TSF-Net: an efficient multiscale time-space-frequency unified network for infrared video nonuniformity correction

Non-uniformity correction (NUC) is essential for infrared imaging. Deep learning has emerged as a promising solution because of its nonlinear modeling capability and scene adaptability. However, existing deep-learning-based NUC methods often struggle to balance correction performance and computational efficiency. Single-frame convolutional neural network (CNN) methods cannot exploit temporal information, whereas multiframe methods often incur high computational costs and perform spatiotemporal fusion at only one scale. To address these limitations, we propose an efficient multiscale time-space-frequency unified network (TSF-Net) for infrared video NUC. TSF-Net performs joint correction across three domains with modules at different resolution scales: (1) in the temporal domain, a lightweight spatiotemporal gated memory (STGM) module efficiently propagates and fuses inter-frame information through an element-wise recurrent state update; (2) in the frequency domain, the discrete wavelet transform (DWT) is used for downsampling to preserve high-frequency texture information during feature extraction; and (3) in the spatial domain, a spatial stripe context module (SSCM) exploits the column-stripe noise characteristics of infrared focal plane arrays. Across all test settings under four levels of synthetically generated nonuniformity noise, TSF-Net achieves the highest PSNR and SSIM values among the evaluated methods. Compared with the best-performing single-frame method, LSC-CNN, TSF-Net improves the average PSNR and SSIM by 2.38 dB and 0.0252, respectively. It also outperforms the best-performing multiframe method, PDB-Unet, by ∼0.47 dB in average PSNR and 0.0089 in average SSIM while using 97.6% fewer parameters. With only 0.25 M parameters and 20.3 GFLOPs per 640×480 frame, TSF-Net demonstrates a favorable balance between correction performance and computational efficiency.