In the quality inspection of photovoltaic (PV) modules, defect detection methods that combine electroluminescence (EL) imaging with deep learning have attracted considerable attention. However, EL images often suffer from complex background interference, minor target defects, and weak feature boundaries, which hinder the accuracy and robustness of existing detection methods. To address these challenges, this paper proposes an improved YOLOv9s algorithm, PV-YOLOv9s, to enhance the performance of defect detection in complex scenarios. First, wavelet gated convolution (WGConv) is proposed and the WGRepNCSPELAN4 module is further constructed to achieve the capture of multi-scale and multi-frequency information, which effectively enhances the background suppression and feature expression. Second, the Multi-Scale Dual Attention (MSDA) module is incorporated to promote multi-scale feature fusion and improve the model’s ability to detect defects of varying sizes. Furthermore, the integration of the Coordinate Attention (CoordAtt) mechanism strengthens the model’s joint perception of spatial positions and channel features, thus improving localization accuracy. Finally, the InnerEIoU loss function is employed to accelerate convergence and enhance bounding box regression precision. Experimental results demonstrate that the proposed PV-YOLOv9s achieves a mean Average Precision (mAP50:95) of 74.2% on a PV module EL image defect detection dataset, outperforming the latest YOLOv13s by 2.1%.