Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 8145 (2026)
769
Metrics details
The accumulation of foreign objects on photovoltaic panels can significantly reduce their photoelectric conversion efficiency and may cause hotspot effects, leading to module aging and permanent damage, which threatens the safe and economic operation of PV power plants. To address challenges in foreign object detection, such as low image contrast, complex morphology, and sample imbalance, this study proposes the DHLNet model. The model is based on an innovative Dual-Flow Feature Pyramid Network, which improves the accuracy of small object detection in complex backgrounds through cross-stage information interaction and multi-level feature reconstruction, while maintaining a lightweight structure. To further enhance performance, a high-frequency enhancement module is integrated into the backbone network to increase the model’s sensitivity to targets with blurred edges. In addition, a large-separable-kernel attention mechanism is introduced to enhance the extraction capability of key features. Experimental results show that DHLNet achieved mAP@0.5 and mAP@0.5:0.95 scores of 79.9% and 58.8%, respectively, representing improvements of 3.1% and 2.2% compared with the original YOLO11n algorithm. The F1 score reached 0.761, an improvement of 1.3%. The model also supports real-time inference, making it suitable for on-site detection and online monitoring. These results verify the effectiveness and practical application potential of DHLNet in automated foreign object detection on photovoltaic panels.
With the global energy structure shifting toward low-carbon and renewable directions, photovoltaic (PV) systems have been widely applied in scenarios such as power plants, distributed systems, and building-integrated photovoltaics (BIPV). However, during actual operation, PV modules are easily obstructed by foreign objects such as bird droppings, leaves, dust, and snow, leading to reduced light absorption efficiency, power attenuation, hotspot effects, and module aging1,2. Accurately identifying the types and locations of these faults remains a challenge. Traditional manual inspection suffers from high cost, low efficiency, and a high rate of missed detections, making it difficult to meet the operation and maintenance requirements of large-scale power plants. Therefore, intelligent detection methods based on computer vision have gradually become mainstream solutions, demonstrating advantages in efficiency and automation3. Nevertheless, challenges still exist under conditions such as complex lighting, diverse defects, and background interference. Building a high-precision and robust automatic detection framework for PV panels is of great significance for improving system reliability, reducing maintenance costs, and promoting the development of green energy.
To ensure the efficient operation and extend the service life of photovoltaic systems, advanced fault detection and diagnosis (FDD) technologies are particularly crucial. Traditional methods suffer from low efficiency, strong dependence on manual labor, and high false detection rates, making them inadequate for the large-scale requirements of modern PV power plants. The development of artificial intelligence (AI) has provided new ideas for fast and accurate PV fault detection. Voutsinas et al.4 proposed an algorithm based on logistic regression and cross-validation, which can achieve early string-level fault identification on the DC side. Natarajan et al.5 used support vector machines (SVM) combined with thermal imaging to classify and identify defects such as cracks, hotspots, and dust accumulation. Herraiz et al.6 developed a hotspot detection model based on region convolutional neural networks (R-CNN), achieving module-level fault localization through the fusion of multi-source images and telemetry data. These studies indicate that deep learning-based visual detection methods show significant potential in PV fault identification, but they are still limited by insufficient data diversity, low model generalization, and high computational complexity.
Although progress has been made in foreign object detection on photovoltaic modules, multiple challenges still exist under complex scenarios. These include large variations in object scale, blurred edges, and complex backgrounds, which easily lead to false detections and missed detections. Existing models lack sufficient sensitivity to small targets and weak-texture defects, with limited feature representation capability. Complex illumination and morphological diversity weaken global perception and reduce detection robustness. High-precision models often incur high computational costs, making it difficult to balance real-time performance and lightweight design. To address these issues, this paper proposes a model named DHLNet, as described below:
To enhance the model’s localization and recognition capabilities, we redesigned the feature fusion network and proposed a novel Dual-Flow Feature Pyramid Network (DFFPN). DFFPN injects the original backbone network features into the neck, enabling more effective fusion of low-level detailed features with high-level semantic information, shortening the information propagation path, and reducing the information loss commonly found in traditional fusion schemes. Therefore, this network provides more discriminative features for subsequent detection tasks.
To address the limitations of the model in feature extraction and representation for foreign object detection on photovoltaic modules, particularly the low sensitivity to small targets and subtle defects, we reconstructed the backbone network of YOLO11n. We optimized the C3K2 module using a computationally efficient High-Frequency Enhancement Residual Block (HFERB) and introduced a novel Cross-Knowledge Hybrid Block (CKHB).
To address the limitation of the model’s global perception—which reduces its detection performance for foreign objects with blurred edges or diverse shapes. We embedded a Large-Separable-Kernel Attention (LSKA) mechanism after the C2PSA module in the neck. This mechanism effectively models global contextual information and expands the receptive field, thereby enhancing the model’s feature-capturing ability and overall detection performance.
Foreign object detection technology for photovoltaic panels has clearly shifted from traditional signal processing methods to advanced deep learning techniques. In the early stages, traditional methods played an important role in detection by directly reflecting the operating state of PV panels7, and they were not affected by image acquisition conditions. These methods used techniques such as Fourier transform and wavelet transform to analyze variations in electrical signals8, enabling foreign object detection and supporting PV panel maintenance. However, with the rapid development of the PV industry, the demand for higher detection accuracy, efficiency, and automation has continued to grow. Consequently, the limitations of these traditional methods, such as insufficient sensitivity to small foreign objects and the inconvenience of requiring electrical connections—have become increasingly evident.
In response, image processing technology has gained widespread application due to its efficiency, accuracy, and automation capabilities9,10. This approach involves using devices such as unmanned aerial vehicles (UAVs) to capture images, followed by image preprocessing to remove noise and enhance contrast. Subsequently, key information such as shape, color, and texture is extracted to enable detection through computer vision techniques11. With the development of artificial intelligence, deep learning-based technologies have emerged as a more promising direction. These methods rely on deep neural networks12 to locate foreign objects on PV panels. By learning patterns and features from large-scale datasets, they can automatically detect and identify surface deposits on PV panels13, achieving higher precision and greater automation. This transformation is not merely a technological upgrade but an inevitable evolution to meet the growing demand for efficient maintenance and maximized power generation in PV systems. It represents a shift from basic signal analysis to highly intelligent and precise detection methods.
