betaFeedback

Reading v1View latest version

WorksComputer Science

Towards Automated Solar Panel Integrity: Hybrid Deep Feature Extraction for Advanced Surface Defect Identification

Abstract

To ensure energy efficiency and reliable operations, it is essential to monitor solar panels in generation plants to detect defects. It is quite labor-intensive, time consuming and costly to manually monitor large-scale solar plants and those installed in remote areas. Manual inspection may also be susceptible to human errors. Consequently, it is necessary to create an automated, intelligent defect-detection system, that ensures continuous monitoring, early fault detection, and maximum power generation. We proposed a novel hybrid method for defect detection in SOLAR plates by combining both handcrafted and deep learning features. Local Binary Pattern (LBP), Histogram of Gradients (HoG) and Gabor Filters were used for the extraction of handcrafted features. Deep features extracted by leveraging the use of DenseNet-169. Both handcrafted and deep features were concatenated and then fed to three distinct types of classifiers, including Support Vector Machines (SVM), Extreme Gradient Boost (XGBoost) and Light Gradient-Boosting Machine (LGBM). Experimental results evaluated on the augmented dataset show the superior performance, especially DenseNet-169 + Gabor (SVM), had the highest scores with 99.17% accuracy which was higher than all the other systems. In general, the proposed hybrid framework offers better defect-detection accuracy, resistance, and flexibility that has a solid basis on the real-life use of the automated PV panels monitoring system.

INTRODUCTION

Electricity is a vital energy source that underpins residential, industrial, and technological applications. It can be generated both by conventional and renewable sources. Oil, coal, nuclear, and natural gas are among the transitional sources [cite:p]{ref-1}, and are linked to severe health and ecological problems because they release hazardous pollutants and industrial waste.

Alternatively, renewable energy sources such as SOLAR, wind, hydropower and geothermal are healthier, more resilient, and have a considerably lower environmental impact [cite:p]{ref-2}.

To achieve global sustainability and establish an acceptable way of life for everyone, the most crucial step is to increase energy efficiency by switching to the use of renewable energy sources [cite:p]{ref-3}. As people are getting more concerned about the environment, health, and society, they are actively moving towards the renewable energy sources for electricity generation. Renewable energy sources can be broadly categorized into five main categories: solar energy harnessed from the sun [cite:p]{ref-4}, wind energy obtained through moving air [cite:p]{ref-5}, biomass power derived from organic sources [cite:p]{ref-6}, hydropower energy generated from flowing water [cite:p]{ref-7}, and geothermal energy obtained from the internal heat of the earth [cite:p]{ref-8,ref-9}.

Among all renewable energy sources, solar is one of the cleanest sources that harnesses the sunlight, transforming it effortlessly to generate power [cite:p]{ref-4,ref-10}. It can be installed on rooftops and other open areas and is most effective in areas with abundant sunlight, such as deserts. Solar energy is flexible, capable of supplying power to individual appliances, residential buildings, and even entire cities without emitting harmful pollutants, making it environmentally friendly. It is also beneficial for off-grid facilities that cannot connect to conventional energy sources or cannot access them at all [cite:p]{ref-11}.

Despite its advantages, solar energy sources have numerous challenges. One of the major challenges is the gradual depreciation of solar panels over time, which results in a decrease in overall efficiency, and highlights the need for effective fault detection and monitoring strategies [cite:p]{ref-12}. Another critical challenge is that solar power generation is highly susceptible to weather conditions - factors such as rain and cloud cover can significantly reduce system performance [cite:p]{ref-13}.

Solar panels may also face an array of challenges that make them less efficient at generating electricity. Some common visual problems (as shown in Figure 1) include dust, bird droppings, snow cover, electrical faults, and physical damage. Dust particles may accumulate on the surface of solar plates, making it harder for sunlight to reach them and reducing the generation capacity of electrical power. The same thing happens when bird droppings build up: they block light from entering. In colder areas, snow may accumulate on solar panels for extended

Image not extracted

Figure 1.

Image not extracted

Figure 1. Visual Samples Depicting Various Surface Conditions of PV Panels for Defect Detection: (a) Clean, (b) Dust, (c) Snow-covered, (d) Bird Droppings, (e) Physical Damage, and (f) Electrical Fault

periods, making it even harder to generate energy. External factors, including hail hitting the solar glass or cells, can cause physical damage, such as cracks. Overheating may result in electrical issues, leading to burnout of solar cells overheating of solar panels leading to the burn out of solar cells [14]. The quality inspection processes heavily rely on defect detection. Defect or anomaly detection (AD) process aims at detecting abnormal features that may lead to quality or functionality issues of a particular object [15, 16, 17].

As solar installations continue to expand globally, it has become essential to ensure they operate effectively and last a long time. Regular physical inspection is one of the most essential steps to maintain the quality and reliability of solar plates. It is quite a challenging and laborious task to perform the manual inspection of large-scale solar plants installed at governmental or industrial scale. Manual inspection is becoming impractical as the number of solar panel installations continues to increase [18].

There is a pressing need to propose an accurate and efficient method for automatic inspection of PV panels to ensure their long operational life span and maintain high energy efficiency. Due to recent advancements in the field of artificial intelligence, computer vision and deep learning, researchers from different research domains are entering in the areas of automated defect detection such as surface crack detection [19, 20], product defect detection [21], and damage detection [22, 23, 24, 25].

