This paper proposes an improved VarifocalNet for detecting defects in photovoltaic (PV) modules, aiming to enhance both detection speed and accuracy. The method introduces two improved bottleneck modules and a feature interactor to enhance performance. The first bottleneck module increases the receptive field without reducing the feature map size, improving defect detection accuracy. The second bottleneck module reduces parameters to improve detection speed. A feature interactor is designed to enhance feature expression in the classification branch, and an improved intersection over union loss function is introduced to better measure the difference between predicted and ground truth boxes. The proposed method achieves the highest detection accuracy and is faster than most other methods except for DDH-YOLOv5 and improved YOLOv7. PV modules are vulnerable to various external factors, and defects can lead to module failure, reducing power output and posing safety risks. Manual inspection and traditional methods are inefficient and cannot meet the demand for large-scale PV systems. Therefore, accurate and efficient defect detection is essential for ensuring the long-term stable operation of PV systems. Defect detection in PV modules can be categorized into two types: analyzing electrical parameters and analyzing visible light, electroluminescence (EL) images, or thermal imaging. EL imaging is widely used for detecting microcracks due to its high resolution. Deep learning methods have become the mainstream approach for detecting defects in EL images due to their strong feature-capturing ability. The paper presents an improved VarifocalNet for detecting defects in PV modules. The backbone network is based on ResNet-101, with two improved bottleneck modules and a feature interactor in the detection head. The improved bottleneck modules enhance detection accuracy and speed, while the feature interactor improves feature expression. An improved GIoU loss function is introduced to better measure the deviation between predicted and ground truth boxes. The proposed method is tested on the PVEL-AD dataset, achieving the highest detection accuracy and recall. The method outperforms other methods in detection accuracy and is faster than most methods. The improved VarifocalNet has the highest detection accuracy and is second only to improved YOLOv7 and DDH-YOLOv5 in detection speed. The method improves both detection accuracy and speed for detecting defective PV modules. Future work includes replacing horizontal prediction boxes with polar coordinate rotated prediction boxes to address feature noise issues.This paper proposes an improved VarifocalNet for detecting defects in photovoltaic (PV) modules, aiming to enhance both detection speed and accuracy. The method introduces two improved bottleneck modules and a feature interactor to enhance performance. The first bottleneck module increases the receptive field without reducing the feature map size, improving defect detection accuracy. The second bottleneck module reduces parameters to improve detection speed. A feature interactor is designed to enhance feature expression in the classification branch, and an improved intersection over union loss function is introduced to better measure the difference between predicted and ground truth boxes. The proposed method achieves the highest detection accuracy and is faster than most other methods except for DDH-YOLOv5 and improved YOLOv7. PV modules are vulnerable to various external factors, and defects can lead to module failure, reducing power output and posing safety risks. Manual inspection and traditional methods are inefficient and cannot meet the demand for large-scale PV systems. Therefore, accurate and efficient defect detection is essential for ensuring the long-term stable operation of PV systems. Defect detection in PV modules can be categorized into two types: analyzing electrical parameters and analyzing visible light, electroluminescence (EL) images, or thermal imaging. EL imaging is widely used for detecting microcracks due to its high resolution. Deep learning methods have become the mainstream approach for detecting defects in EL images due to their strong feature-capturing ability. The paper presents an improved VarifocalNet for detecting defects in PV modules. The backbone network is based on ResNet-101, with two improved bottleneck modules and a feature interactor in the detection head. The improved bottleneck modules enhance detection accuracy and speed, while the feature interactor improves feature expression. An improved GIoU loss function is introduced to better measure the deviation between predicted and ground truth boxes. The proposed method is tested on the PVEL-AD dataset, achieving the highest detection accuracy and recall. The method outperforms other methods in detection accuracy and is faster than most methods. The improved VarifocalNet has the highest detection accuracy and is second only to improved YOLOv7 and DDH-YOLOv5 in detection speed. The method improves both detection accuracy and speed for detecting defective PV modules. Future work includes replacing horizontal prediction boxes with polar coordinate rotated prediction boxes to address feature noise issues.