In the early research on foreign object detection for photovoltaic panels, signal processing methods represented the most fundamental and dominant technical approach. The basic idea is to analyze variations in the output electrical signals, temperature signals, or illumination signals of PV modules to identify performance anomalies caused by obstruction, contamination, or damage. These methods are typically based on electrical models and physical parameter measurements, using changes in time-domain and frequency-domain characteristics of signals such as power, voltage, and current to infer system conditions.
Among them, Wei et al.14 developed a sensor fusion system for detecting shading caused by foreign object obstruction. The system integrates temperature, illumination, and voltage/current sensors with an embedded processing unit, determining foreign object coverage by analyzing deviations between actual power output and the ideal model. Similarly, Srivastava et al.15 proposed a hybrid anomaly detection method that combines sensor data with maximum power point tracking (MPPT) features. This method utilizes voltage sensors for multi-string fault classification and was validated on the OPAL-RT platform, demonstrating robust fault localization even in cases of sensor failure. In another study, Prasshanth16 proposed a hybrid diagnostic framework that integrates multimodal features with rough set theory. By combining texture and frequency-domain features and employing a rough set classifier for decision optimization, the framework achieved high accuracy while reducing data requirements.
In recent years, with the advancement of convolutional neural networks (CNNs) and deep learning technologies, methods using deep learning for PV panel detection have gradually become mainstream and achieved remarkable results. These methods are commonly divided into conventional deep learning approaches, two-stage object detection models represented by R-CNN17, Fast R-CNN18, and Faster R-CNN19, and one-stage object detection models represented by SSD20 and YOLO21.
In conventional deep learning approaches, Guo et al.22 designed a multi-channel one-dimensional convolutional neural network combined with LSTM and AdaBoost ensemble models based on model fusion, achieving a fault classification accuracy of 96.4%. Vlaminck et al.23 proposed a PV module anomaly detection method based on UAV imagery and region convolutional neural networks, which improved the efficiency and accuracy of large-scale PV plant inspection through module localization and defect identification, outperforming existing methods. Dong et al.24 proposed the ISEE system based on edge computing, which captures real-time infrared images of PV panels through cameras and performs defect detection on edge devices using CNNs, effectively addressing the low efficiency and high labor cost issues of traditional offline detection methods.
In two-stage object detection methods, Pierdicca et al.25 proposed the solAIr approach based on Mask R-CNN, which utilizes UAV thermal infrared images to automatically identify abnormal PV cells, achieving excellent performance in terms of IoU and Dice coefficient. Blázquez Folch et al.26 combined UAV infrared imagery with deep neural networks, employing Faster R-CNN along with data augmentation, active learning, and isotonic risk control, achieving a defect detection rate of 60.8%. Salehpour et al.27 proposed a two-stage fault diagnosis and severity assessment scheme based on a deep Q-network (DQN). This method uses reinforcement learning to dynamically select features, offering high sensitivity in lightweight recognition; however, the fusion of temporal and spatial features and the precision of foreign object localization still require improvement.
Compared with two-stage models, one-stage algorithms are faster and more accurate28. Among them, Hong et al.29 proposed an intelligent detection framework that integrates infrared and visible light, utilizing YOLOv5 and ResNet to achieve module segmentation and defect detection, thereby improving recognition accuracy under complex environments. In addition, Sun et al.30 developed a detection model based on the PP-YOLO algorithm, which integrates the architectural advantages of YOLOv5, adopts the lightweight PP-LCNet backbone network and H-Swish activation function, and incorporates a coordinate attention (CA) mechanism to enhance feature perception capability, achieving good performance. Shamta et al.31 proposed a hotspot fault detection framework based on the YOLOv8 segmentation model, which improves generalization through data augmentation and achieves high-precision localization, but shows limited adaptability to foreign objects with diverse shapes.
Although the aforementioned studies have made significant progress in foreign object detection on photovoltaic panels, different methods still vary in detection accuracy, robustness, and engineering applicability. To systematically compare the performance characteristics of various approaches, this paper contrasts the main advantages and disadvantages of traditional signal processing methods, two-stage detection methods, and one-stage detection methods, as shown in Table 1.
As shown in the table, traditional signal processing methods mainly rely on electrical signal features for detection. They offer the advantages of simple implementation and fast response but fail to meet the growing demands for higher spatial resolution and intelligent analysis. With the advancement of deep learning, two-stage detection models have demonstrated outstanding performance in localization accuracy; however, their high computational cost and limited real-time capability make them unsuitable for large-scale UAV inspection scenarios. In contrast, one-stage detection methods have become the mainstream approach due to their efficiency and lightweight design, yet they still face significant challenges in small object detection, adaptation to complex illumination, and recognition of blurred edges.
To overcome the problems of missed and false detections that commonly occur in detecting small-sized, low-contrast, and irregularly shaped foreign objects on PV panel surfaces under complex scenarios, this paper proposes a model named DHLNet. The overall structure of the model is shown in Fig. 1.
The structure of DHLNet model.
Based on the traditional PAN framework, DHLNet designs a Dual-Flow Feature Pyramid Network (DFFPN) to construct multi-dimensional feature propagation paths and cross-stage fusion mechanisms. This enables efficient interaction and collaborative enhancement between shallow spatial details and deep semantic information, thereby significantly improving the representational ability and discriminative power of fused features. Meanwhile, the original feature extraction module in the YOLO network is redesigned into a High-Frequency Enhancement Residual Block. By adjusting the convolution kernel size, introducing depthwise separable convolution, and optimizing the residual path, the model enhances its responsiveness to object edges and fine-grained structures while maintaining a lightweight architecture.
In addition, after the C2PSA module in the neck network, DHLNet further integrates a Large-Selective-Kernel Attention (LSKA) mechanism. This module performs weighted fusion of features from large and small convolution kernels, preserving local details while enhancing global context modeling capability, thereby achieving more precise feature recalibration.
