YOLO-WAD for Small-Defect Detection Boost in Photovoltaic Modules

Wang Yin, Yun Wang, Gang Xie, Zhicheng ZhaoSensors 2025 · DOI

TL;DR

This paper introduces YOLO-WAD, an enhanced object detection model built upon YOLOv10n, specifically designed to improve the detection of small, critical defects in photovoltaic (PV) modules. By incorporating novel modules like Wavelet Transform Convolution (WTConv), Attentional Scale Sequence Fusion (ASF), Efficient Multi-Scale Attention (EMA), and a Dynamic Head (DyHead), YOLO-WAD achieves a significant boost in accuracy for detecting challenging small defects, outperforming existing state-of-the-art models.

What Problem Does This Paper Address?

The performance and lifespan of solar cells are directly tied to the integrity of photovoltaic modules. However, these modules are susceptible to various defects, particularly small ones like "hot spots" and cracks, which can lead to energy loss, reduced efficiency, and even system failure. Accurately detecting these subtle defects is a significant challenge for current computer vision systems, especially when they are small, have weak feature expressions, and can be easily obscured by background noise. Existing object detection models often struggle with the multi-scale nature of these defects and the need for high real-time performance in industrial settings.

Methods

The authors propose YOLO-WAD, an improved version of YOLOv10n, by introducing several novel architectural components:

  • C2f-WTConv (Backbone): Replaces the standard C2f module with a C2f module incorporating Wavelet Transform Convolution (WTConv). WTConv uses wavelet transforms to decompose input images into different frequency components, allowing for better extraction of fine-grained details and features from small targets.
  • ASF (Neck): Introduces an Attentional Scale Sequence Fusion (ASF) module in the neck. ASF effectively fuses features from different levels (P2-P5) extracted by the backbone, enhancing the model's ability to capture both global context and local details crucial for small object detection.
  • C2f-EMA (Neck): Integrates an Efficient Multi-Scale Attention (EMA) mechanism into a C2f-like structure (C2f-EMA). This module redistributes feature weights, prioritizing relevant features and spatial details across image channels to improve feature extraction.
  • DyHead (Head): Employs the Dynamic Head (DyHead) detection head. DyHead offers comprehensive scale, spatial, and channel awareness, significantly improving the model's capacity for classifying and localizing small-target defects.

Datasets

The experiments were conducted on a custom dataset of 2272 infrared aerial images captured at a low altitude (10-15m) of PV modules. These images were manually annotated with four defect types: fracture, hot spot, plant, and battery string. To address sample imbalance, specific undersampling and oversampling strategies were applied. The dataset was split into 2037 images for training, 224 for validation, and 11 for testing.

Key Insights and Ideas

  • Wavelet Transforms for Fine Details: The integration of WTConv highlights the power of multi-frequency analysis in capturing subtle defect features that might be missed by standard convolutional filters.
  • Multi-Scale Fusion is Crucial: The ASF module's success underscores the importance of effectively fusing features from various scales to build a robust representation for small objects.
  • Attention for Prioritization: The EMA and DyHead modules demonstrate how attention mechanisms can dynamically focus on the most informative features and spatial regions, leading to more accurate detection.
  • Dynamic Heads for Adaptability: The DyHead's ability to adapt its processing based on feature characteristics is key to handling the diverse nature of defects and complex backgrounds.
  • Holistic Improvement: The paper shows that combining these specialized modules leads to synergistic improvements, with the full YOLO-WAD model significantly outperforming its individual components and baseline models.

Key Results

The YOLO-WAD model achieved an overall accuracy of 95.6% on their dataset. Crucially, its small-target defect detection accuracy reached 86.3%, which is a significant improvement:

  • 4.1% higher than the baseline YOLOv10n.
  • 9.5% higher than current mainstream models.

In specific comparisons:

  • YOLO-WAD achieved the highest precision (93.4%) and recall (92.7%) among the tested models.
  • For "hot spot" detection, YOLO-WAD achieved an accuracy of 86.3%, substantially higher than other models.
  • Visualizations showed YOLO-WAD providing higher confidence and stability in detecting small targets with fewer false detections compared to RT-DETR-Resnet101 and YOLOv10n.

Limitations and Future Work

  • Computational Cost: While the paper emphasizes efficiency, the introduction of multiple advanced modules might increase computational complexity compared to simpler models. Future work could focus on further optimizing computational resource requirements to adapt the algorithm to a wider range of application scenarios.
  • Generalization to Other Scenarios: The paper mentions a related project on screen defect detection where similar models struggled with increased complexity and defect categories. Future research will aim to optimize the model for such scenarios and overcome limitations in clearly identifying every defect type.
  • Data Availability: The dataset used in this study is not publicly available due to confidentiality agreements with a cooperating unit.
  • Further Optimization: The authors suggest that future research will be committed to further optimizing model performance and promoting the continuous development of PV module defect detection technology.

Summary updated

Abstract

The performance of photovoltaic modules determines the lifetime of solar cells; however, accurate detection remains a challenge when facing smaller defects. To address this problem, in this paper, we propose a YOLO-WAD model based on YOLOv10n. Firstly, we replace C2f (CSP bottleneck with two convolutions) with C2f-WTConv (CSP bottleneck with two convolutions-wavelet transform convolution) in the backbone network to enlarge the receptive field and better extract the features of small-target defects (hot spots). Secondly, an ASF structure is introduced in the neck, which effectively fuses the different levels of output features extracted by the backbone network and enhances the model's ability to detect small objects. Subsequently, an additional detection layer is added to the neck, and C2f is replaced by C2f-EMA (CSP bottleneck with two convolutions-efficient multi-scale attention mechanism), which can redistribute feature weights and prioritize relevant features and spatial details across image channels to improve feature extraction. Finally, the DyHead (dynamic head) detection head is introduced, which enables comprehensive scale, spatial, and channel awareness. This greatly enhances the model's ability to classify and localize small-target defects. The experimental results show that YOLO-WAD detects our dataset with an overall accuracy of 95.6%, with the small-target defect detection accuracy reaching 86.3%, which is 4.1% and 9.5% higher than YOLOv10n and current mainstream models, verifying the feasibility of our algorithm.
BibTeX
@inproceedings{yin2025yolo,
  author = {Wang Yin and Yun Wang and Gang Xie and Zhicheng Zhao},
  title = {{YOLO-WAD for Small-Defect Detection Boost in Photovoltaic Modules}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25061755},
}
Figure