Defect-Specific Physical-Prior-Guided Synthesis of Photovoltaic RGB Defect Images with Pixel-Level Annotations

Reliable visual inspection of photovoltaic modules is limited by the scarcity of diverse defect images and pixel-level annotations. This study proposes a defect-specific RGB image synthesis framework combining local feature disentanglement and physical-prior-guided modeling for shadow, debris, dust, and broken defects, while simultaneously generating spatially aligned pixel-level masks. The generated samples are evaluated in terms of image quality, distribution consistency, structural preservation, and downstream utility. Compared with CycleGAN, MUNIT, and SDXL Inpainting, the proposed method achieves the lowest BRISQUE scores across all four categories, the lowest class-wise Defect-aware Fréchet Distance (DFD) for shadow, debris, and broken, and the lowest MMD2-RBF values across all categories. With generated samples added to real training data, the five-run average classification Accuracy and Macro-F1 reach 88.96% and 87.83%, respectively, while detection performance is also improved. Moreover, a segmentation model trained solely on generated image–mask pairs achieves an mIoU of 60.15% on unseen real images. These results demonstrate that the proposed framework provides structurally consistent, task-useful defect samples and directly usable pixel-level supervision for photovoltaic visual inspection.