PaperlayerView exact revision on arXiv

Source-linked AI summary

RDANet: Relative Degradation Aware Network for Infrared Small Target Detection

Rui Liu, Jing Nie, Ying Fu

arXiv:2608.20870v1cs.CV

TL;DR

Infrared small-target detectors can degrade across target scales and changing backgrounds, motivating RDANet's joint focus on structure preservation and local-contrast stability. Its MSAD and PGSM modules address these issues, and experiments report strong benchmark performance with more stable behavior across scales and scenes.

  • Problem

    Existing detectors show limited stability across target scales and scene conditions, with fragile local contrast and insufficient structure preservation during downsampling.

  • Method

    RDANet combines multi-scale anti-alias downsampling with prototype-guided skip memory to preserve target structure and stabilize local contrast.

  • Results

    RDANet improves IoU to 73.82% while reducing Fa to 7.67 in the complete PGSM comparison.

  • Takeaways & Limitations

    RDANet provides strong benchmark performance with improved generalization to unseen datasets, target-scale distributions, and false-alarm constraints.

  • Takeaways & Limitations

    RDANet does not always achieve the best results in the smallest target interval and depends on training-data scale coverage and scene diversity.

Abstract

from arXiv · show

Infrared small target detection is still challenging in remote sensing imagery, because the targets are extremely small, exhibit weak local contrast, and are often embedded in complex and highly variable backgrounds. In addition to these inherent difficulties, we observe that existing detectors often show unstable performance when the target scale changes or when the scene background varies. This scale- and scene-sensitive degradation indicates that current methods are insufficient in simultaneously preserving target structure during feature downsampling and maintaining discriminative local contrast under background shifts, which finally results in unbalanced detection performance across different conditions. To improve detection robustness, this paper proposes a Relative Degradation Aware Network (RDANet) for infrared small target detection. RDANet consists of two dedicated modules: Multi-Scale Anti-Alias Downsampling (MSAD) and Prototype-Guided Skip Memory (PGSM). MSAD introduces multi-scale anti-alias filtering together with pixel-fold aggregation to reduce aliasing effects during resolution reduction, so that target shape information can be better preserved while irrelevant background responses are suppressed. PGSM further enhances the skip features by retrieving patch-level prototypes from a shared memory and adaptively integrating them into the current representation, which helps maintain stable local contrast cues under diverse scene backgrounds. Experiments on three public benchmarks show that RDANet achieves the best performance on most evaluation metrics, while scale- and background-stratified evaluations indicate more stable behavior across target sizes and scene complexity. The code is available at https://github.com/BIT-RuiLiu/RDANet.

I. INTRODUCTION

Existing infrared small-target detectors show unstable performance across target scales and background changes. RDANet addresses these limitations with anti-aliased structure-preserving downsampling and prototype-guided skip-feature refinement.

  • Motivation: Deep infrared small-target detectors remain limited in maintaining balanced performance across different target sizes.The issue is especially relevant because targets occupy few pixels and often have low signal-to-clutter ratios.
  • Scale and scene degradation: Existing methods can generalize poorly from scale-imbalanced training data to relatively larger targets.IRSTD-1k contains an evident imbalance across target-size intervals, with few large-target samples.
  • Scale and scene degradation: Changing the surrounding background causes pronounced response fluctuations for enlarged targets, while smaller targets produce more stable responses.This links degradation to both target scale and background-induced feature variation.
  • Proposed approach: RDANet combines MSAD and PGSM to improve scale-consistent structure preservation and scene-robust local contrast modeling.MSAD uses multi-kernel low-pass filtering and pixel-fold aggregation, while PGSM retrieves and fuses patch-wise prototypes from shared memory.
  • Contributions: MSAD preserves target structure and suppresses background leakage during resolution transformation.Its design combines multi-kernel anti-alias filtering with pixel-fold aggregation.
  • Contributions: PGSM retrieves patch-wise prototypes from shared memory and adaptively fuses them into skip features to stabilize local contrast across scenes.The contribution targets scene-dependent variation in target-background cues.

