PNG-Net: Physics Noise-Guided Network for Infrared Image Nonuniformity Correction

Infrared thermal imaging is widely used for industrial fault diagnosis because it enables non-contact measurements and reveals temperature anomalies. However, detector nonuniformity degrades image quality and can affect subsequent analysis. Existing nonuniformity correction (NUC) methods face a trade-off between suppressing structured noise and preserving fine image details. To address this issue, a Physics Noise-Guided Network (PNG-Net) is proposed for infrared image nonuniformity correction. First, the input image is decomposed into low-, mid-, and high-frequency components using the fast Fourier transform (FFT), and multi-frequency information is adaptively integrated through a dynamic frequency-band selection module. Second, a learnable PSD-like noise estimator derives a spectral noise prior from the Fourier magnitude spectrum, and an anisotropic mask adds directional information about stripe noise. During feature restoration, the estimated prior noise is incorporated into the proposed Noise-Guided Stack to conditionally modulate multi-scale residual features, thereby progressively recovering structural details and thermal gradients through hierarchical encoder-decoder skip connections. Finally, global residual learning is adopted to predict the nonuniformity correction residual. On the synthetic infrared nonuniformity dataset, PNG-Net achieves a PSNR of 48.92 dB and an SSIM of 0.990, outperforming the compared methods in these two metrics.