Built on the PAN framework, we propose a Dual-Flow Feature Pyramid Network (DFFPN). DFFPN constructs multi-dimensional feature paths and cross-stage fusion to enable efficient interaction between shallow spatial detail and deep semantics. This design improves the representation and discriminability of the fused features. We redesign the C3K2 block as HFERB. By resizing kernels, introducing depthwise separable convolutions, and optimizing the residual pathway, the design increases sensitivity to edges and fine structures while keeping the model lightweight. Additionally, LSKA module is introduced after C2PSA in the neck stage. LSKA integrates responses from large and small kernels through weighted fusion, preserves local detail, strengthens global-context modeling, and produces more accurate feature recalibration. To the best of our knowledge, DHLNet is the first PAN-based detector to jointly introduce a dual-flow FPN, an edge-aware residual block, and large selective-kernel attention in the neck, forming a lightweight yet highly discriminative architecture.
As the depth of convolutional neural networks (CNNs) increases, the semantic representation capability of the model is continuously enhanced, but the spatial resolution decreases layer by layer, and fine-grained features are gradually compressed into channel information. This leads to the weakening of small object and detailed information during the feature fusion process. Feature fusion structures such as FPN and PAN alleviate this problem to some extent through multi-scale feature utilization; however, their static fusion strategies and fixed connection patterns limit the collaborative modeling capability between layers. The BiFPN structure of EfficientDet32 improves fusion efficiency by introducing learnable weights, but it still relies on pixel-wise weighted summation, making it difficult to adequately model complex contextual relationships.
To address the above issues, this paper proposes a Dual-Flow Feature Pyramid Network (DFFPN), as shown in Fig. 2. This structure enhances dynamic fusion and complementary modeling capabilities between different semantic layers through cross-stage feature interaction and multi-level reconstruction mechanisms, thereby improving detection robustness and accuracy under complex scenarios.
Comparison of neck structures between PAN and DFFPN.
FPN enhances high-level semantics through a top-down pathway combined with lateral connections, while PAN introduces a bottom-up pathway to improve localization capability. BiFPN improves fusion efficiency by introducing learnable weights, whereas DFFPN further optimizes information flow through a multi-level reconstruction mechanism. The information flow in DFFPN can be expressed as follows.
The feature layers extracted by the backbone network at different scales are denoted as (P_3), (P_4), and (P_5), where (P_3^{in} in mathbb {R}^{C_3 times H_3 times W_3}) represents the low-level detailed features with the highest resolution, and (P_5^{in} in mathbb {R}^{C_5 times H_5 times W_5}) represents the high-level semantic features with the lowest resolution. The top-down flow process can be described as Eq. (1).
Here, (F_i^{TD}) represents the top-down feature from the previous layer; Upsample() denotes the upsampling operation to match the scale of the i-th layer, typically implemented using nearest-neighbor or bilinear interpolation; (Conv_{TD}) is used for channel transformation and smoothing after upsampling; (Conv_L) denotes convolutional adjustment for the current layer’s input feature; (oplus) indicates element-wise addition or concatenation followed by convolution; (P_i^{in}) represents the input feature of the current layer; and (F_{i+1}^{TD}) refers to the upper-layer feature of the (i+1) stage.
The bottom-up flow is defined as shown in Eq. (2):
Here, (F_i^{BU}) represents the bottom-up feature from the next layer, Downsample() denotes the downsampling operation, (Conv_{BU}) represents convolution after downsampling to adjust the number of channels, and (F_{i-1}^{BU}) indicates the lower-level feature of the (i-1) layer.
To enhance multi-scale fusion among features, DFFPN introduces cross-layer connections, as shown in Eq. (3):
Here, (F_i^{KL}) represents the cross-layer connection feature of the i-th layer, (Conv_{KL}) denotes the convolutional dimensionality reduction applied to the input feature of the j layer, (Align_{i}()) adjusts the feature to the same size as the i-th layer through interpolation or downsampling, and (P_i^{in}) represents the input feature.
By combining the features from the top-down flow, bottom-up flow, and cross-layer flow, the final output of each layer is obtained as Eq.(4):
where, (Conv_{out}) achieves convolution after fusion to unify the number of channels.
Here, (F_i^{out}) represents the output feature, and (Conv_{out}) denotes convolution after fusion to unify the number of channels.
By integrating DFFPN into the Neck part of YOLO, efficient multi-scale feature fusion is achieved through dual-flow and cross-layer connections. The end-to-end detection head of YOLO directly receives the multi-scale outputs from DFFPN for classification and regression, enhancing the network’ ability to detect foreign objects of different sizes and shapes on PV panels. This structure realizes dynamic integration of high-level semantics and low-level details through multi-dimensional information flow channels and adaptive fusion strategies, improving the collaborative representation of semantic and spatial features and effectively alleviating the problem of feature separation between layers.
In foreign object detection on photovoltaic panels, small-sized, low-contrast, and blurred-edge targets are often obscured by complex backgrounds, resulting in insufficient model sensitivity to local details and edge textures. Although the traditional Bottleneck structure enhances nonlinear representation capability through serial convolution and residual connections, it still has limitations in high-frequency feature extraction and texture modeling, making it difficult to accurately characterize fine-grained structures. To address this, this paper proposes the CKHB, as shown in Fig. 3. This module integrates local and global information through a high-frequency enhancement residual structure, strengthening the model’s responsiveness to detailed features.
(a) C3K model structure diagram. (b) CKHB structure design.
The traditional Bottleneck structure consists of two serial 3(times)3 convolution (CBS) layers, with residual connections preserving the original input features. Its advantages lie in deepening the network layers, enhancing nonlinear representation capability, and alleviating gradient vanishing. However, this structure has limited ability in detailed texture modeling and faces performance bottlenecks when detecting small-scale or poorly defined boundary targets. Therefore, this paper introduces the HFERB33 module, whose structure is shown in Fig. 4, to enhance high-frequency detail representation and improve the model’s robustness to blurred-edge targets.
HFERB cooperatively extracts multi-scale high-frequency information through parallel local and global high-frequency extraction branches. By using a residual structure to ensure training stability, it effectively enhances the model’s sensitivity to small targets and blurred boundaries without significantly increasing computational complexity.
Schematic diagram of the HFERB structure.
First, the input feature map (text {HFERB}_{in} in mathbb {R}^{H times W times C}) is evenly divided into two sub-branches along the channel dimension, as shown in Eq. (5).
