GDDS: A Single Domain Generalized Defect Detection Frame of Open World Scenario using Gather and Distribute Domain-shift Suppression Network

TL;DR

This work tackles open-world photovoltaic defect detection under endogenous shift by introducing GDDS, a single-domain generalized, one-stage detector. It integrates the DeepSpine Module for rich multi-scale context and the Gather and Distribution Module for cross-layer global-local feature interaction, augmented by a Normalized Wasserstein Distance (NWD) based loss to stabilize bounding-box regression. The loss combines $L_{CIoU}$ with $L_{NWD}$, where $L_{NWD} = 1 - NWD(N^P,N^G)$ and $NWD$ relies on $W_2^2$ between Gaussian bounding-box representations, with $\beta$ tuned to 0.5. Extensive experiments on the EL endogenous shift dataset and a photovoltaic infrared image dataset show that GDDS achieves superior generalization to endogenous drift and higher speed than competing methods, thanks to the integrated multi-scale fusion and robust loss. These results suggest a practical, scalable path for reliable PV module inspection in real-world manufacturing and maintenance workflows, where data distributions continually vary.

Abstract

Efficient and intelligent surface defect detection of photovoltaic modules is crucial for improving the quality of photovoltaic modules and ensuring the reliable operation of large-scale infrastructure. However, the scenario characteristics of data distribution deviation make the construction of defect detection models for open world scenarios such as photovoltaic manufacturing and power plant inspections a challenge. Therefore, we propose the Gather and Distribute Domain shift Suppression Network (GDDS). It adopts a single domain generalized method that is completely independent of the test samples to address the problem of distribution shift. Using a one-stage network as the baseline network breaks through the limitations of traditional domain generalization methods that typically use two-stage networks. It not only balances detection accuracy and speed but also simplifies the model deployment and application process. The GDDS includes two modules: DeepSpine Module and Gather and Distribute Module. Specifically, the DeepSpine Module applies a wider range of contextual information and suppresses background style shift by acquiring and concatenating multi-scale features. The Gather and Distribute Module collects and distributes global information to achieve cross layer interactive learning of multi-scale channel features and suppress defect instance shift. Furthermore, the GDDS utilizes normalized Wasserstein distance for similarity measurement, reducing measurement errors caused by bounding box position deviations. We conducted a comprehensive evaluation of GDDS on the EL endogenous shift dataset and Photovoltaic inspection infrared image dataset. The experimental results showed that GDDS can adapt to defect detection in open world scenarios faster and better than other state-of-the-art methods.

Figures (13)

• Figure 1: Endogenous shift phenomenon in open world scenarios. (a) In open world scenarios, examples of images acquired from different domains are shown. The left image displays EL images obtained from various production lines in photovoltaic manufacturing scenes. Defective images obtained from different production lines exhibit slight differences in texture, clarity, brightness, and other aspects. The right image shows infrared images acquired at different heights in photovoltaic inspection scenarios. (b) The left and right images respectively depict probability density distribution curves of data from different domains in photovoltaic manufacturing and photovoltaic inspection scenarios. This illustrates the differences in data distribution among different domains. (c) Different detectors were trained on the source domain and tested on multiple other domains. The left and right figures respectively represent scenarios of photovoltaic manufacturing and photovoltaic plant inspection.
• Figure 2: The left and right figures respectively depict the distribution of defect sizes in photovoltaic manufacturing and photovoltaic plant inspection scenarios. The horizontal axis represents the ratio of the actual width of target defect bounding boxes to the width of EL images, while the vertical axis represents the ratio of the actual height of target defect bounding boxes to the height of EL images.
• Figure 3: Overall network architecture.
• Figure 4: Overall network architecture.
• Figure 5: The impact of bounding box deviation on different loss calculation results. $G$ represents the true bounding box, and $P$ represents the predicted bounding box. Points of different colors represent defect targets of different sizes. The horizontal axis values represent the pixel difference between the predicted bounding box and the true bounding box center points, while the vertical axis values represent the corresponding measurement values.