Physics-informed transformer for pseudo-color infrared image super-resolution with application to insulator detection

The accelerated deployment of smart grids has intensified the need for real-time insulator monitoring and health diagnostics, particularly for the early detection of thermal faults. While infrared thermography enables non-contact fault localization, industrial thermal imagers are often constrained by insufficient spatial resolution and inherent sensor artifacts, compromising both image quality and thermal interpretability. To address information compression and undersampling while recovering relative thermal-pattern information and structural representations, we introduce the F idelity-aware I nfra R ed E nhancement Net work ( FIRE-Net ), a physics-informed transformer for artificial intelligence (AI)-driven super-resolution of pseudo-color infrared images. The model incorporates an infrared denoising block to suppress sensor noise, employs learnable unsharp masking and high-frequency attention modules to recover structural details, and is supervised by a thermal-consistency loss function that constrains relative thermal patterns in a normalized palette-derived temperature-proxy domain. On the insulator infrared test set, FIRE-Net achieves peak signal-to-noise ratio values of 31.0021 dB (dB) (×2) and 27.6765 dB (×4), with structural similarity index scores of 0.9143 and 0.8591, respectively. The model also improves downstream detection performance, attaining mean average precision at intersection over union = 0.50 values of 0.857 (×2) and 0.904 (×4), representing improvements of 0.102 and 0.111 over the low-resolution baseline. Its cross-scenario performance is further evaluated on a hold-out photovoltaic dataset. We further analyze computational efficiency and outline future research directions, including domain-extended training and lightweight deployment. The results demonstrate that incorporating physics-inspired priors into super-resolution models effectively improves normalized temperature-proxy consistency and practical detection performance. The source code is publicly available at: https://github.com/Sevil-cy/FIRE-Net .