Here, ({HFERB}_{in}^{LFE}) corresponds to the local high-frequency information extraction branch, which is mainly used to capture fine-grained edge and texture features within the spatial structure. ({HFERB}_{in}^{HFE}) corresponds to the global high-frequency information extraction branch, which focuses on high-frequency information related to the overall contour and contextual semantics of the image.
In the local high-frequency branch ((text {HFERB}_{in}^{LFE})), a 3(times)3 convolution is used to extract fine-grained edge and texture information, and the GELU activation function is applied to enhance nonlinear fitting capability, as shown in Eq. (6).
Here, ({HFERB}_{in1}^{LFE}) represents the high-frequency response feature output of local branch 1, GELU denotes the activation function, and ({Conv}_{3times 3}) represents a convolution with a (3times 3) kernel. This step amplifies the high-frequency responses around texture edges and object contours, improving sensitivity to small occlusions.
In the global high-frequency branch (text {HFERB}_{in}^{HFE}), max pooling is first used to extract spatially compressed features, followed by a (1times 1) convolution and GELU activation to construct a low-cost global response channel, as shown in Eq. (7).
Here, ({HFERB}_{in1}^{HFE}) represents the high-frequency response feature output of global branch 2, ({Conv}_{1times 1}) denotes a convolution with a (1times 1) kernel, and MP represents max pooling. This path performs global context modeling by compressing information, helping stabilize training and enhancing the model’s ability to distinguish high-frequency noise under complex backgrounds.
Subsequently, the features from the two branches are concatenated along the channel dimension to integrate multi-scale high-frequency information from both local and global information flows, as shown in Eq. (8).
({HFERB}_{all}) represents the integrated information, and ({Concat}_c) denotes feature concatenation. By expanding the channel dimension, the model can simultaneously focus on edge details and large-scale contours, thereby obtaining richer and more discriminative feature representations.
Since the concatenation operation introduces information redundancy in the channel dimension, excessive redundant features may increase the burden on subsequent network layers and hinder effective feature fusion. Therefore, to control the feature dimension, a 1(times)1 convolution is further applied to ({HFERB}_{all}) for channel compression and information fusion, as shown in Eq. (9).
Here, ({HFERB}_{all}^prime) represents the feature after channel adjustment. This convolution not only serves the purpose of dimensionality reduction but also effectively mixes the information flow among channels, promoting the fusion and interaction between local and global high-frequency features, thereby generating an integrated high-frequency feature map. On this basis, a residual connection mechanism is introduced, where the original input feature ({HFERB}_{in}) is added element-wise to the fused feature map ({HFERB}_{all}^prime), forming the final enhanced output feature map ({HFERB}_{out}), as shown in Eq. (10).
The residual connection mechanism plays an important role in the overall structural design. On one hand, by directly introducing the input features into the output branch, it preserves the original low-frequency background information, thereby ensuring complete semantic transmission and information fidelity while enhancing high-frequency response capability. On the other hand, this structure provides a shorter path for gradient backpropagation, effectively alleviating the gradient vanishing problem in deep networks and improving numerical stability and convergence efficiency during training.
The design of CKHB, while retaining the lightweight structure and local receptive field advantages of the original C3K2 module, introduces a high-frequency response channel. By applying a high-pass filter to the input feature map to extract edge and texture information, it strengthens the discriminative features of foreign object regions. When deployed in variable outdoor environments such as strong light, shadows, rain, or fog, the high-frequency enhancement mechanism of CKHB enables the model to maintain stable and efficient detection performance across different scenarios, demonstrating strong generalization ability and practical value.
The attention mechanism was first applied in the field of natural language processing and, due to its ability to selectively focus on key features, has been widely introduced into computer vision tasks to enhance feature representation and discriminative capability. Traditional convolution treats all spatial positions equally, whereas the attention mechanism simulates the human visual system’s focus behavior, enabling the model to concentrate on task-relevant regions, thereby improving representational efficiency and decision accuracy.
However, in high-resolution vision tasks, the quadratic computational complexity of self-attention structures makes them difficult to apply in real-time scenarios. Meanwhile, in photovoltaic panel foreign object detection tasks, traditional attention mechanisms are limited by a restricted receptive field, making it difficult to capture both global contextual and local detailed features simultaneously. They are also easily affected by illumination changes, reflections, and background noise. Although the C2PSA module in YOLO series models integrates partial residual and spatial attention mechanisms to enhance attention to salient regions, its global dependency modeling capability remains limited, particularly in complex environments where long-range contextual perception is insufficient.
To address the above issues, this paper introduces a Large Separable Kernel Attention (LSKA)34 mechanism after the C2PSA module, as shown in Fig. 5. This module utilizes large receptive field convolution operations and a selective multi-branch weighting mechanism to effectively expand the receptive field, enhance contextual modeling capability, and improve the model’s robustness to small-scale targets and complex backgrounds.
Schematic diagram of the LSKA structure.
The LSKA module retains the advantages of LKA (Large Kernel Attention) in expanding the receptive field while introducing multi-branch large-kernel convolution paths, multi-scale modeling strategies, and a feature fusion mechanism. It simulates the Query–Key–Value information interaction process of the self-attention mechanism from the spatial dimension, thereby achieving stronger contextual awareness and expressive capability.
First, to reduce the computational burden caused by multi-branch large-kernel operations, the LSKA module performs channel compression on the input feature map (X in mathbb {R}^{C times H times W}), obtaining an intermediate representation (X’ in mathbb {R}^{C times H times W}). This compression process can be formally expressed as Eq. (11):
Here, ({W}_{r}) represents the (1times 1) pointwise convolution weights used for channel mapping.
The compressed feature (X’) is then fed into multiple parallel branches, each employing depthwise separable convolution or dilated convolution with different kernel sizes to extract multi-scale spatial contextual information. The feature extraction process of the i-th branch is defined as Eq. (12).
Here, (K_i in {5, 9, 15, 21}) represents different convolution kernel sizes, and (D_i) denotes the corresponding dilation rate, which can be adjusted to meet the modeling requirements of various receptive fields.
Each branch independently models spatial features with different receptive fields, functionally analogous to each Query–Key pair in the multi-head self-attention mechanism, capturing spatial relationships and contextual dependencies at multiple scales. After obtaining all branch output features ({F_1, F_2, ldots , F_n}), the LSKA module provides two fusion strategies to achieve feature integration and enhanced representation.
