DestripeCycleGAN: Stripe Simulation CycleGAN for Unsupervised Infrared Image Destriping

TL;DR

DestripeCycleGAN tackles infrared image destriping in an unsupervised setting by replacing the CycleGAN auxiliary generator with a Stripe Generation Model that encodes stripe priors. The approach couples a Haar wavelet-based background guidance module and a Multi-level Wavelet U-Net generator to preserve vertical edges and background details, while a gradient-based cycle-consistency enforces robust cross-domain alignment. Empirical results on synthetic and real infrared data show superior PSNR/SSIM and visual quality compared with state-of-the-art methods, with ablations validating the contributions of SGM, HBGM, and MWUNet. The framework offers a practical, data-efficient solution that can generalize across stripe patterns and real-world noise, with potential benefits for downstream infrared vision tasks.

Abstract

CycleGAN has been proven to be an advanced approach for unsupervised image restoration. This framework consists of two generators: a denoising one for inference and an auxiliary one for modeling noise to fulfill cycle-consistency constraints. However, when applied to the infrared destriping task, it becomes challenging for the vanilla auxiliary generator to consistently produce vertical noise under unsupervised constraints. This poses a threat to the effectiveness of the cycle-consistency loss, leading to stripe noise residual in the denoised image. To address the above issue, we present a novel framework for single-frame infrared image destriping, named DestripeCycleGAN. In this model, the conventional auxiliary generator is replaced with a priori stripe generation model (SGM) to introduce vertical stripe noise in the clean data, and the gradient map is employed to re-establish cycle-consistency. Meanwhile, a Haar wavelet background guidance module (HBGM) has been designed to minimize the divergence of background details between the different domains. To preserve vertical edges, a multi-level wavelet U-Net (MWUNet) is proposed as the denoising generator, which utilizes the Haar wavelet transform as the sampler to decline directional information loss. Moreover, it incorporates the group fusion block (GFB) into skip connections to fuse the multi-scale features and build the context of long-distance dependencies. Extensive experiments on real and synthetic data demonstrate that our DestripeCycleGAN surpasses the state-of-the-art methods in terms of visual quality and quantitative evaluation. Our code will be made public at https://github.com/0wuji/DestripeCycleGAN.

Figures (11)

• Figure 1: Overview of the proposed DestripeCycleGAN for infrared image destriping. Our DestripeCycleGAN mainly consists of three modules: stripe generation model (SGM), generator ($G_{\mathcal{S} \rightarrow \mathcal{C}}$), and Haar wavelet background guidance module (HBGM). SGM and $G_{\mathcal{S} \rightarrow \mathcal{C}}$ are employed to model the stripe noise and restore the clean image separately, jointly constructing the unsupervised architecture for $stripy \leftrightarrow clean$. HBGM is utilized to maintain background consistency between the stripy image and the destriped result. Moreover, the discriminator $D_{\mathcal{C}}$ is used to build adversarial loss with $G_{\mathcal{S} \rightarrow \mathcal{C}}$.
• Figure 2: Detailed architecture of the generator $G_{\mathcal{S} \rightarrow \mathcal{C}}$: (a) Multi-level Wavelet U-Net (MWUNet) using the Harr wavelet transform as the sampler to reduce information loss, and (b) group fusion blocks (GFB) that incorporate multi-scale features into the skip connection.
• Figure 3: PSNR of different methods on DLS_50, Set12 and CVC09_50 with Gaussian noise of $\sigma$ = 0.1.
• Figure 4: Comparison of destriping effects among different methods on the test image of DLS_50 he2018single.
• Figure 5: Comparison of destriping effects among different methods on the test image of Set12 zhang2017beyond.