As non-renewable energy resources continue to deplete, the development and utilization of renewable energy have become a critical direction for global energy strategies. Solar photovoltaic systems have shown broad prospects in the renewable energy sector due to their wide resource distribution, environmental friendliness, and sustainable utilization. However, during long-term operation, solar panels are prone to surface defects, which significantly reduce the photoelectric conversion efficiency. These defects typically manifest as abnormal high-temperature regions in infrared thermal images. In recent years, with the rapid advancement of unmanned aerial vehicle (UAV) inspection technology and infrared thermal imaging technology, image-based defect detection methods for solar panels have become a research focus. Since photovoltaic power stations are often deployed in harsh environments and exposed to complex climatic conditions for extended periods, solar panels are vulnerable to damage, seriously affecting the safe and stable operation of the stations. Therefore, establishing an efficient and reliable defect detection system for photovoltaic modules is of great significance for ensuring the operation and maintenance safety of power stations.
With the continuous growth of installed capacity of photovoltaic power stations, traditional manual inspection methods can no longer meet the operation and maintenance needs of large-scale stations. Intelligent inspection systems based on UAV platforms have gradually become mainstream solutions in the industry. Infrared thermal imaging technology monitors the temperature field distribution of solar panels under operating conditions and uses thermal anomaly regions, known as “hot spots,” as key diagnostic indicators for fault identification. Moreover, infrared images possess the advantages of being independent of natural light, insensitive to ambient light changes, and capable of all-weather imaging. The imaging quality is relatively stable across different seasons, climates, and lighting conditions, and the expression consistency of thermal anomaly regions is strong. Therefore, this technology offers advantages in anti-interference detection, enabling rapid panoramic scanning of photovoltaic arrays at long distances without module disassembly or physical contact, making it particularly suitable for efficient inspection of large-scale photovoltaic power stations. Currently, infrared image defect detection methods for solar panels can be mainly divided into two categories: traditional image processing methods and deep learning-based methods. Traditional image processing methods perform defect detection through operations such as image binarization and morphological processing. For example, an improved selfish herd algorithm optimized edge detection mechanisms to enhance hot spot recognition accuracy, while Otsu threshold segmentation combined with multispectral image fusion technology significantly reduced false positive rates in detection. However, traditional methods generally suffer from inherent defects such as low computational efficiency and poor adaptability, and are easily affected by environmental noise.
In contrast, deep learning-based intelligent detection methods provide new technical research directions for photovoltaic module defect detection due to their powerful feature extraction and pattern recognition capabilities. Research in this field has achieved certain progress: SNCF-Net constructs a multi-scale feature correlation model to effectively improve the detection ability of tiny defects under complex backgrounds; U2-Net generates saliency maps that are fused with infrared images, and then attention mechanisms are incorporated into the detection network to improve detection accuracy; RCAG module uses a residual channel attention mechanism to suppress useless background information during feature fusion and highlight defect features, achieving accurate defect detection. An improved MSRCR algorithm enhances the dark regions of infrared defect images of solar panels and improves recognition accuracy. A homomorphic filtering algorithm based on an improved transfer function improves the detection accuracy of solar panels. CEMP-YOLOv10n achieves a balance between lightweight design and detection speed for infrared images of solar panels. FAA-YOLO effectively improves the detection accuracy of complex defects by fusing information from different scales while ignoring the tendency of such features.
However, in object detection tasks under complex backgrounds, the difficulty of feature extraction often leads to missed detections and false positives, thereby reducing the detection accuracy of the model. To address these issues, I propose an object detection method based on an improved RT-DETR-R50. The specific improvements are reflected in the following three aspects:
- In the backbone network, I adopt partial convolution (PConv) to construct an EMA-PConv structure, replacing the original BasicBlock module with EMA-PConv, which effectively reduces the number of model parameters.
- I use a multi-scale feature adaptive pyramid network (MFAPN) for feature fusion. The fusion structure provides richer feature support for small target detection, thereby improving detection performance.
- I propose a multi-path downsampling enhancement module (MDEM) to replace the maximum pooling layer of RT-DETR-R50, solving the problem of semantic information loss and improving the model’s feature extraction capability.
The improved model is denoted as MPMA-DETR. In the following sections, I will detail the design of each module, present extensive experiments on an aerial infrared solar panel dataset, and analyze the results.