To address the challenges associated with manual inspection, we proposed a hybrid framework combining both deep and machine learning techniques for efficient and accurate inspection of solar plates. The proposed framework consists of four major components: data augmentation, data preprocessing, feature extraction network (FEN) and defect classification. A publicly available dataset on KAGGLE [22], consisting of 1,574 images, is used for experimentation purposes. Due to the low volume of the dataset and to reduce the risk of overfitting, three different data augmentation operations, such as rotation, flipping, and image translation were applied. This overall augmentation process resulted in a fourfold increase in dataset size, reaching about 6,060 images.

Different image pre-processing techniques were then employed to improve the image quality by minimizing background noise. A hybrid feature extraction network based on the combination of traditional machine learning approaches

and deep learning models, designed to capture both handcrafted and deep features from PV plate images. For handcrafted features extraction, different methods such as Linear Binary Patterns (LBP) [26], Histogram of Oriented Gradients (HOG) [27], and Gabor Filters [28], were used to effectively represent texture, edge, and frequency-based characteristics of the images.

On the other hand, Convolutional neural networks (CNNs) [29, 30] were used to extract deep features. Handcrafted and deep features were concatenated to generate a combined feature matrix yielding stronger representation and better classification performance. These features were used to train such classifiers as Support Vector Machines (SVM) [31, 32], Artificial Neural Networks (ANN) [33, 34], Logistic Regression (LR) [35, 36], XGBoost [37] to obtain precise defect classification. The block diagram of the proposed system is (as shown in Figure 2). The contributions of this research are as follows:

Block diagram of the proposed system for PV panel defect detection

Figure 2. Block diagram illustrates the overall architecture of the proposed system for PV panel defect detection.

  • An optimized data augmentation pipeline is proposed, resulting in a fourfold increase in data volume to improve model generalization.

  • Proposed an automatic, accurate and efficient inspection system based on a hybrid network that combines handcrafted features (extracted using LBP, HoG and Gabor) with deep learning features extracted by leveraging the use of CNN to enhance the classification accuracy.

  • The novel feature-level fusion method combines both handcrafted and deep features through a concatenation into a single feature matrix, to obtain optimal classification by combining the limited strengths of other features.

  • Fused features were tested with various machine learning classifiers (SVM, XGBoost, ANN, Logistic Regression) and comparable results showed improved performance of fused features compared to current single feature models.

The paper will be structured as follows: Section 2 will provide an overview of recent methodologies applied for detecting defects in SOLAR panels. Section 3 outlines the proposed methodology, which includes a description of the dataset used, data processing, and hybrid feature extraction. Section 4 describes the experimental design, presents the qualitative and quantitative findings, and offers a comparison with previous methods. Section 5 discusses the relevance of the findings and their implications to real-world use of PV monitoring. Lastly, Section 6 concludes the paper by summarizing the main contributions and outlining possible future research directions.

LITERATURE REVIEW

Recent research has utilized both manual feature-based approaches (e.g., texture and edge descriptors) and deep learning-based approaches (e.g., CNNs and Transformers) for PV panel monitoring. The following section will analyze these existing methods in more detail, outlining the major challenges and motivating the design and development of a hybrid feature extraction framework that would integrate the advantages of the two methodologies and enhance the precision of defect detection.

Traditional Approaches:

Earlier approaches to product defect detection primarily focused on traditional image processing techniques that generally concentrated on capturing handcrafted features such as geometric structure, color, shape, edge patterns, and surface roughness or intensity variations, using traditional feature descriptors or statistical techniques. For instance, Tsai et al. [38] used Fourier image reconstruction to detect defects, such as cracks, breaks, and interruptions, in electroluminescence (EL) images of polycrystalline solar modules. They used their methodology of filtering the high-frequency components in the Fourier domain and identifying defects comparing the intensity of the original and reconstructed images. The method, however, demonstrated lower effectiveness when subjected to the defects of a difficult nature. Author proposed another method for solar plate defect-detection based on Independent Component Analysis (ICA) by optimizing the feature extraction and reconstruction steps and making them much more effective in slight variations in texture and intensity in EL images. On the same note, Zhang et al. [39] came up with an Independent Component Analysis (ICA)-based reconstruction algorithm, where a defect-free image was used as a reference to help in the extraction of independent components that would be used in defect detection. The reconstructed images were very effective at highlighting local anomalies; hence, the technique was appropriate for revealing local defects on textured surfaces with periodic features. Both Fourier image reconstruction and ICA-based methods have severe limitations in handling complex, irregular shaped defects. The Fourier-based method is sensitive to noise and light variations, thus limiting its accuracy across diverse environmental conditions. Similarly, ICA also depends on high-quality reference images and experiences challenges with periodic or less visible anomalies, thus limiting its practical usefulness for inspecting solar panels.

Deep Learning Approaches:

Due to the challenges faced by traditional image processing and machine learning approaches, deep learning approaches have emerged as state-of-the-art techniques for tasks related to image analysis. In this regard, Convolutional neural networks CNNs shows promising results by capturing both local as well as global features for identification and classification of defects.

