With the widespread application of deep learning in image detection, solar panel defect identification has increasingly adopted convolutional neural networks (CNNs). This study introduces an optimized framework combining Single Shot MultiBox Detector (SSD) and Residual Networks (ResNet) to address challenges in localization accuracy and classification performance for photovoltaic module defects.
1. Methodology
The proposed dual-stage architecture first locates solar panels using enhanced SSD, then classifies defects through modified ResNet. Key innovations include:
1.1 SSD Network Optimization
We integrate Convolutional Block Attention Module (CBAM) into VGG16 backbone to strengthen multi-scale feature extraction. The default box dimensions are redesigned according to solar panel geometry:
$$S_k = S_{min} + \frac{S_{max} – S_{min}}{m – 1}(k – 1), \quad k \in [1,m]$$
$$w^a_k = S_k\sqrt{a_r}$$
$$h^a_k = \frac{S_k}{\sqrt{a_r}}$$
where \(a_r \in \{1,2,3,\frac{1}{2},\frac{1}{3}\}\) represents aspect ratios optimized for solar panel detection.
1.2 ResNet Enhancement
Squeeze-and-Excitation (SE) blocks are embedded in residual units to amplify critical channel features:
$$X^* = X \cdot \sigma(W_2\delta(W_1z))$$
where \(z\) denotes global pooled features, \(W_1\) and \(W_2\) are FC layer weights, and \(\sigma\) represents sigmoid activation.
2. Experimental Validation
We evaluate performance on a proprietary dataset containing 8,000 solar panel images with three defect types: gridline interruption, microcracks, and shadow occlusion.
| Model | Accuracy (%) | FPS |
|---|---|---|
| Faster R-CNN | 90.37 | 7 |
| RetinaNet | 85.99 | 30 |
| YOLOv3 | 87.59 | 42 |
| Improved SSD | 97.23 | 21 |
The enhanced SSD demonstrates 9.86% higher accuracy than baseline while maintaining real-time processing capability (21 FPS). For defect classification:
| Defect Type | ResNet-50 (%) | SE-ResNet (%) |
|---|---|---|
| Gridline | 90.36 | 96.11 |
| Microcracks | 90.11 | 97.08 |
| Shadow | 87.45 | 95.99 |
3. Performance Metrics
Classification accuracy is calculated as:
$$\text{acc} = \frac{TP + TN}{TP + TN + FP + FN}$$
Our framework achieves 97.23% detection accuracy and 96.39% average classification accuracy across defect types, significantly outperforming conventional approaches for solar panel quality inspection.
4. Computational Efficiency
The modified SSD reduces training convergence time by 40% compared to original implementation. The attention mechanisms add minimal computational overhead:
| Component | FLOPs Increase |
|---|---|
| CBAM in SSD | 0.8% |
| SE in ResNet | 1.2% |
This demonstrates the practical viability of our method for industrial solar panel inspection systems requiring both high precision and throughput.
5. Conclusion
The integration of spatial-channel attention mechanisms with multi-scale detection frameworks significantly improves solar panel defect identification accuracy. Future work will focus on optimizing the architecture for embedded deployment in photovoltaic monitoring systems.