Related Work
The RT-DETR (Real-Time Detection Transformer) model adopts a three-stage cascade structure consisting of a backbone, a hybrid encoder, and a Transformer decoder. The model extracts the last three stages of hierarchical features from the backbone as encoder inputs. In the encoding stage, feature enhancement is achieved through an attention-based intra-scale feature interaction module (AIFI), combined with cross-scale feature fusion technology to convert multi-scale features into serialized image feature representations. The encoder output employs an IoU-aware query selection mechanism to dynamically select the most representative image features as the initial object query set for the decoder. The decoder optimizes object query vectors through multiple iterations and uses an auxiliary prediction head to collaboratively generate bounding box coordinates and corresponding confidence scores.
However, the original RT-DETR-R50 still faces challenges in detecting defects in solar panels under complex backgrounds. The standard max pooling downsampling may overlook fine-grained object features, especially in infrared images where subtle defect regions are highly similar to background thermal interference. Moreover, the convolution operations in the backbone introduce parameter redundancy and computational burden. To address these issues, I propose three novel modules: EMA-PConv, MDEM, and MFAPN, which are integrated into the RT-DETR-R50 framework to form the MPMA-DETR model.
Methodology
In this section, I detail the design of the three proposed modules and their integration into the overall architecture.
EMA-PConv Module
To alleviate the parameter redundancy and computational explosion caused by ordinary convolutions in the backbone while ensuring efficient feature extraction, I introduce partial convolution (PConv) from FasterNet. The core advantage of PConv is that it performs convolution only on a portion of the channels of the input feature map. Compared to traditional convolution, the floating-point operations of PConv are only 1/16 of the original, significantly reducing the overall computational complexity. Based on this, I improve the basic residual block (BasicBlock) and propose a reparameterized lightweight convolution module called EMA-PConv.
EMA-PConv performs a 3×3 convolution operation on only 1/4 of the channels of the input feature map, leaving the remaining channels unchanged. This design is inspired by the low-rank approximation theory: image features have high redundancy along the channel dimension. By performing convolution on only a portion of the channels, key features can be preserved while redundant computations are reduced. Furthermore, the retained un-convolved channels maintain the original semantic structural integrity, facilitating subsequent attention mechanisms to more easily capture important target regions. To further enhance the representational power, I incorporate the efficient multi-head attention (EMA) mechanism, which dynamically learns channel importance weights to guide the model to focus on more discriminative detection regions. The ReLU activation function and layer normalization are used to alleviate gradient issues and improve feature extraction capability. Additionally, a DropPath regularization strategy is introduced in the residual path to prevent overfitting when dealing with complex samples.
The operation of EMA-PConv can be expressed as:
$$
Z = \text{EMA}\left( \text{ReLU}\left( \text{LayerNorm}\left( \text{PConv}_{3 \times 3}(Z_{C/4}) + Z_{3C/4} \right) \right) \right)
$$
where \( Z_{C/4} \) denotes the feature vector with 1/4 of the channels, and \( \text{PConv}_{3 \times 3}(\cdot) \) denotes the 3×3 partial convolution.
The EMA module works as follows: First, the input feature map is divided into \( G \) groups along the channel dimension, and attention computation is performed only within each group to reduce computational overhead. The spatial branch performs global average pooling along the height and width directions for each group, obtains spatial attention weights through 1×1 convolution, and multiplies them element-wise to guide the network to focus on critical regions. The local branch uses 3×3 convolution to extract fine-grained local information. Subsequently, the two branches generate attention weights through global average pooling and softmax operations, and perform cross-product and fusion (MatMul) to obtain the final output. The EMA operation can be expressed as:
$$
Y = X_{\text{Group}} \times \sigma\left\{ \text{MatMul}\left( \text{Softmax}\left( \text{AvgPool}(X_1) \right), X_2 \right) + \text{MatMul}\left( \text{Softmax}\left( \text{AvgPool}(X_2) \right), X_1 \right) \right\}
$$
where \( X_{\text{Group}} \) is the grouped representation of feature map \( X \) with shape \( B \times C \times H \times W \); \( X_1 \) and \( X_2 \) are the feature maps after convolution processing in the spatial and local branches respectively; AvgPool denotes global average pooling for generating spatial global features; Softmax is used for feature fusion to generate normalized weights; MatMul performs matrix multiplication to combine different feature weights; and \( \sigma \) is the sigmoid activation function for generating final weights.