Shaik et al. [40] proposed a novel approach to enhance the monitoring and defect-detection for solar panels, addressing the challenges of traditional manual inspections, particularly at scale. By utilizing Unmanned Aerial Vehicles (UAVs) and satellite imagery, a more rapid and efficient monitoring system has been established. A deep learning-based segmentation U-NET model integrated with an Atrous Spatial Pyramid Pooling (ASPP) module was proposed, achieving an impressive overall accuracy of 98% and a Mean Intersection-Over-Union (IoU) Score of 95%. The proposed system efficiently detects five classes of damage, including undamaged solar panels, demonstrating high level of effectiveness in defect identification.

Another author [41] proposed a deep learning-based framework that utilizes the thermal images for the identification and classification of defects in PV plates. The proposed model, based on Convolutional Neural Network with VGG16 backbone, achieves an impressive F1-Score of 84.12% and Mean Minimum Precision (MMP) of 71%, which may be explained by using localization heatmaps to identify errors with high precision. Furthermore, the model’s adaptability and strong competitiveness were demonstrated through a comparative analysis with advanced models including Yolo-v8 and transformer-based architectures such as Visual Transformers (ViT) and Swin Transformers. Both techniques, as suggested [40], [41] rely heavily on modalities of images such as EL or thermal imaging and hence, their usability in a variety of environmental situations is limited. UAV- and satellite-based systems, although efficient, face the problem of scaling, which is driven by high computational and operational costs. A framework based on MobileNet-V3 was proposed by Mansour et al. [42], achieving 100% accuracy in identifying and classifying PV panel defects, such as bird-dropping and snow cover. Another author, Kazmi et al. [43] suggested using transfer learning with a pretrained VGG-16 model as a backbone and achieved an accuracy of 74%. These findings indicate remarkable progress toward the automatic diagnosis of faults in solar panels, opening a new horizon for energy efficiency and fewer manual inspections for privacy. DenseNet-121, also suggested by another author [44], achieves an accuracy of 86.44%.

Nagar et al. [45] explored the DenseNet169 model using a dataset of 11,000 images across six classes: bird droppings, dust, electrical damage, physical damage, clean panels, and snow-covered panels. Impressively, the DenseNet169 model achieved high training accuracy and demonstrated robust performance on the test dataset, demonstrating its powers in enhancing the efficiency of solar panel inspection and maintenance procedures. Numerous deep learning algorithms as suggested by [42], [43], [44], [45] including VGG16, DenseNet and MobileNetV3 requires large, annotated data sets and are not generalized to new defect types or real-world changes in brightness, particulate and specular reflection. In addition, models that have been trained using constrained data tend to perform poorly under diverse field conditions. Finally, privacy-preserving strategies, such as federated learning, are frequently compromised for detection accuracy, indicating the need for more balanced and dynamic models. Recent studies have analyzed the use of YOLO 9, 10, and 11 models for defect-detection in photovoltaic panels, using three datasets that include both thermal infrared and RGB images. The datasets cover a range of defects, including dirt, snow, bird droppings, and electrical-mechanical and structural failures. Comparative analysis shows that YOLOv10 and v11 have significantly better performance compared to traditional classifier, including the support vector machine (SVM) and region-based convolutional neural network (R-CNN), with YOLOv11 being the most efficient one, achieving a 89.7% precision, 87.7% recall, and 92.7 mean average precision (mAP) and as a result highlighting its superior accuracy and operational capability in detecting defects in photovoltaic panel [46], [47].

Recent research has made a valuable contribution to the methodology for detecting and classifying the defects in solar panels; however, several critical issues remain. The existing literature is limited primarily to focusing on a limited list of types of defects. It is based on relatively limited volume of images, typically fewer than 1,500 samples, thus limiting the ability of the models to apply to a wide range of environmental conditions. Even though network architectures like VGG16 and YOLO have shown significant promise in the deep-learning paradigm, their performance is highly dependent on large, labelled data sets and they often struggle with real-world variability. In addition, the potential advantages of hybrid models combining handcrafted feature engineering with deep-learning models are not studied thoroughly, which ends up in poor accuracy and robustness. To address these shortcomings, the current paper proposes a hybrid model, that combines handcrafted and deep feature extraction methods and is further supported by extensive data augmentation. The proposed strategy will increase datasets, thereby improving the effectiveness and accuracy of defect detection systems.

Materials and Methods

Data Acquisition:

The PV Panel Defect Dataset is publicly available on KAGGLE, generally designed and developed for detection and classification of different faults in photovoltaic (PV) panels. It contains six classes representing real-world conditions for solar panels, including clean, dusty, bird-dropping, electrical damage, physical damage and snow-covered panels. Each image in the dataset has distinct visual qualities that influence solar energy production, creating a dynamic and demanding setting for model training and evaluation.

The dataset contains high-resolution RGB photographs taken under various lighting and environmental conditions, ensuring the resilience and generalizability of deep learning and computer vision-based fault detection systems. Table 1 presents an overview of the PV Panel Defect Dataset.

