HCFNet: A lightweight hybrid convolution–attention fusion network for photovoltaic cell defect detection in electroluminescence images

Under the goals of carbon neutrality and energy structure transition, photovoltaic systems are rapidly advancing toward large-scale deployment and high reliability. However, various defects inevitably occur in PV cells during manufacturing and long-term operation, leading to reduced power generation efficiency and potential safety risks. To address the challenges of diverse defect morphologies, weak texture characteristics in electroluminescence images, class imbalance, and the limited deployability of conventional deep learning models, this paper proposes a lightweight PV defect detection network, termed hybrid convolution–attention fusion network (HCFNet), which integrates convolutional and attention-based features. ConvNeXt is employed as the backbone to extract local texture features, while squeeze-and-excitation channel attention is introduced to enhance defect-related information. In addition, a hierarchical multi-head attention (HMHA) module is designed to model global spatial dependencies, improving the perception of weak-texture defects such as micro-cracks and fine line interruptions. Furthermore, a cross-feature fusion module is constructed to effectively integrate convolutional and attention features, enhancing feature discriminability while maintaining model efficiency. To mitigate class imbalance, data augmentation strategies and the Focal Loss function are adopted. Experimental results on the PVEL-AD dataset demonstrate that HCFNet achieves an overall accuracy of 96.06% and a precision of 96.10% on an eight-class defect classification task, significantly outperforming several mainstream CNN models. Ablation studies further verify the effectiveness of the proposed modules, indicating that the method exhibits strong robustness and promising engineering applicability for PV defect detection.