An improved feature aggregation network for photovoltaic cell defect detection

Detecting defects in photovoltaic cells is essential for maintaining the reliability and efficiency of solar power systems. Existing methods face challenges such as (1) the interaction and fusion of features at different layers in the feature extraction network of the object detection model are not sufficient, (2) ineffective detection of uniquely structured curved defects, such as line-crack, star-crack, and (3) difficulty in differentiating defects from the noisy backgrounds of photovoltaic cell. To address these issues, we propose an improved feature aggregation network (IFA-Net) for photovoltaic cell defect detection, which enhances feature fusion, utilizes dynamic snake efficient layer aggregation network, and integrates squeeze-and-excitation attention to improve detection accuracy and robustness. Our extensive comparative and ablation studies on photovoltaic electroluminescence anomaly detection (PVEL-AD) dataset demonstrate that the IFA-Net outperforms the baseline model YOLOv7-tiny, achieving improvements by 3.5% in mAP@0.5, 3.2% in mAP@0.5:0.95, and 3.6% in Recall.