ClassDescriptionNo. of Images
CleanImages of clean solar panels.289
Snow-coveredImages of solar panels covered with snow.262
DustyImages of solar panels covered with dust275
Electrical FaultImages of solar panels damaged due to electrical fault.225
Physical DamageImages of physically damaged solar panels225
Bird DroppingsImages of solar panels covered with bird droppings298

Table 1. PV Panel Defect Dataset Summary

Data Preprocessing:

During the preprocessing phase, all augmented images undergo two transformations: noise reduction and image enhancement.

Noise Reduction

The methods employed to improve the image quality are Non-Local Means (NLM) denoising and Bilateral filtering. Non-Local Means (NLM) denoising was effective at reducing noise without losing important textural and structural features needed to detect defects accurately, whereas Bilateral filtering improved image smoothness while preserving important edge details. A bilateral filter with a diameter of 9, sigmaColor of 75, and sigmaSpace of 75 was applied, which was sufficient to reduce image noise while retaining the important features such as edges. NLM denoising was also applied with the following parameters: h=10, hColor=10, the size of the template window = 7 and the size of the search window = 21. This overall denoising operation helped successfully remove noise without damaging any important textural and structural information of the image.

Image Enhancement

Two different image enhancement techniques were employed to improve contrast and brightness consistency, including Contrast Limited Adaptive Histogram Equalization (CLAHE) and Gamma Correction. Firstly, the images were downsized to 640x640 pixels, followed by converting the BGR to the LAB to increase luminance while preserving chromaticity. The L channel was then processed with Contrast-Limited Adaptive Histogram Equalizations (CLAHE) with a clip limit of 2.0 and a tile grid of 8×8. Lastly, the equalized L channel was reassembled with the respective A and B channels and then lastly converted to the BGR color model once again, giving the images a brighter contrast and clarity improving the ability to detect defects.

In addition, gamma correction was applied to improve image contrast. It is a nonlinear enhancement method that changes the value of pixel intensities according to power-law function. in accordance with a power-law function. The gamma value in our research was 1.5, used to lighten mid-tone areas without altering the overall image structure. A lookup table was built to effectively transform the original pixel values into gamma adjusted ones. All the images were resampled to 640x640 pixel size and subjected to this transformation, yielding a collection of contrast-enhanced images in a different direction.

Feature Extraction Network:

In our research work, we proposed a hybrid feature extraction model based on combination of handcrafted feature extraction techniques and deep learning-based algorithms to take advantage of both approaches. This combination will enable the model to capture features simultaneously both at low level (textual and edge-related data) and high level (semantic.), resulting in development of an accurate and efficient defect detection system.

Handcrafted Features

For handcrafted features extraction, we employed three different approaches: Linear Binary Patterns (LBP), Histogram of Gradients (HoG) and Gabor Filters. LBP is applied for textural analysis, HoG is used for capturing shape and edges, and Gabor filters are used for analyzing the spatial frequency and orientation. The combination of these methods can provide an extensive overview of surface aspects, thus enhancing overall classification accuracy.

Deep Learning Features

For deep features extraction, we used Convolutional Neural Networks (CNNs) with backbone of DenseNet169, that provides high level semantic features. This helped us to extract complex and intricate features that couldn’t be captured by handcrafted approaches.

Feature Fusion

To leverage the use of both handcrafted features and deep features, both were concatenated in the form of single feature matrix, which integrated low-level visual characteristics with high-level abstractions.

Classification:

The fused features are then fed to classification network, comprising three different classifiers: Support Vector Machines (SVMs), Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (Light GBM). Each of the ML classifiers are selected due to its unique strength in handling complex and high-dimensional. We employed SVM because of its excellent ability to build optimal decision boundaries between the classes by the application of kernel-based optimization. It is well suited for small-to-medium size datasets, and it is stronger at high classification accuracy even with restricted training samples, thus suitably applicable to PV defect classification where inter-class distances may be very fine.

Evaluation Metrics:

Performance of our posed model was being evaluated using different evaluation metrics such as accuracy, precision, recall and F1-score. These metrics were selected to provide a comprehensive assessment of the model’s classification capability. Accuracy is used to determine the general validity of predictions, whereas precision is used to measure the percentage of correct samples of defects out of all predicted defects. Recall is used to test the sensitivity of the model (in its ability to detect actual defects) and the F1-score (harmonic-mean between precision and recall) is a balanced measure of model performance. Collectively, these indicators provide a credible apparatus of assessing the efficiency and strength of the proposed PV panel fault detection system [48].

Accuracy=TP+TNTP+TN+FP+FN(1)\text{Accuracy} = \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}} \tag*{(1)}
Precision=TPTP+FP(2)\text{Precision} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}} \tag*{(2)}
Recall=TPTP+FN(3)\text{Recall} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}} \tag*{(3)}
F1-Score=2×Precision×RecallPrecision+Recall(4)\mathrm{F1\text{-}Score} = 2 \times\frac{\text{Precision} \times\text{Recall}}{\text{Precision}+\text{Recall}} \tag*{(4)}
Data augmentation techniques applied to PV panel images

Figure 03. Illustration of data augmentation techniques applied to PV panel images. (a) Clean panels, (b) Panels affected by bird droppings, (c) Snow-covered panels, (d) Panels with physical defects, and (e) Panels exhibiting electrical faults.

Additional augmented PV panel images

Figure 04.

