PaperlayerView exact revision on arXiv

Source-linked AI summary

Seeing through Denoising: Frequency-guided One-step Denoising for Infrared Small Target Detection

Xuelin Qian, Zixuan Wang, Chang Xu, Dingwen Zhang, Junwei Han

arXiv:2610.29361v1cs.CV

TL;DR

IRSTD remains difficult because tiny, low-SNR targets are obscured by clutter, and generative modeling is underexplored. DiffIRSTD uses diffusion denoising to reconstruct target-free backgrounds, then provides them as detector context; the authors report effectiveness across multiple benchmarks and improvements over an auto-encoder across three datasets.

  • Problem

    Tiny, low-SNR infrared targets are often obscured by clutter, while most existing methods directly map images to masks and generative modeling remains underexplored.

    Source: §I · p. 1

  • Method

    DiffIRSTD treats targets as noise-like signals, uses diffusion to reconstruct a target-free background, and supplies that background as context to a detector.

    Source: §I · p. 1, A. Overview · p. 3

  • Results

    The authors report effectiveness and robustness across multiple benchmarks; the one-step diffusion model outperforms the auto-encoder across three datasets, with IRSTD-1K IoU improving from 67.01 to 72.13.

    Source: §VI · p. 11, C. Ablation Study · p. 11

  • Takeaways & Limitations

    The reconstruction component is modular and can be integrated with different detection architectures as contextual guidance.

    Source: §VI · p. 11

    The reconstruction network still has many parameters, so further lightweight optimization is needed for deployment of the full framework on edge devices.

    Source: B. Limitations and Future Work · p. 11

Abstract

from arXiv · show

Infrared small target detection (IRSTD) is a challenging task due to the extremely small size of targets, low signal-to-noise ratios, and heavy background clutter in infrared imagery. While the progress of deep learning has advanced this field, existing methods primarily adopt discriminative paradigms that directly map images to target masks, overlooking the potential of generative modeling. In this paper, we propose DiffIRSTD, a novel generative framework that reformulates infrared small target detection as a background reconstruction problem. Instead of explicitly predicting target regions, we reinterpret small targets as noise-like signals and leverage a diffusion-based model to recover a clean, target-free background from a perturbed input image. The residual between the reconstructed background and the original image is then exploited to highlight candidate target regions, providing explicit cues for subsequent detection. To address the computational cost typically associated with diffusion sampling, we introduce a one-step denoising strategy that enables efficient background reconstruction while maintaining practical inference speed. Furthermore, we present a frequency-guided perception loss to preserve high-frequency background details and focus more on the reconstruction quality of target regions. More importantly, our DiffIRSTD operates as a plug-and-play framework, compatible with existing IRSTD models without increasing the computational complexity of detectors. Extensive experiments on IRSTD-1K, NUAA-SIRST, and NUDT-SIRST datasets demonstrate the superiority of our method, achieving state-of-the-art performance across three benchmarks.

I. INTRODUCTION

DiffIRSTD recasts infrared small target detection as target-free background reconstruction, using denoising to expose target cues for detection. It combines one-step diffusion with frequency-guided supervision and is reported to outperform state-of-the-art approaches on three benchmarks.

I. INTRODUCTION

  • Targets can be only a few pixels, dim and point-like, and obscured by low signal-to-noise ratios and clutter.These conditions make robust and accurate detection challenging.Source: §I · p. 1
  • Most existing methods directly map images to target masks, while generative modeling remains underexplored in IRSTD.Earlier diffusion-based approaches generate training samples or iteratively refine masks, with the latter adding computational overhead that may limit deployment in resource-constrained settings.Source: §I · p. 1
  • DiffIRSTD treats targets as noise-like signals and denoises a perturbed image to reconstruct a target-free background whose difference from the input highlights candidate regions.The paper argues that reconstructing the background is expected to be easier than directly generating target masks.Source: §I · pp. 1–2, Figure 1 · p. 2
  • One-step denoising reduces sampling steps, while frequency-guided perception loss preserves high-frequency background details and emphasizes reconstruction quality in target regions.The high-frequency constraint is intended to retain background edges and textures.Source: §I · p. 2
  • The reconstructed background is concatenated with the original image as a dual-channel input, and the framework can plug into existing detectors without increasing their computational complexity.The background provides an explicit prior for localization.Source: §I · p. 2
  • Experiments on three IRSTD benchmarks report that DiffIRSTD outperforms state-of-the-art approaches.The supplied introduction passage ends before completing its description of the detection performance.Source: §I · p. 2

II. RELATED WORK

Prior IRSTD work includes discriminative architectures and background-target decoupling, while diffusion methods have been used for data generation and detection. DiffIRSTD instead uses generative denoising to produce background context for existing detectors.