The first strategy is concatenation fusion, in which the feature maps extracted from each branch are concatenated along the channel dimension. Subsequently, a 1(times)1 pointwise convolution is applied to the concatenated result for dimensional expansion, restoring the original number of channels and achieving efficient fusion of multi-scale information, as shown in Eq. (13).
The second strategy is weighted fusion, which introduces a set of learnable fusion weight parameters ({alpha _1, alpha _2, ldots , alpha _n}) to perform weighted summation of the output features from each branch, simulating the information selection and aggregation process in the attention mechanism, as shown in Eq. (14)
Finally, the fused feature map F is connected to the original input X through a residual connection to form the final output, which enhances feature representation while preserving the original information to facilitate gradient propagation and network stability.
Integrating the LSKA mechanism into the YOLO framework can significantly improve the performance of foreign object detection on photovoltaic module surfaces. By combining large convolution kernels with self-attention, LSKA enhances global contextual modeling capability, expands the global receptive field, and captures global spatial correlations. As a result, it improves target recognition accuracy in complex scenes, reduces background false detections, and enhances robustness against various types of foreign objects while maintaining stable detection accuracy under dynamic conditions. At the same time, it suppresses redundant feature interference and dynamically preserves key information. This integration effectively alleviates problems such as missed detections of small objects, false alarms from background interference, and insufficient scale adaptability, providing strong support for intelligent operation and maintenance in industrial PV systems.
To ensure the reliability of DHLNet model validation, all training and testing tasks in this study were conducted on a single GPU. The experiments were performed in a Windows 11 environment with the following hardware configuration: an NVIDIA RTX 4090 graphics card (24 GB VRAM), 128 GB of memory, and an Intel Core i7-13700K processor. The software environment included Python 3.10, PyTorch 2.1.1, and CUDA 12.6. This configuration efficiently handles high-resolution PV module images, supports large-batch training and complex data augmentation strategies, ensures training stability and convergence efficiency, and improves the reproducibility of experimental results. Optimization was performed using Stochastic Gradient Descent (SGD), and the specific training parameters are shown in Table 2.
To verify the effectiveness of the proposed DHLNet model, a representative photovoltaic module defect dataset was constructed. This dataset consists of images collected from multiple PV power plants as well as high-resolution photographs obtained from online sources. The images cover a wide range of environmental conditions, including different seasons, lighting intensities, and weather scenarios, realistically reflecting the complexity of real-world PV inspection environments.
All images were captured using standard RGB cameras. During data collection, the photovoltaic module target occupied at least two-thirds of the image frame; in some samples, the target region nearly or completely covered the entire frame to ensure diversity in target scale and viewing angles. After collection, all images were uniformly resized to 640 (times) 640 pixels to meet the input requirements for model training and inference. In addition, to address potential overlap between on-site captured images and images obtained online, all samples were manually visually inspected during data curation to remove duplicate or highly similar images. This ensured that there was no data overlap among the training, validation, and test sets, thereby effectively preventing data leakage from affecting the experimental results.
To ensure annotation accuracy and consistency, all images were manually labeled using the LabelImg tool in YOLO format, with each bounding box precisely corresponding to a single instance of a foreign object or defect. In addition, data augmentation techniques such as random flipping, brightness adjustment, and Gaussian blur were applied to enhance the model’s robustness and generalization ability.
Finally, the dataset was randomly divided into training, validation, and testing sets in a ratio of 8:1:1. The training set contains 2903 images, the validation set contains 402 images, and the test set contains 331 images. All images were resized to 640(times)640 pixels as model input. The specific number of targets is shown in Table 3, and sample examples are illustrated in Fig. 6.
Dataset display.
To evaluate the performance of the model in the task of foreign object detection on photovoltaic modules, the following key metrics were adopted: F1 score, Precision, Recall, number of parameters, mAP@0.5, and mAP@0.5:0.95. The F1 score characterizes the balance between Precision and Recall. Precision represents the proportion of correctly predicted positive samples among all samples predicted as positive, while Recall measures the model’s ability to detect actual positive samples. The definitions of these metrics are given in Eqs. (15) and (16).
TP represents the number of true targets successfully detected, that is, the number of predicted boxes matching the actual defect regions. FN represents the number of true targets that were not detected, while FP refers to the number of incorrectly detected targets, meaning predicted boxes that have no overlap or insufficient overlap with any real target. The number of parameters reflects the model’s complexity; mAP@0.5 is the mean average precision under the condition of IoU = 0.5, while mAP@0.5:0.95 evaluates the model’s detection accuracy and generalization capability comprehensively across multiple IoU thresholds. The formulas are shown in Eqs. (18) and (19).
Here, N represents the total number of categories, and (AP_i) denotes the average precision of the i-th category.
To comprehensively evaluate the performance of the proposed DHLNet in the photovoltaic module foreign object detection (FOD) task, we compared it with several representative traditional object detection algorithms, including YOLOv5n, YOLOv6n, YOLOv7t, YOLOv8n, YOLOv9t, YOLOv10n, YOLOv12n, and the original YOLO11n.
To further analyze the performance of each algorithm during training, we plotted the loss and mean Average Precision (mAP) curves with respect to the number of training iterations. The results are shown in Fig. 7a and b.
Loss function and mAP variation curves of the YOLO series.
As shown in Fig. 7a, the loss functions of all models exhibit a downward trend during training, indicating gradual network convergence. Among them, DHLNet shows a rapid decrease in loss within the first 50 epochs, converging faster than mainstream lightweight models such as YOLOv5n–YOLOv12n. Particularly in the later stages, the loss of DHLNet stabilizes around 0.5, outperforming comparison models such as YOLOv5n, YOLOv8n, and YOLOv11n, demonstrating superior training stability and generalization capability. This advantage mainly stems from the model’s precise capture and effective preservation of high-frequency information, as well as the attention mechanism’s adaptive focus and enhanced representation of critical details such as defect edges and subtle foreign objects, enabling robust recognition of PV panel fault features even in complex scenarios.