Preprocessing techniques applied to PV panel images, showing panels covered with dust, snow-covered panels, panels with physical defects, panels exhibiting electrical faults, and panels affected by bird droppings

Figure 04. Illustration of preprocessing techniques applied to PV panel images. (a) Clean panels, (b) Panels covered with Dust, (c) Snow-covered panels, (d) Panels with physical defects, and (e) Panels exhibiting electrical faults, (f). Panels affected by bird droppings.

Overall architecture of the proposed hybrid PV panel defect detection framework

Figure 05. Overall architecture of the proposed hybrid PV panel defect detection framework.

Results and Discussion

This section demonstrates the experimental results of various handcrafted, deep learning, and hybrid methods for feature extraction in detecting defects in PV Panels. Different evaluation metrics such as accuracy, precision, recall and F1-score are used to evaluate the performance of proposed methods as presented in Table 1-Table 7.

Feature Extraction Method (s)ClassifiersAccuracy (%)Precision (%)Recall (%)F1-Score (%)
DenseNet-169 + LBP + HOGSVM89.2189.5089.2189.11
DenseNet-169 + LBP + HOGXGBoost86.6787.0186.6786.66
DenseNet-169 + LBP + HOGLGBM89.0189.0188.4788.46
DenseNet-169 + Gabor + HOGSVM88.8089.1688.8188.87
DenseNet-169 + Gabor + HOGXGBoost86.7487.0186.7486.78
DenseNet-169 + Gabor + HOGLGBM90.1290.6690.2390.16

Table 7. Comparative performance analysis of triple hybrid feature extraction techniques combined with various classifiers

Performance of Handcrafted Feature Extraction Techniques:

Performance of three different handcrafted feature extraction methods such as Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG), and Gabor Filters in combination with three machine learning classifiers, which include SVM, XGboost, and LightGBM (LGBM). is presented in Table 2. The LBP + LGBM combination had the highest accuracy of 91.85%, which points to the high discriminative power of LBP to local texture differences, including the pattern of dust or snow. Comparatively, HOG-based models recorded relatively low accuracy (approximately between 80 ~ 85%) because of their low sensitivity to subtle textual variations at different levels of illumination. The gabor filters were moderately sensitive to the background noise and lighting change as well as frequency and orientation analysis with an average accuracy of 89.38%.

Feature Extraction Method(s)Classifier (s)Accuracy (%)Precision (%)Recall (%)F1-Score (%)
LBPSVM90.0191.1690.8190.81
LBPXGBoost91.3591.5291.8591.84
LBPLGBM91.8592.0191.8591.84
GaborSVM86.5086.8386.6086.60
GaborXGBoost89.5089.8099.6089.50
GaborLGBM89.3890.0189.3089.36
HOGSVM84.7784.9984.7784.80
HOGXGBoost79.4281.0179.4279.44
HOGLGBM81.9782.3281.9781.94

Table 2. Comparative performance analysis of handcrafted feature extraction techniques combined with various classifiers

Performance of Dual Handcrafted Feature Extraction Techniques

As indicated by the summarized results in Table 3, the combination of two handcrafted features increased classification accuracy through the integration of complementary visual data. The best performance was obtained using the LightGBM classifier with LBP + Gabor combination with the accuracy of 94.23%. The enhanced performance is indicative of the fact that the combination of textural and spatial frequency features leads to better ability of the model to detect various PV panel defects, such as microcracks, bird droppings, and snow coverages. Nevertheless, the results of combinations like LBP + HOG and HOG + Gabor were moderate, indicating that the decision on feature fusion needs to take into consideration redundancy and inter-feature correlation.

Feature Extraction Method (s)ClassifiersAcc. (%)Precis. (%)Recall (%)F1-Score (%)
LBP + HOGSVM85.1085.8385.6685.66
LBP + HOGXGBoost83.7084.5084.3384.16
LBP + HOGLGBM88.1489.0188.1488.08
HOG + GaborSVM84.6085.3385.0185.01
HOG + GaborXGBoost84.0384.5084.1684.16
HOG + GaborLGBM86.0185.6085.3086.01
LBP + GaborSVM91.1792.0191.7691.76
LBP + GaborXGBoost91.4491.3391.4491.42
LBP + GaborLGBM94.2394.3394.2394.42

Table 3. Comparative performance analysis of dual handcrafted feature extraction techniques combined with various classifiers

Performance of Tripple Handcrafted Feature Extraction Techniques:

Table 4 shows the results of the LBP + HOG + Gabor feature combination. The model had the highest accuracy of 88.80% when using LightGBM classifier, which was less than the case of the dual feature. This decrease in performance can be explained by redundancy of features and higher dimensionality which caused noise and lower discriminative effectiveness of fused feature space. This configuration may yield more descriptive information although overfitting can be caused by over-fusion without dimensionality reduction particularly in situations where the dataset is small.