A. Infrared Small Target Detection

  • The related-work survey also describes CNN, transformer, Mamba, and SAM-based methods that refine feature representation or mask reconstruction.Examples include nested multi-scale designs, edge-focused modules, transformer blocks, and SAM adaptation.Source: A. Infrared Small Target Detection · p. 2
  • Low-rank decomposition methods model infrared images through background-target decoupling and detect targets using residual extraction.The image is modeled with background, target, and noise components; methods estimate the background before extracting the residual.Source: A. Infrared Small Target Detection · p. 2
  • Unlike LRRNet’s low-rank decomposition and residual extraction, DiffIRSTD uses diffusion to generate target-free background context rather than directly suppressing background.The generated background is an auxiliary cue for existing detectors, not itself the detection result.Source: A. Infrared Small Target Detection · pp. 2–3

B. Diffusion-based Object Detection Models

  • Diffusion-based IRSTD methods have pursued data augmentation and detection or feature modeling, including iterative mask generation.Examples include diffusion-based image resampling and generation, continuous feature representations, and iterative target-mask output.Source: B. Diffusion-based Object Detection… · p. 3
  • DiffIRSTD differs by treating targets as noise-like signals and removing them to generate target-free backgrounds, with reconstruction and detection networks separated for adaptation to different detectors.This differs from approaches that use diffusion to generate images or masks.Source: B. Diffusion-based Object Detection… · p. 3

III. METHODOLOGY

DiffIRSTD assumes small targets behave as noise-like perturbations and learns target-free backgrounds conditioned on observed images. The generated background serves as contextual input to a target detector, alongside the original image.

A. Overview

  • DiffIRSTD assumes targets are noise-like perturbations and uses diffusion denoising to learn the conditional distribution of target-free backgrounds.The paper contrasts this with approaches that generate target masks or use diffusion itself to localize targets.Source: A. Overview · p. 3
  • The generated clean background is conditioned on the observed image and supplied to the detector as contextual information rather than used for residual computation.This distinguishes the framework’s background use from residual-based approaches discussed in the passage.Source: A. Overview · p. 3
  • The methodology section covers diffusion processes, image denoising with one-step sampling and frequency-guided loss, detector use of generated backgrounds, pseudo-label generation, and training and inference workflows.Source: A. Overview · p. 3

B. Preliminary

DDPMs model generation through forward noise addition and reverse denoising. The forward process uses a predefined noise schedule, while a neural network learns to predict the added noise.

  • DDPMs define data generation with forward diffusion and reverse denoising Markov chains.Source: B. Preliminary · pp. 3–4
  • The forward process incrementally injects Gaussian noise, with βt controlling the noise level at each step.Source: B. Preliminary · p. 4
  • A neural network ϵθ(xt, t) predicts the added noise, and the diffusion training objective minimizes its mean squared error against the ground-truth noise.Source: B. Preliminary · p. 4

C. Reconstruction Network

The reconstruction network treats small targets as noise-like deviations and uses diffusion denoising to estimate a target-free background. One-step denoising reduces inference cost, while frequency-guided supervision helps preserve background detail.

  • Small targets are treated as noise-like signals against a smooth background, making background recovery analogous to diffusion denoising.Source: C. Reconstruction Network · p. 4
  • The diffusion process perturbs the input and denoises it to produce a reconstructed image approximating the scene without targets.Source: C. Reconstruction Network · p. 4
  • One-step denoising skips iterative transitions to accelerate inference, but the noise level must balance target removal against background fidelity.Too large a noise level can degrade image quality, while too small a level can leave targets difficult to remove in one step.Source: C. Reconstruction Network · p. 5
  • The frequency-guided loss compares high-frequency representations of reconstructed and ground-truth backgrounds after frequency transformation and high-pass masking.The mask retains components beyond a cutoff radius, and the loss is a mean squared error between masked spectra.Source: C. Reconstruction Network · p. 5
  • The frequency-guided loss emphasizes target-related high-frequency discrepancies while preserving background structure and texture during one-step denoising.Source: C. Reconstruction Network · p. 5

D. Detection Network

The detector receives the original image alongside its reconstructed target-free background, learning their discrepancy to localize targets. The framework can use different detector architectures without modifying its reconstruction network.

  • The detector uses the original image and reconstructed background as a dual-channel input, learning their discrepancy to highlight target regions.Source: D. Detection Network · p. 5
  • The baseline detector is a U-Net with skip connections, and channel attention rescales feature maps using pooled contextual information.The decoder outputs the target mask after encoder features are progressively down-sampled and spatial resolution is restored.Source: D. Detection Network · pp. 5–6
  • The reconstruction component is plug-and-play across detector architectures, and the loosely connected networks can be deployed separately or paired with lighter or more powerful detectors.Source: D. Detection Network · p. 6

E. Background Acquisition Strategy

Because infrared datasets provide target masks but not ground-truth backgrounds, the method synthesizes pseudo-label backgrounds by masking targets and inpainting the removed regions.