II. RELATED WORK

Related work has pursued scale-aware modeling, scene robustness, improved downsampling, and memory-based feature enhancement. However, prototype retrieval within skip connections and its coordination with downsampling remain insufficiently explored for joint scale and background robustness.

  • Existing IRSTD methods: Traditional IRSTD methods use filtering, local-contrast, and low-rank decomposition priors, while deep learning has become dominant.Traditional approaches are lightweight and interpretable but can degrade under low contrast, clutter, and scene variation.
  • Scale robustness: Recent methods improve scale-aware representations through multi-scale heads, dynamic losses, attention, context modeling, and scale-aligned fusion.These approaches target variation in target size through different representation and fusion strategies.
  • Scene robustness: Scene-robustness studies improve adaptability through confidence regulation, cross-view alignment, scene-conditioned parameters, low-rank modeling, pruning, and feature compensation.The cited approaches address complex scenes, cross-scene generalization, noise, clutter, or efficiency.
  • Downsampling and memory: Alternative downsampling designs seek to reduce aliasing and preserve fine-grained spatial details before convolution.SPDConv is described as rearranging local information into channels to retain spatial detail.
  • Downsampling and memory: Memory mechanisms store and retrieve informative patterns for feature recovery, temporal consistency, and prototype learning in IRSTD.Existing examples include memory-augmented recovery and frequency-aware enhancement for video sequences.
  • Research gap: Prototype retrieval within skip connections and its coordination with downsampling remain largely unexplored.This leaves scale imbalance and background-induced contrast shifts insufficiently addressed.

A. Overview of RDANet

RDANet is a lightweight encoder–decoder framework designed to stabilize representations across target scales and scene conditions. It replaces encoder stride-two downsampling with MSAD and refines skip connections with PGSM before decoder fusion.

  • Motivation: Existing detectors remain unstable when target scales expand beyond the dominant training distribution or backgrounds change.The paper treats these effects as a broader limitation across IRSTD models.
  • Design rationale: Repeated encoder downsampling can introduce aliasing, blur target morphology, and leak background interference into deeper features.These effects become more evident when target scales vary.
  • Architecture: RDANet replaces each stride-two encoder downsampling operation with MSAD and refines each skip connection with PGSM before fusion.MSAD targets morphology preservation and leakage suppression, while PGSM retrieves representative patch-wise prototypes for scene-invariant local contrast.

B. Multi-Scale Anti-Alias Downsampling

MSAD performs multi-scale anti-alias filtering, pixel-unshuffle aggregation, lightweight refinement, residual compensation, and channel recalibration to preserve target morphology during downsampling while suppressing background leakage.

  • Motivation: Conventional strided convolutions and pooling can introduce aliasing, blur target boundaries, distort morphology, and amplify background responses.These effects are particularly harmful when infrared target cues are weak.
  • Multi-scale filtering: MSAD applies three fixed binomial low-pass kernels with sizes 3, 5, and 7 to obtain scale-specific responses.The different smoothing ranges address target and background structures at different spatial extents.
  • Multi-scale filtering: Adaptive channel-wise fusion selects smoothing weights so smaller kernels preserve weak target details while larger kernels suppress broader high-frequency clutter.The fusion balances responses instead of applying one fixed smoothing strength to all features.
  • Pixel-fold aggregation: Pixel-unshuffle reorganizes each local 2 × 2 neighborhood into channels, retaining sub-pixel information during resolution reduction.A grouped convolution then averages the four folded sub-channels as a parameter-free aggregation strategy.
  • Feature refinement: A depthwise separable convolution provides local feature refinement with low computational overhead.The operation consists of depthwise 3 × 3 convolution followed by pointwise 1 × 1 convolution.
  • Residual compensation: The residual shortcut supplies a stable low-frequency reference and helps preserve feature distributions during resolution change.It directly downsamples the input through average pooling before combining with the refined main branch.
  • Channel recalibration: Channel recalibration with an SE block emphasizes target-relevant channels and suppresses residual background activations.MSAD therefore integrates filtering, sub-pixel aggregation, residual compensation, and channel recalibration in one lightweight module.