Performance of Deep Learning Feature Extraction Techniques:

DenseNet-169 showed the best results compared to any handcrafted model, which was found at Table 5. When using SVM, the accuracy exhibited by the model was 99.17% which is an indication that deep hierarchical features have the ability of well approximating both local and global spatial dependencies. The good performance of DenseNet-169 in terms of semantic patterns demonstrates its capacity to capture significant semantic patterns, which is very appropriate in complex defects of the surface. Nevertheless, these networks are very expensive in terms of computational resources, and some large balanced datasets are necessary to maximize their performance

Performance of Hybrid Feature Extraction Techniques:

A hybrid feature extraction framework was designed to integrate the benefits of both deep and handcrafted representations. The combined characteristics of DenseNet-169 and LBP, HOG and Gabor descriptors were categorized through the SVM, XGBoost and LightGBM classifiers. According to Table 6, dual hybrid systems, especially DenseNet-169 + Gabor (SVM), had the highest scores with 99.17% accuracy which was higher than all the other systems.

This fusion is effectually efficient to capture rich semantic properties as well as textural and frequency-related indications, offering great strength to change of lighting and any form of smog. On the same parameter, the DenseNet-169 + LBP and DenseNet-169 + HOG models also provided the highest accuracy (more than 98%), confirming that handcrafted features are useful when it comes to stabilizing and adapting the model.

Feature Extraction Method (s)ClassifiersAccuracy (%)Precision (%)Recall (%)F1-Score (%)
LBP + HOG + GaborSVM85.4386.1685.4385.42
LBP + HOG + GaborXGBoost88.3988.6688.3988.34
LBP + HOG + GaborLGBM88.8089.0188.8788.01

Table 4. Comparative performance analysis of triple handcrafted feature extraction techniques combined with various classifiers

Feature Extraction Method (s)ClassifiersAccuracy (%)Precision (%)Recall (%)F1-Score (%)
DenseNet-169 + LBPSVM99.0199.0098.0199.01
DenseNet-169 + LBPXGBoost97.0198.0098.0098.00
DenseNet-169 + LBPLGBM98.1098.0198.1098.10
DenseNet-169 + GaborSVM99.1799.1699.0199.17
DenseNet-169 + GaborXGBoost97.6997.6697.5597.66
DenseNet-169 + GaborLGBM97.9498.0197.9497.93
DenseNet-169 + HOGSVM92.7593.0192.7092.75
DenseNet-169 + HOGXGBoost97.2897.0197.2897.27
DenseNet-169 + HOGLGBM98.1898.3398.1798.18

Table 6. Comparative performance analysis of dual hybrid feature extraction techniques combined with various classifiers

Table 7 and Table 8 are triple and quadruple combinations of traditional as well as deep learning methods, combining DenseNet-169 with LBP and HOG with Gabor and tested using a variety of classifiers. These combinations obtained marginally lower accuracy with LightGBM performing best at 92.88 percent. The decline in performance relating to Table 5 could be attributed to a variety of reasons:

Feature Extraction Method (s)ClassifiersAccuracy (%)Precision (%)Recall (%)F1-Score (%)
DenseNet-169SVM99.1799.0199.1099.01
DenseNet-169XGBoost97.0397.0197.0397.02
DenseNet-169LGBM97.9498.1697.9497.93

Table 5. Comparative performance analysis of deep learning feature extraction techniques combined with various classifiers

Feature Extraction Method (s)ClassifiersAccuracy (%)Precision (%)Recall (%)F1-Score (%)
DenseNet-169 + LBP + HOG + GaborSVM89.0189.0189.0188.80
DenseNet-169 + LBP + HOG + GaborXGBoost90.0490.1690.0190.01
DenseNet-169 + LBP + HOG + GaborLGBM92.8893.0192.4292.45

Table 8. Comparative performance analysis of hybrid feature extraction techniques combined with various classifiers

  • Feature Redundancy: As the number of handcrafted features increases, more descriptors that duplicate information are added, but they do not add to the discriminative power.

  • Curse of Dimensionality: This is caused by the fact that additional features prepend more dimensions, and this can cause overfitting, particularly when there are few training examples. The expansion of the features to a high dimension demands larger datasets to be done effectively with generalization, which was not performed here.

  • Noise Propagation: With the aggregation of multiple hand-crafted descriptors, a small portion of noise inherent in each form of feature type will add up resulting in a misshapen feature distribution and lowering classification accuracy.

  • Computational Complexity: Increased feature volume leads to longer training times and higher computational costs, which may affect model convergence and stability.

CONCLUSIONS

This paper introduces a novel hybrid network by concatenating the handcrafted features extracted through LBP, HoG and G.F., and deep features extracted by using DenseNet169. The comparative analysis reveals that the fusion of the dual hybrid feature (DenseNet-169 + Gabor) feature is the most balanced trade-off between the accuracy, efficiency, and strength feature. This finding demonstrates that deep semantic and manual texture characteristics complement each other in solar defect analysis. The application of powerful classifier models such as XGBoost and LightGBM also enhance stability and stop overfitting with the help of effective gradient boosting and regularization. In general, the hybrid framework proposed has a better defect detecting accuracy, resistance, and flexibility that have a solid basis on the real-life use of the automated PV panels monitoring system. The triple and quadruple hybrid combinations had better performance than single model methods, suggesting that multiple feature extraction methods can result in higher model capabilities at distinguishing various defect characteristics. Nevertheless, they were slightly less accurate than the dual-hybrid combinations mainly because they are more complex in models, feature redundancy and could easily overfit because of the high dimensionality of the feature spaces. Despite these shortcomings, the results of the experiments prove the hypothesis that the suggested hybrid structure can be taken as the one that remains stable and shows consistent performance. The combination of high-level features and handcrafted features is effective in increasing generalization and detect defects reliably even in changing environmental conditions and illumination differences.