Figure 7b shows that the proposed DHLNet performs excellently in terms of mAP@0.5, achieving high detection accuracy early in the training process and eventually stabilizing at 0.799, which surpasses models such as YOLOv5n, YOLOv8n, and YOLO11n. This result indicates that DHLNet has stronger representational capability in feature extraction and object recognition, improving small-object detection accuracy under complex conditions. The stable plateau of its mAP curve further reflects strong generalization ability. Overall, the proposed model architecture and training strategy significantly enhance detection accuracy and convergence speed, providing reliable support for practical deployment.
To comprehensively evaluate the performance of the proposed method in the task of foreign object detection on photovoltaic panels, this experiment uses the YOLO11n algorithm as the baseline for improvement and selects several mainstream foreign object detection algorithms for comparison. The comparative experiments were conducted on the same test set. The evaluation includes the average precision (APs) metric for small targets smaller than 32(times)32 pixels to ensure the comprehensiveness and objectivity of the results. The experimental results of each algorithm in the PV panel foreign object detection task are shown in Table 4.
From the perspective of key quantitative evaluation metrics, DHLNet demonstrates stronger discriminative capability on the benchmark dataset. Its mAP@0.5 reaches 0.799, representing an improvement of 1.0 percentage point over the comparable YOLOv8n and 3.1 percentage points over the baseline architecture YOLO11n. Under the more stringent mAP@0.5:0.95 evaluation protocol, DHLNet achieves a score of 0.588, improving by 2.2 percentage points compared with both YOLOv8n and YOLO11n. These results empirically indicate that DHLNet possesses higher localization accuracy and semantic classification capability under high IoU threshold conditions, making it well-suited for fine-grained target detection in complex operational environments.
Although the Transformer-based RT-DETR model achieves a relatively high precision of 0.776, its recall is comparatively low at 0.676, resulting in an F1-score of 0.721 and indicating insufficient overall balance. In addition, its mAP@0.5:0.95 reaches only 0.519, reflecting limited detection stability and generalization capability in complex scenarios. In contrast, D-FINE outperforms RT-DETR in both F1-score and mAP@0.5, suggesting a more reasonable trade-off between detection accuracy and recall rate, although its performance still falls short of that of DHLNet. Conversely, DEIM performs relatively poorly; despite having an acceptable recall, its mAP@0.5:0.95 is only 0.485, indicating limited generalization ability in diverse environments.
The detection results are shown in Fig. 8, illustrating the visual outputs of these models on typical test images, which include various types of defects such as scratches, shadows, bird droppings, and cracks. In comparison, YOLOv8 and YOLOv9 exhibit instances of misclassification and redundant bounding boxes, demonstrating their limited ability to distinguish targets in complex scenarios. The multi-scale contextual awareness mechanism and enhanced feature fusion strategy introduced in DHLNet effectively improve its detection capability for small objects and densely distributed defects.
Comparison of detection results of different models.
The model performance results are presented in Table 5. Here, PT Inf Time denotes the average inference time when using the original weight file, reflecting the model’s baseline inference efficiency without additional deployment optimization. ONNX Inf Time refers to the average inference time after exporting the model to the ONNX format and applying inference acceleration, which is used to demonstrate the inference speed gains in engineering deployment scenarios. In addition, the Memory metric indicates the average GPU memory consumption when performing PT-based inference, and can be used to evaluate the model’s hardware resource usage during runtime.
It can be observed that the traditional two-stage detector Faster R-CNN, although having 41.5M parameters and a computational cost as high as 370.2 GLOPs, suffers from extremely low inference efficiency: its FPS is only 2.9, its PT inference time reaches 344.83 ms, and it also incurs the highest memory usage at 1351.64 MB, making it difficult to meet real-time detection requirements. In contrast, lightweight YOLO models generally keep their parameter size within the 2–6M range, with inference times of about 2–3 ms and memory usage stably within 910–985 MB, demonstrating strong real-time performance and good suitability for deployment.
More specifically, DHLNet achieves the best overall performance while remaining lightweight. It contains 2.75M parameters and requires 6.7 GLOPs, and reaches an FPS of 462.1, the highest among all models. Its PT and ONNX inference times are reduced to 2.16 ms and 1.97 ms, respectively. With a GPU memory footprint of 922.50 MB, which is comparable to other lightweight models, these results indicate that the proposed method improves detection performance without significantly increasing GPU resource consumption, showing strong potential for real-time deployment.
DHLNet is an enhanced architecture built upon the YOLO framework and achieves state-of-the-art performance across all evaluation metrics. Its superior performance primarily benefits from the proposed DFFPN, which effectively extracts multi-scale spatial and semantic information while improving the representation capability for small-scale and occluded targets.
The visualization results of the DHLNet model for foreign object detection on PV panels are shown in Fig. 9. The model accurately identifies various types of defects and occlusions, including bird droppings, leaves, dust, and snow. It performs stably when localizing objects with blurred edges, irregular shapes, or overlapping regions, with detection boxes tightly aligned to the target areas—demonstrating stronger spatial localization capability.
DHLNet model detection results.
Moreover, the model maintains high detection confidence under complex lighting conditions and achieves reliable recognition in low-contrast regions and partially reflective surfaces. This indicates that the proposed feature fusion and high-frequency enhancement mechanisms effectively improve the model’s robustness and generalization ability. Overall, the results verify that DHLNet delivers excellent detection performance for multi-scale targets, complex backgrounds, and diverse foreign objects in real-world photovoltaic scenarios.
Details of small-object detection.
To verify DHLNet’s capability in recognizing tiny foreign objects under complex backgrounds, Fig. 10 presents the detection results of different categories of small targets along with locally magnified regions. The first and second examples correspond to the Bird Drop category. It can be observed that bird droppings are usually small in scale, with blurred edges and textures similar to the background, yet the model is still able to accurately localize them and provide high-confidence predictions. The third example shows the Dust category, where such foreign objects often appear as weak-texture contamination with inconspicuous target features; nevertheless, the model can effectively capture fine-grained details, thereby reducing missed detections.
The fourth example belongs to the Physical Damage category, which has complex shapes and is easily affected by component reflections. Even in this case, the model maintains good detection accuracy. The locally enlarged results demonstrate that the proposed model possesses stronger feature representation capability and robustness for small-object foreign object detection, and can meet the requirements of accurate multi-category foreign object recognition in complex photovoltaic panel scenarios.
