In recent years, the global shift toward renewable energy has intensified, with solar power emerging as a pivotal solution due to its simplicity and environmental benefits. Solar panels, the core components of photovoltaic systems, are often exposed to harsh outdoor conditions, leading to defects such as cracks, hotspots, and delamination. These imperfections can significantly reduce energy efficiency and, in severe cases, cause electrical failures or fires. Traditional inspection methods rely heavily on manual labor, which is time-consuming, prone to human error, and inefficient for large-scale solar farms. To address these challenges, this study proposes an improved Single Shot MultiBox Detector (SSD) algorithm tailored for defect detection in solar panels. By integrating advanced neural network architectures and optimization techniques, the enhanced model achieves higher accuracy and faster processing speeds, making it suitable for real-world applications in photovoltaic maintenance and monitoring.
The original SSD algorithm, while efficient for general object detection, faces limitations when applied to photovoltaic panel defects, including suboptimal precision, slower inference times, and issues with missed or false detections. These shortcomings arise from its reliance on the VGG-16 backbone, which has high computational complexity, and its handling of class imbalance during training. Our improvements focus on three key aspects: replacing VGG-16 with an ECA-enhanced ResNet50 network to boost feature extraction capabilities, substituting the Conv7 convolutional layer with an Involution operator for lightweight processing, and incorporating the Focal Loss function to mitigate the imbalance between positive and negative samples. Experimental results demonstrate that the refined algorithm achieves a mean Average Precision (mAP) increase of 4.41 percentage points and a detection speed improvement of 6.55 frames per second (FPS) compared to the baseline SSD, underscoring its efficacy for photovoltaic applications.
To provide context, the standard SSD algorithm utilizes a VGG-16 network as its backbone, modified by converting fully connected layers to convolutional layers and adding extra feature layers for multi-scale detection. The loss function combines localization and confidence losses, as defined by:
$$L(x, c, l, g) = \frac{1}{N} \left( L_{\text{conf}}(x, c) + \alpha L_{\text{loc}}(x, l, g) \right)$$
where \(L_{\text{conf}}\) represents the classification loss, \(L_{\text{loc}}\) denotes the localization loss, \(N\) is the number of matched default boxes, \(x\) indicates the matching between predictions and ground truth, \(c\) is the class confidence, \(l\) refers to the predicted box parameters, and \(g\) represents the ground truth boxes. Although this framework is effective, it struggles with the intricate textures and small defects common in solar panels, necessitating the enhancements detailed in this work.
The cornerstone of our improved SSD algorithm is the integration of a ResNet50 architecture augmented with an Efficient Channel Attention (ECA) module. Residual networks, like ResNet50, alleviate the vanishing gradient problem in deep networks by introducing skip connections, allowing the model to learn residual functions. For an input \(x\), the output \(H(x)\) of a residual block is expressed as \(H(x) = F(x) + x\), where \(F(x)\) is the residual mapping. This enables smoother gradient flow during backpropagation, as shown by the gradient calculation:
$$\frac{\partial \text{loss}}{\partial x_L} = \frac{\partial \text{loss}}{\partial x_L} \cdot \left( 1 + \frac{\partial}{\partial x_L} \sum_{i=l}^{L-1} F(x_i, W_i) \right)$$
Here, \(x_L\) is the output of the \(L\)-th layer, and the term \(1\) ensures that gradients do not vanish, even if the residual term \(\frac{\partial}{\partial x_L} \sum_{i=l}^{L-1} F(x_i, W_i)\) becomes small. To further enhance feature representation, we embed the ECA mechanism after residual blocks. Unlike the Squeeze-and-Excitation (SE) attention, which uses fully connected layers, ECA employs a one-dimensional convolution with an adaptive kernel size \(k\) determined by:
$$k = \frac{\log_2(C) + b}{\gamma}$$
where \(C\) is the number of channels, \(\gamma\) is set to 2, and \(b\) is set to 1. This design reduces computational overhead while emphasizing informative channels, making it ideal for detecting subtle defects in photovoltaic panels. The ECA-ResNet50 backbone replaces the VGG-16 network in SSD, leading to a more robust feature extraction process with fewer parameters.
Another critical modification involves replacing the Conv7 convolutional layer in the original SSD with an Involution operator. Conventional convolution uses spatial-agnostic kernels across channels, which can be computationally intensive. In contrast, Involution employs spatial-specific kernels shared across channels, inverting the design principles of convolution. The kernel generation function \(\Phi(\cdot)\) for Involution is defined as:
$$H_{m,n} = \Phi(X_{m,n}) = W_1 \sigma(W_0 X_{m,n})$$
where \(X_{m,n}\) is the feature vector at position \((m,n)\), \(W_0\) and \(W_1\) are linear transformation matrices, and \(\sigma\) denotes a combination of ReLU and batch normalization. The number of parameters for convolution (\(x_1\)) and Involution (\(x_2\)) can be compared as follows:
$$x_1 = \frac{C^2 K^2 G}{r} + C^2 K^2$$
$$x_2 = K^2 C^2$$
$$\frac{x_1}{x_2} = \frac{G}{rK^2} + \frac{1}{rC}$$
Here, \(C\) is the number of channels, \(K\) is the kernel size, \(G\) is the group number, and \(r\) is the channel reduction ratio. Since \(C\) is typically large in deep networks and \(r \geq 1\), Involution significantly reduces parameter counts, enhancing the model’s efficiency for solar panel defect detection.
