Photovoltaic Cell Surface Defect Detection Based on Wavelet-Aware Perception and Selective Reconstruction

Defect detection in electroluminescence (EL) images of solar cells remains challenging because defects are often small, exhibit subtle grayscale variations, and are obscured by complex cell-texture backgrounds. These conditions can attenuate defect features as they propagate through deep layers and make them difficult to distinguish from background patterns, thereby limiting the ability of detection models to accurately localize and classify defects. To address these issues, this study proposes a Wavelet-Aware Selective Reconstruction Network (WASR) for solar cell defect detection. A frequency-aware encoding strategy built around the Wavelet-Aware Downsampling Encoder (WADE) module preserves fine-grained defect cues by jointly modeling spatial and frequency-domain characteristics. In addition, a Cross-Scale Context Fusion Module (CCFM) improves multilevel feature transmission and fusion, whereas a Selective Reconstruction Detection Head (SRD-Head) performs selective feature reconstruction with residual enhancement. Together, these modules enhances defect perception capability under low-contrast conditions and complex backgrounds. Compared with the YOLO11 baseline, WASR increases mAP@0.5 by 2.4 percentage points, reduces the parameter count by 0.62 M, and improves inference speed by 23 FPS. The results show that WASR provides a favorable balance between detection accuracy and computational efficiency. Ablation experiments and visualization results further confirm the contribution of the proposed components under challenging conditions.