by 1.8% and 2.5%, respectively. The MBL-YOLOv8n model demonstrates a better precision in defect detection of PV modules, and provides an effective solution for intelligent quality inspection in PV industry.