The EMA-PConv module effectively reduces model parameters while maintaining strong feature extraction capability. I conducted ablation experiments to study the impact of the channel convolution ratio (1/8, 1/4, 1/2) on model performance. The results are summarized in Table 1.
| Experiment | Channel Ratio | P (%) | R (%) | mAP50 (%) | Parameters (×107) | GFLOPs | FPS |
|---|---|---|---|---|---|---|---|
| I | 1/2 | 72.1 | 67.6 | 73.1 | 1.81 | 54.3 | 149 |
| II | 1/4 | 72.3 | 68.7 | 73.3 | 1.69 | 51.5 | 151 |
| III | 1/8 | 69.2 | 67.4 | 72.9 | 1.66 | 50.7 | 154 |
From Table 1, I observe that the 1/4 channel ratio achieves the best mAP50 (73.3%) while maintaining a good balance between parameters (1.69×107), GFLOPs (51.5), and FPS (151). Therefore, I adopt this configuration as the default in all subsequent experiments.
Multi-Path Downsampling Enhancement Module (MDEM)
The original RT-DETR model uses max pooling for downsampling, which may ignore fine-grained object features when extracting global context. This is particularly problematic in infrared solar panel images where subtle defect regions are highly similar to background thermal interference. To address this, I design the Multi-Path Downsampling Enhancement Module (MDEM) to enhance defect target information expression while suppressing background interference.
MDEM performs downsampling through three parallel paths: strided convolution, depthwise separable convolution, and max pooling. The strided convolution strengthens spatial gradient changes at target edges, the depthwise convolution focuses on local difference structures, and the max pooling suppresses high-frequency noise background. The outputs of these three paths are fused to produce a feature map that not only preserves target structural information but also effectively weakens background interference responses.
MDEM adopts a two-stage process for efficient feature extraction and dimensionality reduction. First, a 7×7 convolution is applied to the input feature vector to extract initial features, compressing the output channels to C/4 while maintaining the original spatial resolution. This gives the initial feature vector \( X_{\text{init}} \). In the first stage, the module runs two parallel paths: a strided path using strided convolution to achieve 2× downsampling and expand channels to C/2, and a depth path using depthwise separable convolution for local feature extraction and channel expansion to C/2. The outputs are concatenated, followed by a 1×1 convolution and batch normalization. Max pooling is omitted at this stage because shallow feature maps often contain significant noise, and using max pooling would retain noise interference.
The process in the first stage can be expressed as:
$$
X_{\text{init}} = \text{Conv}_{7 \times 7}(X) \in \mathbb{R}^{C/4 \times H \times W}
$$
$$
F = \text{Conv}_{1 \times 1}\left( \text{Concat}\left( \text{Conv}_{3 \times 3}(X_{\text{init}}), \text{DWConv}_{3 \times 3}(X_{\text{init}}) \right) \right)
$$
In the second stage, a three-modal processing mechanism is introduced: the depth path continues to use depthwise separable convolution to expand channels to C; the strided path performs downsampling while preserving original information; and the newly added pooling path extracts high-response region features via max pooling. The three paths are concatenated and fused using a 1×1 convolution, which integrates channel-dimension information while implicitly modeling channel-wise weights through the training process. This implicit fusion strategy avoids additional parameter redundancy and model size increase.
The second stage operation is:
$$
V = \text{BN}\left( \text{Conv}_{1 \times 1}\left( \text{Concat}\begin{pmatrix} \text{Conv}_{3 \times 3}(F) \\ \text{DWConv}_{3 \times 3}(F) \\ \text{MaxPool}(F) \end{pmatrix} \right) \right)
$$
MDEM effectively enhances feature extraction by combining multiple downsampling strategies, improving the model’s ability to discriminate defects in complex backgrounds.
Multi-Scale Feature Adaptive Pyramid Network (MFAPN)
In infrared solar panel defect detection, the interference caused by different resolution images and varying defect sizes poses challenges. Traditional object detection methods struggle to balance the capture of fine details and the expression of global semantic information, leading to low detection accuracy. To address this, I propose the Multi-Scale Feature Adaptive Pyramid Network (MFAPN), which strengthens the model’s adaptability to multi-scale features through collaborative design of feature extraction and fusion.
