DSDINet: Deep semantic decoupling and integration photovoltaic solar cell defect segmentation network

Defects in photovoltaic (PV) solar cells can significantly degrade photoelectric conversion efficiency and shorten service life. Solar cell defect segmentation enables accurate localization of defects, thereby supporting the intelligent operation and maintenance of photovoltaic power plants. However, solar cell defect segmentation faces two major challenges. On the one hand, multiscale low-contrast defect information is difficult to extract from the foreground features due to distracting information such as flocculent regions and busbars in the background. On the other hand, high-level semantic information is prone to being lost in the feature fusion process, leading to the missed detection of some edge-located defects and tiny defects. To address these two challenges, a deep semantic decoupling and integration defect segmentation network (DSDINet) is introduced which can effectively extract rich multiscale information from the foreground features and improve the representation capability of high-level features. Specifically, we design a Semantic Information Decoupling and Aggregation (SIDA) module to decouple and process the foreground and background of enriched fine-grained low-level feature maps, and then fuse them with high-level features. For defects of different scales, a Multiscale Feature Exploration (MFE) module is embedded as a built-in submodule of SIDA to mine discriminative features from decoupled foreground defect regions. In addition, to enhance the features of edge-located defects and tiny defects, we propose a Global and Local Attention (GLA) module that simultaneously models global semantic dependencies and local fine-grained details. Experimental results on the public polycrystalline silicon defect segmentation dataset PSCDE demonstrate that DSDINet achieves state-of-the-art performance, demonstrating its effectiveness for PV solar cell defect segmentation. Source code is available at: https://github.com/Taraseu/PV-Defect-Segmentation-DSDINet.