ASCNet: Asymmetric Sampling Correction Network for Infrared Image Destriping

TL;DR

Infrared destriping is challenged by cross-level column semantic gaps and insufficient global column modeling. ASCNet addresses this by combining an asymmetric sampling strategy with a Residual Haar Discrete Wavelet Transform downsampler and a Column Non-uniformity Correction Module, enabling robust global column reasoning within a U-shaped network; degraded image formation is modeled as $I_D = I_N + I_C$. The approach comprises three core components: RHDWT, Pixel Shuffle upsampling, and CNCM with RCSSC, and is validated on synthetic and real infrared data, showing PSNR/SSIM gains and improved real-stripe perception. The work also demonstrates task-level gains in infrared small target detection, suggesting practical impact for IR surveillance and related applications, especially under diverse stripe distributions.

Abstract

In a real-world infrared imaging system, effectively learning a consistent stripe noise removal model is essential. Most existing destriping methods cannot precisely reconstruct images due to cross-level semantic gaps and insufficient characterization of the global column features. To tackle this problem, we propose a novel infrared image destriping method, called Asymmetric Sampling Correction Network (ASCNet), that can effectively capture global column relationships and embed them into a U-shaped framework, providing comprehensive discriminative representation and seamless semantic connectivity. Our ASCNet consists of three core elements: Residual Haar Discrete Wavelet Transform (RHDWT), Pixel Shuffle (PS), and Column Non-uniformity Correction Module (CNCM). Specifically, RHDWT is a novel downsampler that employs double-branch modeling to effectively integrate stripe-directional prior knowledge and data-driven semantic interaction to enrich the feature representation. Observing the semantic patterns crosstalk of stripe noise, PS is introduced as an upsampler to prevent excessive apriori decoding and performing semantic-bias-free image reconstruction. After each sampling, CNCM captures the column relationships in long-range dependencies. By incorporating column, spatial, and self-dependence information, CNCM well establishes a global context to distinguish stripes from the scene's vertical structures. Extensive experiments on synthetic data, real data, and infrared small target detection tasks demonstrate that the proposed method outperforms state-of-the-art single-image destriping methods both visually and quantitatively. Our code will be made publicly available at https://github.com/xdFai/ASCNet.

Figures (15)

• Figure 1: (a) Wavelet U-shaped neural network. (b) DWT decomposition and reconstruction of the stripy image $\mathbf{P}$. (c) DWT decomposition and reconstruction of the stripy feature $\mathbf{S}$ in wavelet U-shaped neural network. (d) Visualization of each channel of the crosstalk feature $\mathbf{K}_{c}$.
• Figure 2: Insightful feature visualizations between (a) conventional Symmetric Sampling: DWT/IDWT, and (b) Asymmetric Sampling: DWT/PS during the last two stages of upsampling. The Column-wise Mean Responding Curves (CMRC) of feature maps (maximum the channel-wise response) are provided to showcase the column semantic changes introduced by upsampling. The cross-level column semantic gap in symmetric sampling is highlighted in red boxes. We can observe that asymmetric sampling has stable semantic fluctuations and clearer pointers for noise reconstruction.
• Figure 3: Overview of the proposed Asymmetric Sampling Correction Network (ASCNet). Three core modules of ASCNet are: (a) Pixel Shuffle (PS) to achieve semantic-bias-free image reconstruction, (b) Residual Haar Discrete Wavelet Transform (RHDWT) that enriches structural and semantic feature representation, and (c) Column Non-uniformity Correction Module (CNCM) nests the Residual Column Spatial Self-Correction (RCSSC) block to enhance column characteristics, spatial information, and long-range dependencies.
• Figure 4: Architecture of the Residual Column Spatial Self-Correction (RCSSC) block. (a) RCSSC incorporates residual connection into the Column Spatial Self-Correction (CSSC) block. The core modules of CSSC are: (b) Spatial Attention Branch (SAB) that utilizes spatial correlation to enhance structural characterization of key areas, (c) Column Attention Branch (CAB) that strengthens column characteristics to eliminate the stripe noise column differences and (d) Self-Calibrated Branch (SCB) that build remote dependencies to fine-tune the global uniformity.
• Figure 5: Image destriping results of different methods on ICSRN with Gaussian-distributed simulated noises. Yellow arrows and circles represent the stripe noise residue and blurred image details.