First, input feature maps of different resolutions (denoted as L, M, S for large, medium, and small receptive field features) are resized to the same spatial resolution and concatenated along the channel dimension. I adopt a semantic-intensity-first strategy: the higher-level, larger-receptive-field features \( X_L \) are concatenated first to emphasize global semantic dominance, followed by mid-scale features \( X_M \), and finally local structure features \( X_S \). This order is inspired by semantic-guided fusion mechanisms, using stable high-level features as semantic support and gradually incorporating low-level details to alleviate multi-scale semantic inconsistency and improve feature fusion discriminability.
Specifically, the large feature map \( X_L \) is compressed to the size of \( X_M \) using both adaptive max pooling and adaptive average pooling, and the two pooled maps are summed to fuse context information from different pooling strategies. The small feature map \( X_S \) is upsampled to the size of \( X_M \) using nearest-neighbor interpolation. The three processed feature maps are then concatenated along the channel dimension to form a multi-scale feature vector with consistent feature dimensions. This approach retains the spatial semantic information of each scale without adding extra computational cost.
This process is formulated as:
$$
X’_L = \text{MaxPool}(X_L) + \text{AdaPool}(X_L)
$$
$$
X’_S = \text{Upsample}(X_S)
$$
$$
Y = \text{Concat}(X’_L, X_M, X’_S)
$$
Inspired by sequential modeling, I further explore the correlation between features from different scales by stacking the multi-scale feature vector along the scale dimension to construct a 5D tensor \( F \in \mathbb{R}^{B \times C \times 3 \times H \times W} \), where the third dimension represents the scale sequence. A 3D convolution kernel of size \(1 \times 1 \times 1\) is applied along the scale dimension for feature interaction, simulating the temporal dependency capture mechanism in sequence modeling. After batch normalization and ReLU activation, a \(3 \times 1 \times 1\) 3D max pooling compresses the scale dimension and outputs the final feature, which contains core information from all scales, preserving details while fusing global semantics.
The operations are:
$$
F = \text{Reshape}(Y) \in \mathbb{R}^{B \times C \times 3 \times H \times W}
$$
$$
F_{\text{out}} = \text{MaxPool}_{3 \times 1 \times 1}\left( \text{ReLU}\left( \text{BN}\left( \text{Conv}_{1 \times 1 \times 1}(F) \right) \right) \right)
$$
MFAPN strengthens information interaction across different feature layers, enriching semantic features and improving detection performance, especially for small defects.
I conducted ablation experiments on the concatenation order of different resolution features in MFAPN. The results are shown in Table 2.
| Experiment | Concatenation Order | P (%) | R (%) | mAP50 (%) | Parameters (×107) | GFLOPs | FPS |
|---|---|---|---|---|---|---|---|
| I | L, M, S | 73.3 | 80.4 | 76.9 | 4.34 | 142.2 | 88 |
| II | L, S, M | 73.1 | 75.1 | 76.6 | 4.34 | 142.2 | 88 |
| III | M, L, S | 72.8 | 76.0 | 74.3 | 4.31 | 142.2 | 87 |
| IV | M, S, L | 72.9 | 73.6 | 75.2 | 4.31 | 142.2 | 87 |
| V | S, L, M | 71.5 | 78.5 | 73.6 | 4.31 | 142.2 | 88 |
| VI | S, M, L | 68.8 | 78.0 | 74.1 | 4.31 | 142.2 | 88 |
The results demonstrate that the order (L, M, S) yields the best mAP50 (76.9%), confirming that semantic-intensity-first concatenation effectively improves multi-scale feature fusion. Therefore, I adopt this order as the default setting.
Overall Architecture of MPMA-DETR
The improved RT-DETR structure, named MPMA-DETR, integrates the three proposed modules. The EMA-PConv modules replace the BasicBlock modules in the backbone, reducing parameters and computation. The MDEM module replaces the max pooling layers to enhance downsampling. The MFAPN is applied in the feature fusion stage to improve multi-scale feature representation. The encoder and decoder remain similar to the original RT-DETR but benefit from the enhanced features. The complete model is designed to achieve high detection accuracy with reduced computational cost.
The above image illustrates an example of solar panels inspected in a real-world scenario, highlighting the importance of robust defect detection methods.
Experimental Results and Analysis
Experimental Setup
All experiments were conducted on an NVIDIA 4080 GPU with Python 3.10 and PyTorch 1.12.1. For fair comparison, no pre-trained weights were used for any model. The batch size was set to 8, the number of training epochs was 200, with early stopping patience of 50 epochs. The initial learning rate was 0.0001. All other hyperparameters were kept at their default values.