To systematically analyze the individual effects and synergistic contributions of each improved module, ablation experiments were conducted by progressively introducing each component. This approach enables a quantitative evaluation of the actual performance contribution of each module. The experimental results on the test set are presented in Table 6.
Using YOLO11n as the baseline model, its mAP@0.5 on the standard test set is 0.768 and mAP@0.5:0.95 is 0.566; however, it shows limitations in detecting small targets and occluded foreign objects. After introducing the HFERB module, the F1-score increases to 0.760 and mAP@0.5 improves by 1.0 percentage point, indicating that the high-frequency enhancement structure effectively improves feature representation and edge recognition capability. With the further integration of the LSKA module, the Recall rises to 0.732, and the model’s ability to perceive small-scale foreign objects in complex backgrounds is significantly enhanced. Subsequently, the inclusion of the DFFPN architecture raises the Recall to 0.756 and the mAP@0.5:0.95 to 0.576, verifying the effectiveness of the multi-scale semantic fusion strategy in foreign object detection.
Visualization of detection results before and after the introduction of each module in the ablation study.
As shown in Fig. 11, the progressive integration of HFERB, LSKA, and DFFPN modules enables the model to achieve superior performance in boundary localization and feature discrimination, with a noticeable reduction in false detections. The improvements are particularly evident in detecting small, low-contrast, and occluded foreign objects.
Overall, DHLNet achieves 1.3% increase in F1-score, and improvements of 3.1% and 2.2% in mAP@0.5 and mAP@0.5:0.95, respectively, with only a slight increase in computational complexity. The model significantly enhances detection accuracy and robustness while maintaining lightweight design and real-time performance, demonstrating strong potential for engineering applications in photovoltaic inspection systems.
This study proposes a novel lightweight object detection model, DHLNet, designed specifically for foreign object detection on photovoltaic (PV) panels. The model introduces structural improvements based on the YOLOv11n framework to better address the challenges of PV inspection scenarios, such as small object sizes, complex backgrounds, and diverse types of foreign objects. Specifically, the HFERB module is introduced to enhance the feature extraction capability of the C3K2 structure, improving the model’s perception of fine-grained textures and edge information. The large kernel self-attention mechanism (LSKA) is integrated into the backbone network to significantly expand the receptive field, effectively enhancing small-object detection performance under complex lighting and reflection conditions. In addition, the multi-scale feature fusion module (DFFPN) is incorporated to improve the model’s robustness and feature integration across different object scales.Experimental results demonstrate that the improved DHLNet achieves mAP@0.5 = 0.799 and mAP@0.5:0.95 = 0.588, representing a substantial improvement over the original YOLOv11n. Meanwhile, the model maintains a lightweight structure with 462.1 FPS inference speed and 2.16 ms per image latency, indicating excellent efficiency and strong adaptability for edge deployment. Overall, DHLNet provides an efficient, accurate, and deployable solution for real-time intelligent detection of foreign objects on PV panels, showing great potential for widespread application in smart energy maintenance and automated inspection systems.
Future work will focus on further enhancing the lightweight design and reducing computational complexity to better satisfy the constraints of resource-limited edge devices. In addition, we plan to deeply integrate the detection framework with UAV path planning algorithms to achieve collaborative optimization between target detection and autonomous inspection, thereby improving the practicality and intelligence of the proposed method in real-world PV operation and maintenance scenarios. Although the current DHLNet framework demonstrates high accuracy and strong robustness in detecting known foreign object categories, it has not yet incorporated adaptive learning for unseen defect types. In real PV applications, new or rare defect patterns may emerge that deviate from the training data distribution. Addressing this limitation will be a key direction of our future research. Specifically, we intend to integrate open-set recognition and incremental learning mechanisms into DHLNet, enabling it to dynamically adapt to and recognize novel or previously unseen defect categories. This improvement will further enhance the model’s generalization capability, making it more applicable to the continuously evolving PV inspection environment.
Data will be made available on request.
El-Banby, G. M. et al. Photovoltaic system fault detection techniques: A review. Neural Comput. Appl. 35(35), 24829–24842 (2023).
Article Google Scholar
Alshareef, M. J. A comprehensive review of the soiling effects on PV module performance. IEEE Access 11, 134623–134651 (2023).
Article Google Scholar
Polymeropoulos, I. et al. Enhancing solar plant efficiency: A review of vision-based monitoring and fault detection techniques. Technologies 12(10), 175 (2024).
Article Google Scholar
Voutsinas, S., Karolidis, D., Voyiatzis, I. & Samarakou, M. Development of a machine-learning-based method for early fault detection in photovoltaic systems. J. Eng. Appl. Sci. 70(1), 27 (2023).
Article Google Scholar
Natarajan, K., Bala, P. K. & Sampath, V. Fault detection of solar PV system using SVM and thermal image processing. Int. J. Renew. Energy Res. 10(2), 967–977 (2020).
Google Scholar
Herraiz, A. H., Marugán, A. P. & Márquez, F. P. G. Photovoltaic plant condition monitoring using thermal images analysis by convolutional neural network-based structure. Renew. Energy 153, 334–348 (2020).
Article Google Scholar
Guerriero, P. et al. Monitoring and diagnostics of PV plants by a wireless self-powered sensor for individual panels. IEEE J. Photovolt. 6(1), 286–294 (2015).
Article Google Scholar
Taher, M. A. et al. Wavelet and signal analyzer based high-frequency ripple extraction in the context of MPPT algorithm in solar PV systems. IEEE Access 12, 113726–113740 (2024).
Article Google Scholar
Høiaas, I. et al. Inspection and condition monitoring of large-scale photovoltaic power plants: A review of imaging technologies. Renew. Sustain. Energy Rev. 161, 112353 (2022).
Article Google Scholar
Monicka, S. G. et al. Image processing based hot-spot detection on photovoltaic panels. Int. J. Intell. Syst. Appl. Eng 11, 510–518 (2023).
Google Scholar
Zhao, X. et al. HRNet-based automatic identification of photovoltaic module defects using electroluminescence images. Energy 267, 126605 (2023).