F. Training and Inference

Training jointly supervises background reconstruction and target-mask detection. At inference, the image and its reconstructed target-free background are fed to the detector to predict the mask.

F. Training and Inference

  • The reconstruction network is trained with pixel-level reconstruction and frequency-guided perceptual losses to generate high-quality backgrounds.The high-frequency loss guides denoising toward target-free backgrounds, while λ controls its importance.Source: F. Training and Inference · p. 6
  • The detection network learns target localization and segmentation under SoftIoU and BCE losses.SoftIoU optimizes intersection over union, while BCE supports pixel-level classification accuracy.Source: F. Training and Inference · p. 6
  • At inference, the input image and its reconstructed target-free image are fed together to the detection network to predict a target mask.Source: F. Training and Inference · p. 6

IV. EXPERIMENTS

Experiments evaluate DiffIRSTD on IRSTD-1K, NUDT-SIRST, and NUAA-SIRST using dataset-specific train/test splits. DiffIRSTD achieves the best reported IoU on all three datasets, while visual comparisons show fewer false alarms and missed detections than comparison methods.

A. Experimental Setups

B. Comparisons with the State-of-the-arts

  • 72.13 on IRSTD-1K, 95.91 on NUDT-SIRST, and 80.73 on NUAA-SIRST are DiffIRSTD’s IoU scores, the best on all three datasets.The method also achieves competitive false-alarm performance, particularly on NUDT-SIRST, while balancing localization accuracy and false positives.Source: B. Comparisons with the… · p. 7, Table I · p. 7
  • The upper-bound setting substantially outperforms existing methods on IRSTD-1K and NUDT-SIRST, while the gap to DiffIRSTD leaves room for better background generation.It bypasses the generation module and trains the detector directly with pseudo-label background images.Source: B. Comparisons with the… · pp. 7–8
  • Across five visualized samples, DiffIRSTD suppresses false alarms and missed detections seen in comparison methods, with stronger ground-truth alignment in the first two rows.The figure marks correct detections in green, false alarms in red, and missed targets in orange.Source: B. Comparisons with the… · p. 8, Figure 4 · p. 8

C. Ablation Study

The ablations examine reconstruction choices, diffusion steps, background inputs, frequency-guided loss settings, and detector compatibility. They report gains from one-step diffusion and show how design choices affect detection and background reconstruction.

Efficacy of Designs in Reconstruction Network

  • Adding the reconstruction network raises IRSTD-1K IoU from 67.13 to 69.99.The baseline activates only the U-Net detection network.Source: C. Ablation Study · p. 8

Reconstruction Loss and Diffusion Steps

  • Image supervision improves detection over standard noise supervision, while t=50 or 70 tends to perform better within the 30–90-step range.The experiments use t=50; increasing t blurs background details without substantially improving target removal.Source: Table III · p. 9, C. Ablation Study · p. 9, Table IV · p. 9

Designs of Dual-channel Inputs

  • Concat(x, xB) achieves the highest IoU, 72.13, among two-channel configurations; high-pass and explicit-difference inputs perform worse.The authors suggest explicit difference maps may amplify noise or reconstruction errors.Source: Table V · p. 9, C. Ablation Study · p. 9

Coefficients in Frequency-guided Perceptual Loss

  • λ=0.1 achieves the highest IoU (72.13) and lowest Fa; too little or too much weighting can impair detail preservation or increase false alarms.The reported performance fluctuations across λ are not significant.Source: Table VI · p. 9, C. Ablation Study · p. 10

Choice of High-pass Cutoff Frequency rc

  • rc=0.7 gives the best reported balance between retaining high-frequency details and maintaining background structure.The authors report that smaller or larger cutoff settings can degrade detection performance.Source: Table VII · p. 10, C. Ablation Study · p. 10

Discussion on Reconstruction Methods

Generalization to Other Detectors

A. Broader Impact

DiffIRSTD frames infrared small-target detection as background reconstruction followed by target detection, and the authors discuss possible extensions and deployment constraints. They report benchmark effectiveness and detector adaptability while noting that edge deployment requires further lightweight optimization.

A. Broader Impact

  • The proposed “background reconstruction + target detection” decomposition could improve reliability in maritime search and rescue and missile early warning, with lower false alarms and higher detection accuracy.The authors also suggest that the target-erasure idea could extend to remote-sensing tasks such as cloud removal and occluder elimination.Source: A. Broader Impact · p. 11

B. Limitations and Future Work

  • The reconstruction network still has many parameters despite single-step sampling; deploying the full framework on edge devices requires further lightweight optimization.The authors suggest running reconstruction independently in the cloud as a preprocessing step.Source: B. Limitations and Future Work · p. 11

VI. CONCLUSION

  • Experiments on multiple benchmarks support the framework’s effectiveness and robustness, while its modular reconstruction component can integrate with different detection architectures.The conclusion presents this adaptability as evidence of practical value.Source: §VI · p. 11

Related papers

Loading 2610.29361v1…