Adaptive generative augmentation for long-tailed photovoltaic infrared defect recognition

This is the authors' abstract. We don't add key points for this paper.

Photovoltaic infrared defect recognition is often affected by highly imbalanced class distributions, where underrepresented defect categories provide limited training data for supervised classifiers. To improve recognition under this naturally long-tailed setting, this study proposes ADLS-IRGAN, a data-centric generative augmentation method that supplements tail classes without modifying the downstream classifier. An exploratory metric–utility analysis on two binary defect-recognition tasks examines associations between generated-image quality metrics and downstream recognition utility. FID, MS-SSIM, and LPIPS show the most consistent associations and serve as empirical cues for model design. ADLS-IRGAN introduces an adaptive deformable latent-space mechanism to improve latent sampling flexibility and combines structural and perceptual constraints to enhance the diversity and class-related characteristics of generated thermograms. Experiments are conducted on a duplicate-controlled photovoltaic infrared dataset containing 19,972 images from 12 classes using five fixed train/validation/test splits. On the naturally imbalanced 12-class benchmark, ADLS-IRGAN achieves an Accuracy of 0.8332, Macro-F1 of 0.7082, Weighted-F1 of 0.8290, Balanced Accuracy of 0.6909, and Tail Macro-F1 of 0.6322. Relative to real-only training, Macro-F1, Balanced Accuracy, and Tail Macro-F1 improve by 12.65%, 13.30%, and 26.04%, respectively. Compared with Baseline-GAN augmentation, the corresponding gains are 4.93%, 4.63%, and 9.49%. Component-wise ablation shows that removing latent-space or structure–perception components reduces downstream recognition performance. Positive gains are observed with MobileNetV3-Large and Swin-T, indicating that the augmentation benefit is not restricted to ResNet50. ADLS-IRGAN provides a practical data-side augmentation strategy for improving recognition of underrepresented photovoltaic infrared defect categories.