Table 2 Overview of related studies in PV panel defect detection research.

²¹

Knowledge distillation with YOLOX and CSPHN networks.

Employs a bi-branch collaborative training method with knowledge distillation for PV hot-spot identification, emphasizing improvements in detection precision and computational efficiency.

Attained an AP50 metric of 82.2%, indicating rapid and precise hot-spot defect identification over diverse adverse situations.

²²

DL using surveillance camera images.

Created SoilingEdge, a DL methodology for estimating power loss from the soiling of PV panels from images obtained from edge devices such as security cameras.

Accomplished precise power loss assessment from soiling by advanced image processing techniques on edge platforms, enabling reliable inference for outdoor PV monitoring.

²³

Ghost Convolution Including YOLOv5 with BottleneckCSP and Tiny Target Prediction Head (GBH-YOLOv5)

The study introduces GBH-YOLOv5, a new method for detecting tiny defects on PV panels using advanced image processing.

Achieving at least a 27.8% increase in accuracy compared to state-of-the-art methods.

²⁴

End-to-end approach using Red Green Blue (RGB) imagery from Unmanned Aerial Vehicles (UAVs) and YOLOv4 architecture for DL-based defect detection.

The study develops a method using UAV-acquired RGB imagery and YOLOv4 to detect, identify, and locate defects in solar PV modules.

The defect detection achieved 83% accuracy on the validation set and 73% on the test set in large-scale PV installations.

²⁵

DL using ResNet152-Xception and a coordinate attention mechanism.

Deep-learning model using ResNet152-Xception and attention for PV cell defect detection, addressing data scarcity and imbalance.

The model achieves 96.17% accuracy in binary classification and 92.13% in multiclassification of PV cell defects, outperforming several common models.

²⁶

U-net semantic segmentation for EL image analysis of PV modules.

The study utilizes U-net architecture to detect and quantify defects in solar PV modules through EL imaging, spanning various module designs and image qualities.

Defects in silicon wafer-based solar cells that are mono- or multi-crystalline are efficiently identified and measured by the U-net model.

²⁷

Synchronized Thermography (ST) using a portable IR-camera.

The research investigates PV fault identification by Infrared Radiation (IR) thermography, emphasizing the adaptation of observations to varying outside situations.

IR thermography is effective in detecting PV panel defects in various harsh conditions, providing consistent information with common conditions.

²⁸

Optical Stepped Thermography with post-data processing algorithms.

The study utilizes optical stepped thermography, halogen lamps, and IR camera monitoring to enhance defect signatures in PV panels.

This method effectively identifies defects in PV panels, with processed images being more evaluative than raw thermal images.

²⁹

Support Vector Machine (SVM) and Back Propagation Neural Network (BPNN)

identification of solar cell microcracks by EL image analysis.

Achieved 92.67% accuracy with SVM and 93.67% with BPNN in classifying solar cells as cracked or not.

³⁰

Steerable evidence filtering, local thresholding, and a minimum spanning tree for crack detection.

Identifying fissures in multicrystalline solar cells with improved contrast crack saliency maps and segmentation techniques for comprehensive crack extraction.

94.4% average detection rate for various types of cracks.

³¹

Machine learning (ML) - SVM, Random Forest (RF) with Hough transform for pattern detection.

Automatic splitting of EL images into cells, defect detection, and feature computation for precise defect categorization in PV panels.

Achieved high accuracy (0.997) but lower recall (0.274) using SVM in defect type identification.