C. Prototype-Guided Skip Memory

Prototype-Guided Skip Memory (PGSM) retrieves patch-level prototypes from a learnable memory and integrates aligned structural references into skip features. This design stabilizes local contrast and preserves target-related details under background changes.

  • Prototype retrieval: PGSM tokenizes projected skip features into local patch queries, enabling retrieval of target and background structures at fine granularity.The query tokens are formed after 1×1 projection and patch unfolding, rather than from global feature statistics.
  • Prototype retrieval: A global learnable memory bank stores prototypical feature patches that summarize representative local structures across scenes.Query-memory similarity uses frequency-aware representations derived from Hann-windowed, low-frequency-cropped log-magnitude spectra.
  • Similarity and alignment: Frequency-aware cosine matching emphasizes structural magnitude patterns and reduces sensitivity to small positional deviations and background texture changes.The similarity uses an inner product normalized by the ℓ2 norms of the spectral representations.
  • Similarity and alignment: Top-K prototypes are aligned with phase-correlation displacement estimates and aggregated using attention weights before fusion.Bilinear warping compensates residual offsets so retrieved patterns remain spatially compatible with the query patch.
  • Feature fusion: The retrieved prototypes act as structural references while the original skip features remain the primary source of target detail.Aggregated patches are folded into a full-resolution map, concatenated with the skip feature, and fused by 1×1 convolution, BatchNorm, and ReLU.
  • Feature fusion: PGSM provides lightweight memory-enhanced skip refinement that improves local-contrast stability and supplies more reliable fine-grained decoder features.The fusion enriches skip features with context-aware prototype cues while suppressing scene-specific background interference.

A. Experimental Details

RDANet is evaluated on three public IRSTD benchmarks using pixel-level and object-level metrics, with scale and background robustness assessed through stratified analyses. The experiments use standardized image sizes, dataset splits, and training settings.

  • Datasets: The evaluation covers IRSTD-1k, NUDT-SIRST, and NUAA-SIRST, containing 1,001, 1,327, and 427 images respectively.All images are resized to 512×512; the dataset split ratios are 4:1, 1:1, and 4:1 respectively.
  • Evaluation metrics: IoU measures predicted-mask overlap, while Pd and Fa quantify target detection probability and false-alarm rate.The study combines pixel-level and object-level evaluation.
  • Robustness analysis: Scale robustness is evaluated by grouping IRSTD-1k targets into mask-area bins.The supplied passage lists bins beginning with,,, [64, 127], and [128, 255].
  • Robustness analysis: Background robustness is measured by local complexity and BG-Gap across low-, medium-, and high-complexity groups.Groups are divided at the 33.33th and 66.67th percentiles, with BG-Gap defined as maxg IoUg − ming IoUg.
  • Implementation: RDANet is trained for 1,000 epochs with AdamW, cosine annealing, an initial learning rate of 1 × 10^-3, and batch size 4.Implementation uses PyTorch on a single NVIDIA GeForce RTX 3090 GPU.

B. Comparison with the State-of-the-Art Methods

RDANet is compared with traditional and deep-learning-based IRSTD methods using overall, scale-stratified, background-stratified, and visual evaluations. It achieves strong benchmark performance and more stable behavior across target scales and background complexity.

