Table 9 Summary of dataset collection, preprocessing, and annotation.
From: ResNet-based image processing approach for precise detection of cracks in photovoltaic panels
Aspect | Description |
|---|---|
Data collection | 2,000 EL images (monocrystalline & polycrystalline). |
Labeling method | Images labeled into 0%, 33%, 67%, 100% defect probability. |
Image preprocessing | Converted to 8-bit grayscale, and normalized for lighting. |
Contrast adjustment | Histogram equalization applied to standardize brightness and visibility. |
Data augmentation | Random flips, rotation, cropping, zoom, and Gaussian noise (via PyTorch). |
Standardization | Pixel intensity scaled to [0, 1]; orientation and framing standardized. |
Class imbalance | Balanced using oversampling and class-weighted “BCEWithLogitsLoss”. |
