The photovoltaic (PV) industry is crucial to the global development of renewable energy. Electroluminescence (EL) imaging technology has been widely employed for detecting internal defects in photovoltaic modules. However, traditional methods are suffered from the low detection efficiency and the false positive rates are unsatisfying. This paper proposes an improved YOLOv9 network for photovoltaic module defect detection, integrating multi-scale feature fusion and self-attention mechanisms to effectively learn and capture defect features in PV modules. The proposed network introduces the DySample technique to enhance sampling accuracy and efficiency. Additionally, the Efficient Multi-Scale Attention Module (EMA) is employed to enhance spatial information integration through the cross-space mechanism. Furthermore, the Attention Convolution Mix (ACMix) module is integrated into the detection head to improve the fusion of global and local features. Experimental results demonstrate that the mAP@0.5 and mAP@0.95 of our method are 94.43% and 84.50%, respectively. The proposed method exhibits better performance compared to other deep learning methods, showing considerable potential in the field of photovoltaic defect detection.