Limitations

Although the proposed hybrid PV panel defect detection framework has good outcomes, several limitations must be realized. First, the dataset, though it has been enriched to make it more diverse, is still relatively modest in the context of large-scale benchmarks that are most often necessary when training a deep learning model, which could constrain the extrapolation of the system to highly different real-world conditions, e.g., extreme lighting, heavy shadows, or infrequent defect patterns. Second, the handcrafted feature extractors adopted in this paper like LBP, HOG and Gabor filters are susceptible to noise and changes in illumination and can thus affect their performance when images are taken with different camera sensors or even in various environmental conditions. Third, hybrid feature fusion greatly expands the dimensionality of the feature space which may lead to redundancy and increase the computational cost of training and inference. This was also experienced in the reduced results on multi-feature fusion models. Finally, the system is yet to be tested against actual real-time deployment platforms, including UAV-based inspection pipelines or edge devices, which can provide more hardware limitations. Future research will aim at correcting these limitations by increasing the dataset, decreasing the redundancy of features, adding other types of imaging, and optimizing the model to be applied in real-time to an industrial environment.

Funding: Not Applicable

Data Availability: Data will be made available on request.

Declarations

Competing Interests: The authors declare no competing interests

Contributions

Muhammad Junaid Asif and Muhammad Saad Rafaqat conceived and designed the experiments.

Muhammad Junaid Asif and Muhammad Saad Rafaqat performed the experiments.

Muhammad Saad Rafaqat and Usman Nazakat analyzed the data, wrote the code, designed the software or performed the computation work.

Muhammad Junaid Asif, Usman Nazakat and Uzair Khan prepared the figures and/or tables.

Muhammad Junaid Asif and Muhammad Saad Rafaqat drafted the work or revised it critically for important content.