Dataset
The dataset used in this study comes from a photovoltaic power plant, where infrared images were captured by UAVs equipped with infrared cameras from different power units. A total of 3,694 infrared images with a resolution of 640×640 were collected. The dataset contains four types of solar panel defects: diode short (hotspot), PID effect (golden-spot), minor crack (light-golden-spot), and shading (shadow). The images were annotated using the Labelme tool and divided into training set (2,586 images), validation set (382 images), and test set (726 images), in a ratio of approximately 7:1:2.
Evaluation Metrics
I use precision (P), recall (R), mean average precision (mAP50 and mAP50@95), number of parameters, floating-point operations (GFLOPs), weight size (MB), and inference speed (FPS) as evaluation metrics.
Precision and recall are defined as:
$$
P = \frac{TP}{TP + FP}, \quad R = \frac{TP}{TP + FN}
$$
where TP, FP, and FN denote true positives, false positives, and false negatives, respectively. The average precision (AP) for a class is the area under the precision-recall curve, and mAP is the mean of APs over all classes.
Ablation Study on MPMA-DETR
To validate the effectiveness of each proposed module, I conducted ablation experiments based on the original RT-DETR-R50 model. The results are summarized in Table 3.
| Experiment | EMA-PConv | MFAPN | MDEM | P (%) | R (%) | mAP50 (%) | Parameters (×107) | GFLOPs | FPS |
|---|---|---|---|---|---|---|---|---|---|
| I (Baseline) | — | — | — | 73.7 | 75.9 | 75.0 | 4.30 | 129.6 | 100 |
| II | √ | — | — | 72.3 | 68.7 | 73.3 | 1.69 | 51.5 | 151 |
| III | — | √ | — | 73.3 | 80.4 | 76.9 | 4.34 | 142.2 | 88 |
| IV | — | — | √ | 74.5 | 83.4 | 76.4 | 4.31 | 50.2 | 154 |
| V | √ | √ | — | 75.1 | 75.3 | 76.6 | 3.57 | 131.2 | 72 |
| VI | √ | — | √ | 75.1 | 73.7 | 75.5 | 3.47 | 47.5 | 70 |
| VII | — | √ | √ | 75.7 | 76.4 | 75.8 | 4.32 | 53.4 | 149 |
| VIII (Ours) | √ | √ | √ | 76.3 | 79.6 | 78.3 | 3.58 | 50.6 | 135 |
From Table 3, I observe the following: (1) EMA-PConv alone (Experiment II) reduces parameters by 60.7% but decreases mAP50 to 73.3%, indicating that lightweight modules may suffer from feature representation degradation when used alone. (2) Adding MFAPN (Experiment III) improves mAP50 to 76.9%, showing its effectiveness in multi-scale fusion. (3) Adding MDEM (Experiment IV) improves recall to 83.4% and mAP50 to 76.4%. (4) Combining all three modules (Experiment VIII) achieves the highest mAP50 of 78.3% and mAP50@95 of 45.5% (not shown in table, but reported later), with a 16.8% reduction in parameters compared to the baseline. The overall improvement of 3.3 percentage points in mAP50 demonstrates the synergy of the three proposed components.
Comparison with State-of-the-Art Methods
I compared MPMA-DETR with several popular object detection algorithms: YOLOv10x, YOLOv11, YOLOv12, Faster-RCNN, RT-DETR-R18, RT-DETR-R101, and DE-DETR. All models were trained under the same conditions on the same dataset. The results are shown in Table 4.
