Hyperparameters for training and testing.

Solar panel defect detection, a crucial quality control task in the manufacturing process, often faces challenges such as varying defect sizes, severe image background interference, and imbalanced data sample distribution. To address these issues, this paper proposes the EBBA-Detector. The core of the model lies in an enhanced balanced attention framework, which includes an Enhanced Bidirectional Feature Pyramid Network (EBFPN) and a Balanced-Attention Module (B-A Module). The EBFPN captures defect features of different sizes, significantly improving the recognition ability for small defects, while the B-A Module suppresses background interference, guiding the model to focus more on defect locations. Additionally, this paper designs a Scaled Dynamic Focal Loss (SDFL) function, which enables the model to pay more attention to minority and hard-to-identify defect samples under imbalanced data distribution. Through experimental validation on a large-scale electroluminescence (EL) dataset, the proposed method has achieved significant improvements in detection performance, with a mean Average Precision (mAP) of 89.85%, outperforming other models in multiple defect category detections. Therefore, the EBBA-Detector not only effectively detects small target objects but also demonstrates good handling capabilities for large targets and imbalanced data, providing an efficient and accurate solution for solar panel defect detection.