Table 3 Comparative analysis of ML approaches for photovoltaic defect detection in EL imagery: traditional methods, conventional DL, and ResNet34-Based optimization.
From: ResNet-based image processing approach for precise detection of cracks in photovoltaic panels
Aspect | Traditional ML | Conventional DL | Proposed Method |
|---|---|---|---|
Input type | Handcrafted features | Raw EL/IR/RGB images | Raw EL images (high resolution) |
Feature extraction | Manual engineering | Automatic (convolutional layers) | Automatic via residual blocks with deep feature representation |
Model complexity | Low to moderate | Moderate to high | Moderate (ResNet34 strikes balance) |
Accuracy | Low to moderate | High | High |
Training data requirements | Small to medium datasets | Large datasets | Optimized for medium-sized dataset with class balancing |
Generalization | Limited (noise-sensitive) | High (but overfits on imbalance) | High; robust across defect types due to residual learning and augmentation |
Computation efficiency | Very efficient | Varies; heavier models (e.g., YOLOv5, ResNet152) require high compute | Efficient; suitable for edge deployment with ONNX export |
Detection types | Binary | Binary or multiclass (limited microcrack sensitivity) | Fine-grained (micro/macro, dormant) |
Deployment readiness | Offline tools | Resource-intensive | Edge devices (low power) |
Robustness to image quality | Low (needs clean input) | Moderate | High due to preprocessing + skip connections in ResNet |
Scalability to real-world PV systems | Low | Moderate to high | High (modular, portable) |