To address the class imbalance issue—where background samples (negative) vastly outnumber defect samples (positive)—we integrate the Focal Loss function into the SSD confidence loss. The original confidence loss uses softmax cross-entropy:
$$L_{\text{conf}} = -\sum_{i \in \text{Pos}} x_{ij}^p \log(\hat{c}_i^p) – \sum_{i \in \text{Neg}} \log(\hat{c}_i^0)$$
where \(\hat{c}_i^p\) is the predicted confidence for class \(p\). Focal Loss adds a modulating factor \((1 – p_t)^\gamma\) to down-weight easy examples and focus on hard negatives:
$$\text{FL}(p_t) = -(1 – p_t)^\gamma \log(p_t)$$
where \(p_t\) is the model’s estimated probability for the true class, and \(\gamma \geq 0\) is a tunable focusing parameter. When \(\gamma = 0\), Focal Loss reverts to standard cross-entropy. For our improved algorithm, the confidence loss becomes:
$$L_{\text{conf}} = \sum_{i \in \text{Pos}} \text{FL}(p_i) + \sum_{i \in \text{Neg}} \text{FL}(p_i)$$
This adjustment improves the model’s sensitivity to rare defect instances in photovoltaic panels, reducing false negatives and enhancing overall detection performance.
For experimental validation, we compiled a dataset of 1,800 images of solar panels, including various defects such as cracks and hotspots. The dataset was split into 1,500 training images and 300 test images, annotated in the PASCAL VOC format. The experimental setup involved the following configuration, as summarized in the table below:
| Component | Specification |
|---|---|
| Operating System | Windows 11 |
| CPU | Intel i7-11800H |
| GPU | NVIDIA RTX 3050 |
| Memory | 16 GB |
| Python Version | 3.7.7 |
| Deep Learning Framework | PyTorch 1.7.0 |
| CUDA Version | 11.6 |
We evaluated the model using the mean Average Precision (mAP) and frames per second (FPS) as primary metrics. Precision and recall are defined as:
$$\text{Precision} = \frac{\text{TP}}{\text{TP} + \text{FP}}$$
$$\text{Recall} = \frac{\text{TP}}{\text{TP} + \text{FN}}$$
where TP, FP, and FN denote true positives, false positives, and false negatives, respectively. The average precision (AP) for each class is computed from the precision-recall curve, and mAP is the mean across all classes. To assess computational efficiency, we compared the number of parameters (Params) and floating-point operations (FLOPs) between the original and improved SSD algorithms, as shown in the following table:
| Algorithm | FLOPs (10^9 operations) | Model Size (MB) |
|---|---|---|
| Original SSD | 86.46 | 95.45 |
| Improved SSD | 68.74 | 79.54 |
The reduction in FLOPs and model size underscores the lightweight nature of our approach, achieved through the ECA-ResNet50 backbone and Involution operator. During training, we employed a learning rate schedule starting at 0.0005 and gradually decreasing over epochs, which optimized loss convergence. The loss and learning rate curves demonstrated stable training dynamics, with the model achieving lower loss values compared to baseline methods.
Comparative analysis with other object detection algorithms, including YOLOv3 and YOLOv5, highlighted the superiority of our improved SSD for photovoltaic panel defect detection. The mAP results over training epochs are illustrated in the table below, showing consistent improvements:
| Algorithm | mAP (%) | Detection Speed (FPS) |
|---|---|---|
| Original SSD | 67.95 | 48.36 |
| YOLOv3 | 58.83 | 52.33 |
| YOLOv5 | 62.17 | 55.64 |
| Improved SSD | 72.36 | 54.91 |
Our improved SSD algorithm achieved an mAP of 72.36%, a 4.41 percentage point increase over the original SSD, and a detection speed of 54.91 FPS, which is 6.55 FPS faster. While YOLOv5 slightly outperformed in speed (55.64 FPS), our method excelled in accuracy, surpassing it by 10.19 percentage points in mAP. This makes the enhanced SSD particularly suitable for photovoltaic applications where precision is critical, such as identifying micro-cracks or hotspots in solar panels that could compromise system integrity.
In conclusion, this study presents a robust improved SSD algorithm for defect detection in photovoltaic panels, leveraging ECA-ResNet50, Involution operators, and Focal Loss to address key challenges in accuracy and efficiency. The experimental results confirm significant gains in both mAP and inference speed, demonstrating the model’s potential for real-time solar farm inspections. Future work will explore transfer learning techniques to further reduce data requirements and enhance generalization across diverse photovoltaic environments. By advancing automated detection methods, we contribute to the reliability and sustainability of solar energy systems, supporting global efforts in clean energy adoption.