  • Quantitative results: RDANet achieves the highest IoU on IRSTD-1k at 73.82% and on NUDT-SIRST at 95.43%.On NUAA-SIRST, it reaches Pd = 100% and the lowest Fa of 1.91 despite slightly lower IoU than MSDANet.
  • Scale robustness: On NUDT-SIRST, RDANet ranks first in three of four scale intervals and achieves overall IoU of 95.43%.On IRSTD-1k, it obtains the highest SB-IoU of 71.60% and best average rank of 1.67.
  • Visual results: Visual comparisons use red boxes for correct detections, blue boxes for missed targets, yellow boxes for false alarms, and corner close-ups.The comparisons include traditional and deep-learning-based methods on challenging cases.
  • Background robustness: RDANet achieves mean IoU of 73.82% and BG-Gap of 3.97 on IRSTD-1k, ranking first in both low- and high-complexity groups.These background-stratified results indicate more stable performance under varying background complexity.
  • Visual results: RDANet detects both small targets on a textured stone surface and produces the most accurate, complete target shape in another challenging case.The qualitative results also report stable responses when the same target is composited onto different backgrounds.

C. Discussion

The discussion attributes RDANet’s gains to complementary MSAD and PGSM components that improve structural preservation, local contrast, and robustness across scales and scenes. Ablations and robustness analyses support strong performance with modest overhead, while limitations remain for the smallest targets, underrepresented conditions, and inference efficiency.

  • Overall Effectiveness of MSAD and PGSM: Removing either MSAD or PGSM degrades performance, while removing both returns results to the baseline, confirming complementary joint benefits.MSAD mainly preserves structure during downsampling, whereas PGSM improves target–background separation under scene variation.
  • Design Choices: 73.82% IoU and 7.67 Fa are achieved by complete PGSM, improving over frequency retrieval without alignment at 72.60% IoU while reducing false alarms.The comparison supports the joint contribution of frequency-aware retrieval and phase-correlation alignment.
  • Mechanism Analysis of MSAD and PGSM: MSAD suppresses peripheral high-frequency components and preserves dominant low-frequency structure, producing clearer, more localized target responses across encoder stages.These visualizations are consistent with reduced aliasing and background leakage during resolution transformation.
  • Mechanism Analysis of MSAD and PGSM: PGSM makes skip features more target-focused and structurally complete, providing clearer boundaries and more stable local contrast cues under complex scenes.The refinement reduces ambiguous background responses compared with directly propagating shallow encoder features.
  • Generalization and Robustness Analysis: RDANet achieves the best performance in both cross-dataset transfer settings without fine-tuning and improves IoU and Pd while maintaining low Fa.The settings transfer models between IRSTD-1k and NUAA-SIRST, indicating stronger generalization to unseen scenes and background statistics.
  • Generalization and Robustness Analysis: After augmented training, RDANet remains competitive across most scale intervals and achieves the best overall balance among IoU, detection probability, and false-alarm rate.Its scale robustness is attributed not only to data rebalancing but also to more stable feature representations across target-size distributions.
  • Limitations: RDANet does not always lead in the smallest target interval, depends on training-data scale and scene coverage, and is not the fastest method.Anti-alias filtering may weaken extremely subtle responses, prototypes may be insufficiently representative in underrepresented conditions, and memory-guided refinement adds computation.

V. CONCLUSION

The conclusion frames relative degradation as instability caused by insufficient structural preservation during resolution changes and fragile local contrast modeling under scene variation. RDANet addresses these issues with MSAD and PGSM, yielding more robust representations across target scales and background conditions.

  • V. CONCLUSION: RDANet targets detector instability caused by changing target scales and background conditions in infrared small target detection.The paper attributes this relative degradation to insufficient structural preservation and fragile local contrast modeling.
  • V. CONCLUSION: MSAD reduces aliasing and background leakage while preserving target morphology, whereas PGSM stabilizes local contrast cues across scenes.Together, the components improve feature robustness across both target scales and background conditions.
  • V. CONCLUSION: Experiments report competitive quantitative performance, more stable qualitative results, and advantages in scale stability, scene robustness, and cross-dataset generalization.The conclusion identifies finer-grained scale-adaptive enhancement and more efficient memory interaction as future directions.
Loading 2608.20870v1…