Scientists in China have developed a new deep learning model based on the so-called ‘You Only Look Once’ algorithm, which requires only one forward propagation pass through the neural network to make predictions.
Researchers in China have developed a new deep learning model to detect defects in photovoltaic panels.
The approach uses high-resolution visible light imaging to identify defects using an algorithm based on the You Only Look Once (YOLO) deep learning framework, which makes predictions with just a single forward pass through the neural network.
The new method also includes a hybrid attention transformer in the cross-stage partial bottleneck, combined with two convolutional backbones.
“Building on the foundation of YOLOv8, we introduce a new HAT-C2f neural network module and redesign the backbone component,” the researchers explain. “In the neck section, the conventional C2f module is replaced by the RepNCSPELAN4 architecture – an efficient layer design with specified channel sizes, repetitions and convolutions – and a SKAttention mechanism is added before the detection head.”
The HAT-C2f module improves the backbone’s ability to extract fine image details, while the integration of RepNCSPELAN4 into the neck improves feature aggregation, allowing the system to detect objects of different sizes. By adding SKAttention to the sensing head, the model can adapt to different scales.
The model, called YOLO-HRS, was evaluated based on 6,500 labeled visible-light images from the data science competition platform and online community Kaggle, divided into four classes: clean, dust, cracked and bird droppings. About 80% of the images were used for training and the remaining 20% for validation. YOLO-HRS has been tested with previous YOLO models and advanced object detection algorithms. Ablation studies were also performed, where individual components of the model were independently evaluated.
The analysis showed that the YOLO-HRS achieved an accuracy of 86.87%, a recall of 84.6%, an average average precision (mAP) at an intersection over the union (IoU) of 0.5 of 88.98%, and [email protected]:0.95 of 77.08%.
Ablation studies showed substantial performance improvements in object detection, while comparisons with other models showed that YOLO-HRS performed better. For example at [email protected]only YOLOX approached its performance with 85.59%, while RT-DETR, Faster-RCNN, NanoDet and RetinaNet scored 79.34%, 66.29%, 64.16% and 69.54% respectively.
Compared to baseline YOLOv8, YOLO-HRS showed a 3% improvement [email protected].
“In summary, the model was experimentally validated using visible light images of PV panels, confirming its high reliability and precision,” the team concluded. “YOLO-HRS provides accurate defect detection in visible light PV panels, providing a more reliable solution for real-world applications.”
Looking ahead, the researchers plan to further optimize and expand YOLO-HRS.
“First, we will refine and test the model architecture on various datasets, including infrared and electroluminescence images, to evaluate its applicability in different scenarios. Second, we aim to develop lightweight structures, including novel down-sampling and feature extraction methods, to improve the balance between accuracy and model size. Finally, exploring cross-domain applications and self-supervised learning techniques could reduce dependence on large annotated datasets,” they said.
The new method was introduced in “A novel deep learning model for defect detection in photovoltaic panels using visible light imaging”, published in Technical applications of artificial intelligence. The research was conducted by scientists from China’s Zhejiang University of Finance and Economics and Hangzhou Dianzi University.
This content is copyrighted and may not be reused. If you would like to collaborate with us and reuse some of our content, please contact: editors@pv-magazine.com.