| Model | P (%) | R (%) | mAP50 (%) | mAP50@95 (%) | Parameters (×106) | GFLOPs | Weight (MB) | FPS |
|---|---|---|---|---|---|---|---|---|
| RT-DETR-R50 (Baseline) | 73.7 | 75.9 | 75.0 | 44.3 | 43.0 | 129.6 | 86.1 | 100 |
| RT-DETR-R18 | 72.4 | 71.9 | 71.5 | 41.6 | 20.0 | 57.0 | 40.5 | 192 |
| RT-DETR-R101 | 73.4 | 82.8 | 74.4 | 44.4 | 74.0 | 247.1 | 153.8 | 71 |
| Faster-RCNN | 68.3 | 71.3 | 71.9 | 40.3 | 41.3 | 133.9 | 161.9 | 42 |
| YOLOv10x | 75.3 | 68.4 | 75.3 | 44.8 | 31.6 | 169.8 | 64.1 | 161 |
| YOLOv11 | 75.2 | 79.4 | 76.4 | 45.1 | 2.6 | 6.3 | 5.5 | 435 |
| YOLOv12 | 73.6 | 74.7 | 73.4 | 42.0 | 2.7 | 6.4 | 5.6 | 357 |
| DE-DETR | 71.2 | 74.3 | 72.2 | 41.9 | 40.3 | 86.0 | 175.0 | 167 |
| MPMA-DETR (Ours) | 76.3 | 79.6 | 78.3 | 45.5 | 35.8 | 50.6 | 76.1 | 135 |
From Table 4, MPMA-DETR achieves the highest mAP50 (78.3%) and mAP50@95 (45.5%) among all compared methods. It also obtains the best precision (76.3%) and competitive recall (79.6%). In terms of model complexity, MPMA-DETR has 35.8 million parameters, which is significantly lower than RT-DETR-R101 (74.0 million) and YOLOv10x (31.6 million? Actually YOLOv10x has 31.6 million, but ours is slightly higher; however, our GFLOPs is only 50.6, much lower than YOLOv10x’s 169.8). The inference speed of 135 FPS is faster than RT-DETR-R101 (71 FPS) and Faster-RCNN (42 FPS), while being slower than lightweight models like YOLOv11 (435 FPS). Considering the trade-off between accuracy and efficiency, MPMA-DETR provides the best overall performance, especially in challenging scenarios with complex backgrounds.
Visualization Analysis
To intuitively demonstrate the detection capability of MPMA-DETR, I visualize the Grad-CAM heatmaps generated from the last multi-head attention output layer of the Transformer decoder. This layer is close to the detection head and fully reflects the model’s attention distribution during the object discrimination stage. The heatmaps are generated by backpropagating the gradient information to obtain the corresponding weights of the activation maps and then performing a weighted summation.
The qualitative comparison between the baseline RT-DETR-R50 and MPMA-DETR (with MDEM) shows that the improved model focuses more precisely on defect regions while effectively suppressing background noise. For example, in images where the baseline model’s heatmap responses are scattered over non-target areas (e.g., background thermal patterns), MPMA-DETR’s heatmaps are highly concentrated on the actual defects. In cases of small or subtle defects, the baseline model often fails to highlight them, whereas MPMA-DETR clearly captures them. This confirms that the MDEM module enhances the model’s ability to extract discriminative features and suppress irrelevant background information.
Furthermore, comparing MPMA-DETR with the baseline in images containing multiple defects demonstrates that the improved model consistently produces sharper and more localized attention maps, leading to fewer missed detections. The integration of EMA-PConv and MFAPN further contributes to the stability and richness of the feature representations.
Conclusion
In this paper, I have proposed MPMA-DETR, an improved infrared defect detection method for solar panels based on the RT-DETR-R50 framework. The method incorporates three novel modules: (1) EMA-PConv, which leverages partial convolution and efficient multi-head attention to reduce model parameters and computational cost while preserving feature quality; (2) MDEM, a multi-path downsampling enhancement module that captures both global and local information to strengthen feature extraction and suppress background interference; and (3) MFAPN, a multi-scale feature adaptive pyramid network that enhances information interaction among different feature scales, enriching semantic feature representation.
Extensive experiments on an aerial infrared solar panel dataset demonstrate that MPMA-DETR achieves significant improvements over the baseline RT-DETR-R50. The mAP50 increased by 3.3 percentage points to 78.3%, and mAP50@95 increased by 1.2 percentage points to 45.5%. The model also reduces parameters by 16.8% and achieves a favorable balance between accuracy and computational efficiency. Comparisons with state-of-the-art detectors, including YOLOv10x, YOLOv11, YOLOv12, Faster-RCNN, RT-DETR-R101, and DE-DETR, show that MPMA-DETR outperforms all of them in terms of mAP50 and precision, while maintaining competitive inference speed.
The proposed method effectively addresses the challenge of detecting defects in solar panels under complex backgrounds, providing a robust solution for intelligent photovoltaic inspection. Future work could explore further lightweight design for edge deployment and extend the approach to other types of photovoltaic component defects.