References

  1. [14]Y. Ledmaoui, A. El Maghraoui, M. El Aroussi and R. Saadane, "Enhanced fault detection in photovoltaic panels using cnn-based classification with pyqt5 implementation," Sensors, vol. 24, p. 7407, 2024.
  2. [15]M. Asad, W. Azeem, H. T. Mustafa, Y. Fang, J. Yang, Y. Zuo and W. Liu, "Inter-modality feature prediction through multimodal fusion for 3D shape defect detection," Neural Networks, vol. 193, p. 108057, 2026.
  3. [16]M. Asad, W. Azeem, A. A. Malik, H. Jiang, A. Ali, J. Yang and W. Liu, "3D-MMFN: Multi-level multimodal fusion network for 3D industrial image anomaly detection," Advanced Engineering Informatics, vol. 65, p. 103284, 2025.
  4. [17]M. Asad, W. Azeem, H. Jiang, H. Tayyab Mustafa, J. Yang and W. Liu, "2M3DF: Advancing 3D Industrial Defect Detection With Multi-Perspective Multimodal Fusion Network," IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, pp. 6803-6815, 2025.
  5. [18]F. H. K. Al-Zaabi, A. M. S. Al Washahi, R. K. M. Al-Maaini, M. K. Boddu and others, "Advancing Solar PV Component Inspection: Early Defect Detection with UAV Based Thermal Imaging and Machine Learning," in 2023 Middle East and North Africa Solar Conference (MENA-SC), 2023.
  6. [19]R. Singh, R. Kashyap and A. Kumar, "Prominent solution for solar panel defect detection using AI-based computer vision technology with IoT sensors in the solar panel manufacturing industry," International Journal of Information Technology, p. 1–17, 2024.
  7. [20]U. Hijjawi, S. Lakshminarayana, T. Xu, G. P. M. Fierro and M. Rahman, "A review of automated solar photovoltaic defect detection systems: Approaches, challenges, and future orientations," Solar Energy, vol. 266, p. 112186, 2023.
  8. [21]Y. S. Afridi, K. Ahmad and L. Hassan, "Artificial intelligence based prognostic maintenance of renewable energy systems: A review of techniques, challenges, and future research directions," International Journal of Energy Research, vol. 46, p. 21619–21642, 2022.
  9. [22]A. Lenarczyk, PV Panel Defect Dataset, 2025.
  10. [23]Z. Sedaghatjoo, H. Hosseinzadeh and B. S. Bigham, "Local binary pattern (lbp) optimization for feature extraction," arXiv preprint arXiv:2407.18665, 2024.arxiv.org/abs/2407.18665
  11. [24]S. Divya, L. P. Suresh and A. John, "Image feature generation using binary patterns—LBP, SLBP and GBP," in ICT Analysis and Applications, Springer, 2022, p. 233–239.
  12. [25]N. Kaur, N. Nazir and others, "A review of local binary pattern based texture feature extraction," in 2021 9th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO), 2021.
  13. [26]M. Pietikäinen, "Image analysis with local binary patterns," in Scandinavian Conference on Image Analysis, 2005.
  14. [27]V. A. Saeed, "A framework for recognition of facial expression using HOG features," International Journal of Mathematics, Statistics, and Computer Science, vol. 2, p. 1–8, 2024.
  15. [28]E. A. Siddiqui, V. Chaurasia and M. Shandilya, "Detection and classification of lung cancer computed tomography images using a novel improved deep belief network with Gabor filters," Chemometrics and Intelligent Laboratory Systems, vol. 235, p. 104763, 2023.
  16. [29]M. J. Asif, "Crowd Scene Analysis Using Deep Learning Techniques," arXiv preprint arXiv:2505.08834, 2025.arxiv.org/abs/2505.08834
  17. [30]A. A. Elngar, M. Arafa, A. Fathy, B. Moustafa, O. Mahmoud, M. Shaban and N. Fawzy, "Image classification based on CNN: a survey," Journal of Cybersecurity and Information Management, vol. 6, p. 18–50, 2021.
  18. [31]M. A. Chandra and S. Bedi, "Survey on SVM and their application in image classification," International Journal of Information Technology, vol. 13, p. 1–11, 2021.
  19. [32]H. Tanveer, M. A. Adam, M. A. Khan, M. A. Ali and A. Shakoor, "Analyzing the performance and efficiency of machine learning algorithms, such as deep learning, decision trees, or support vector machines, on various datasets and applications," The asian bulletin of big data management, vol. 3, p. 126–136, 2023.
  20. [33]A. Mirakowski, "Comparison of the Effectiveness of ANN and CNN in Image Classification," in International Conference on Computational Collective Intelligence, 2024.
  21. [34]S. Sharma and K. Guleria, "Deep learning models for image classification: comparison and applications," in 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE), 2022.
  22. [35]P. Dinesh, A. S. Vickram and P. Kalyanasundaram, "Medical image prediction for diagnosis of breast cancer disease comparing the machine learning algorithms: SVM, KNN, logistic regression, random forest and decision tree to measure accuracy," in AIP Conference Proceedings, 2024.
  23. [36]M. S. Islam, S. Sultana, F. A. Farid, M. N. Islam, M. Rashid, B. S. Bari, N. Hashim and M. N. Husen, "Multimodal hybrid deep learning approach to detect tomato leaf disease using attention based dilated convolution feature extractor with logistic regression classification," Sensors, vol. 22, p. 6079, 2022.
  24. [37]A. Maleki, M. Raahemi and H. Nasiri, "Breast cancer diagnosis from histopathology images using deep neural network and XGBoost," Biomedical Signal Processing and Control, vol. 86, p. 105152, 2023.
  25. [38]D.-M. Tsai, S.-C. Wu and W.-Y. Chiu, "Defect detection in solar modules using ICA basis images," IEEE Transactions on Industrial Informatics, vol. 9, p. 122–131, 2012.
  26. [39]X. Zhang, H. Sun, Y. Zhou, J. Xi and M. Li, "A novel method for surface defect detection of photovoltaic module based on independent component analysis," Mathematical problems in engineering, vol. 2013, p. 520568, 2013.
  27. [40]A. Shaik, A. Balasundaram, L. S. Kakarla and N. Murugan, "Deep learning-based detection and segmentation of damage in solar panels," Automation, vol. 5, p. 128–150, 2024.
  28. [41]N. Appavu, "AI-Powered VGG16-based Method for Detecting, Classifying, and Visualizing Solar Panel Defects," in 2025 4th International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE), 2025.
  29. [42]S. Mansurov, Z. Çetin, E. Aslan and Y. Özüpak, "A Deep Learning Approach for Fault Detection in Photovoltaic Systems Using MobileNetV3," Gazi University Journal of Science Part A: Engineering and Innovation, vol. 12, p. 197–212, 2025.
  30. [43]K. Kazemi, "Federated Transfer Learning for Image-Based Solar Panel Fault Detection," in 2025 12th Iranian Conference on Renewable Energies and Distributed Generation (ICREDG), 2025.
  31. [44]S. R. Joshua, S. Park and K. Kwon, "Visual Diagnostics: Deep Learning with DenseNet121 for Identifying Faults in Solar Panels," in 2024 15th International Conference on Information and Communication Technology Conference (ICTC), 2024.
  32. [45]S. Nagar, M. K. Mishra and P. K. Rai, "Using Densenet169 for Image-Based Classification of Solar Panel Defects," in 2024 7th International Conference on Contemporary Computing and Informatics (IC3I), 2024.
  33. [46]A. Ghahremani, S. D. Adams, M. Norton, S. Y. Khoo and A. Z. Kouzani, "Detecting defects in solar panels using the YOLO v10 and v11 algorithms," Electronics, vol. 14, p. 344, 2025.
  34. [47]A. Mohamed, Y. Nacera, B. Ahcene, A. Teta, E. O. Belabbaci, A. Rabehi, Y. A. Alsahab and M. Benghanem, "Optimized YOLOv8 based model for photovoltaic defect detection in electroluminescence images," Scientific Reports, vol. 15, p. 32955, 2025.
  35. [48]H. Khalid, S. Saqib, M. J. Asif and D. A. Dewi, "Strategic Customer Segmentation: Harnessing Machine Learning For Retaining Satisfied Customers," Lahore Garrison University Research Journal of Computer Science and Information Technology, vol. 8, p. 31–41, 2024.

Paper details

vlnc:W_62y5yx45
Published
Added
License
CC BY 4.0
Usage

views · 18

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents