Infrared Image Deturbulence Restoration Using Degradation Parameter-Assisted Wide & Deep Learning

TL;DR

Infrared deturbulence introduces spatially varying geometric distortions and blur that challenge conventional restoration. The authors present DparNet, a parameter-assisted wide & deep network that learns a degradation parameter matrix $P$ from degraded sequences via a parameter-prediction module and uses $P$ to guide a parallel wide branch while a BRNN-based deep branch performs multi-frame restoration. The two branches are fused to yield the final restored sequence, achieving state-of-the-art restoration quality with a modest increase in model size. Experiments on a dedicated infrared deturbulence dataset (49,744 images) and a denoising dataset (109,536 images) show PSNR improvements of $0.6$–$1.1$ dB over SOTA methods and demonstrate strong efficiency, suggesting degradation priors can generalize to other modalities and restoration tasks.

Abstract

Infrared images captured under turbulent conditions are degraded by complex geometric distortions and blur. We address infrared deturbulence as an image restoration task, proposing DparNet, a parameter-assisted multi-frame network with a wide & deep architecture. DparNet learns a degradation prior (key parameter matrix) directly from degraded images without external knowledge. Its wide & deep architecture uses these learned parameters to directly modulate restoration, achieving spatially and intensity adaptive results. Evaluated on dedicated infrared deturbulence (49,744 images) and visible image denoising (109,536 images) datasets, DparNet significantly outperforms State-of-the-Art (SOTA) methods in restoration performance and efficiency. Notably, leveraging these parameters improves PSNR by 0.6-1.1 dB with less than 2% increase in model parameters and computational complexity. Our work demonstrates that degraded images hide key degradation information that can be learned and utilized to boost adaptive image restoration.

Figures (5)

• Figure 1: Wide & deep learning boosts efficient image restoration. (a) is drawn according to relative PSNR and average running time to restore a 256 x 256 x 3 image for our method and SoTA restoration methods. Area of each circle in (a) is proportional to the number of model parameters. (b) shows the promotion of wide & deep learning on improving restoration performance.
• Figure 2: Overall structure of the proposed method. The learned degradation parameter matrix ( P) is integrated with degraded images ( I) via our DparNet in (a), which adopts a wide & deep architecture, to generate restored images ( Y).
• Figure 3: Visual comparison of image deturbulence.
• Figure 4: Stability comparison of image sequences.
• Figure 5: Visual comparison of image denoising.