Article Google Scholar
Mathew, A., Amudha, P. & Sivakumari, S. Deep learning techniques: An overview. International conference on advanced machine learning technologies and applications 599–608 (Springer Singapore, 2020)
Dong, S., Wang, P. & Abbas, K. A survey on deep learning and its applications. Comput. Sci. Rev. 40, 100379 (2021).
Article MathSciNet Google Scholar
Wei, Y., Shao, C. & Zeng, G. Detection system of foreign objects coverage on PV panels. IOP Conf. Series: Mater. Sci. Eng. IOP Publishing 394(4), 042103 (2018).
Srivastava, D. & Singh, N. Anomaly detection and circuit-based categorisation of fault in PV systems. Int. J. Power Electron. 20(2), 152–179 (2024).
Article Google Scholar
Prasshanth, C. et al. Fault detection in photovoltaic systems using unmanned aerial vehicle-captured images and rough set theory. Solar Energy 290, 113348–113348 (2025).
Article Google Scholar
Girshick, R., Donahue, J., Darrell, T. et al. Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition 580–587 (2014).
Xu, X. et al. Crack detection and comparison study based on faster R-CNN and mask R-CNN. Sensors 22(3), 1215 (2022).
Article ADS PubMed PubMed Central Google Scholar
Ren, S. et al. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137–1149 (2016).
Article ADS PubMed Google Scholar
Zhai, S. et al. DF-SSD: An improved SSD object detection algorithm based on DenseNet and feature fusion. IEEE Access 8, 24344–24357 (2020).
Article Google Scholar
Yin, W. et al. PV-YOLO: Lightweight YOLO for photovoltaic panel fault detection. IEEE Access 11, 10966–10976 (2023).
Article Google Scholar
Guo, F. et al. Research on fault diagnosis method for photovoltaic array based on model fusion. Electr. Eng. 107, 8189–8199 (2025).
Article Google Scholar
Vlaminck, M. et al. Region-based CNN for anomaly detection in PV power plants using aerial imagery. Sensors 22(3), 1244 (2022).
Article ADS PubMed PubMed Central Google Scholar
Dong, M. et al. ISEE: Industrial Internet of Things perception in solar cell detection based on edge computing. Int. J. Distrib. Sens. Netw. 17(11), 15501477211050552 (2021).
Article Google Scholar
Pierdicca, R. et al. Automatic faults detection of photovoltaic farms: solAIr, a deep learning-based system for thermal images. Energies 13(24), 6496 (2020).
Article ADS Google Scholar
Blá¡zquez Folch, J. Detection of artifacts in infrared images and video of PV solar panels (Universitat Politècnica de Catalunya, 2024).
Salehpour, S. et al. Two-stage deep Q-network reinforcement learning based ultra-efficient fault diagnosis and severity assessment scheme for photovoltaic protection. Energy AI 20, 100512–100512 (2025).
Article Google Scholar
Adarsh, P., Rathi, P. & Kumar, M. YOLO v3-Tiny: Object detection and recognition using one stage improved model. In: 2020 6th international conference on advanced computing and communication systems (ICACCS). 687–694. (IEEE, 2020)
Hong, F. et al. A novel framework on intelligent detection for module defects of PV plant combining the visible and infrared images. Solar Energy 236, 406–416 (2022).
Article ADS Google Scholar
Sun, T., Gao, H., Fan, S. et al. Enhancing Dust Detection on Photovoltaic Panels with PP-YOLO: A Deep Learning Approach. In: 2024 4th International Conference on Machine Learning and Intelligent Systems Engineering (MLISE) 24–28 (IEEE, 2024).
Shamta, I., Demir, F. & Demir, E. B. Predictive fault detection and resolution using YOLOv8 segmentation model: A comprehensive study on hotspot faults and generalization challenges in computer vision. Ain Shams Eng. J. 15(12), 103148–103148 (2024).
Article Google Scholar
Chen, J., Mai, H.S., Luo, L. et al. Effective feature fusion network in BIFPN for small object detection. In: 2021 IEEE international conference on image processing (ICIP) 699–703 (IEEE, 2021).
Li, A., Zhang, L., Liu, Y. et al. Feature modulation transformer: Cross-refinement of global representation via high-frequency prior for image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision pp. 12514–12524 (2023).
Lau, K. W., Po, L. M. & Rehman, Y. A. U. Large separable kernel attention: Rethinking the large kernel attention design in CNN. Expert. Syst. Appl. 236, 121352 (2024).
Article Google Scholar
Zhao, Y., Lv, W. et al. DETRs beat YOLOs on real-time object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 16965–16974 (2024).
Zhang, S., Chi, C., Yao, Y., Lei, Z. & Li, S.Z. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: Proc, lEEE/CVF Conf Comput. Vis. Pattern Recognit. (CVPR), 9756–9765 (2020).
Peng, Y., Li, H., Wu, P., Zhang, Y., Sun, X. & Wu, F. D-FINE: Redefine regression task in DETRs as fine-grained distribution refinement. arXiv preprint, arXiv:2410.13842, (2024).
Huang S., Lu Z., Cun X., Yu Y., Zhou X., Shen X. et al. DEIM: DETR with improved matching for fast convergence. arXiv preprint, arXiv:2412.04234, (2024).
Download references
This work was supported by the National Natural Science Foundation of China [Project No. 62173171], the Natural Science Foundation of Liaoning Province [Project No. 2022-MS-397], and the Fuxin Gongda Chengpu Electric Co., Ltd. [Project No. 2025-H0015].
Faculty of Software, Liaoning Technical University, Huludao, 125105, Liaoning, China
Haibo Jin
Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao, 125105, Liaoning, China
Mengjiao Li, Xiaoyun Lv & Jishen Peng
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
Haibo Jin: Conceptualization, Methodology, Funding acquisition; Mengjiao Li: Software, Investigation, Formal analysis; Xiaoyun Lv: Writing—original draft, Software, Validation; Jishen Peng: Writing—review and editing.
Correspondence to Haibo Jin.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Jin, H., Li, M., Lv, X. et al. Foreign object detection on photovoltaic panels based on DHLNet. Sci Rep 16, 8145 (2026). https://doi.org/10.1038/s41598-026-39074-6
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41598-026-39074-6
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.
