RDANet: Relative Degradation Aware Network for Infrared Small Target Detection

arXiv:2608.20870 · cs.CV · Submitted 2026-08-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "RDANet: Relative Degradation Aware Network for Infrared Small Target Detection".

Jane: The RDANet proposes a new network architecture, consisting of two dedicated modules, to improve infrared small target detection by addressing relative degradation caused by target scale changes and scene variations.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we’re looking at this paper today called "RDANet: Relative Degradation Aware Network for Infrared Small Target Detection". It tackles a real problem where small targets in infrared pictures are hard to find because they are tiny and the background is super messy.

Jane: Right, so the main issue they point out is that existing detectors often get unstable when the target size changes or when you look at different backgrounds, which means their performance gets unbalanced across those different conditions.

Lu: Exactly. The paper says this scale- and scene-sensitive degradation shows that current methods aren't good enough because they can't keep the target structure stable while downsampling features and also maintain that sharp local contrast when the background shifts <ref:2608.20870#pg1>

Meng: So, what’s the fix? What does this RDANet actually propose to solve this instability with scale changes and background variations?

Tom: They propose a new network called RDANet which has two main modules, Multi-Scale Anti-Alias Downsampling or MSAD, and Prototype-Guided Skip Memory or PGSM <ref:2608.20870#pg1>

Jane: That’s the core idea. The abstract says MSAD introduces multi-scale anti-alias filtering along with pixel-fold aggregation to reduce aliasing effects during resolution reduction <ref:2608.20870#pg5>

Lu: And that filtering is specifically designed so that the target shape information can be better preserved while irrelevant background responses are suppressed <ref:2608.20870#pg5> They use three fixed binomial kernels with kernel sizes three, five, and seven as depthwise convolutions to get scale-specific responses <ref:2608.20870#pg6>

Meng: I need to know how that helps the actual detection. How does that pixel-unshuffle operator with stride two work for keeping those fine structures during resolution reduction?

Tom: It reorganizes each local two by two neighborhood into the channel dimension after anti-alias filtering, which keeps local sub-pixel information and avoids discarding fine target structures <ref:2608.20870#pg6> It’s about preserving that detail when you shrink the image size

Jane: That sounds complicated, so what about PGSM? What is Prototype-Guided Skip Memory doing to help with the scene variations part of this problem?

Lu: PGSM enhances those skip features by retrieving patch-level prototypes from a shared memory and adaptively integrating them into the current representation <ref:2608.20870#pg5> It's designed to maintain stable local contrast cues even when the scene is diverse

Paper summary: Meng: So, this shared memory bank stores representative feature patches from across different scenes, and PGSM uses that to guide the skip features? What’s the mechanism for retrieving those prototypes?

Lalam: PGSM first constructs a query representation by projecting the input skip feature map into a reduced channel space and unfolding it into a set of patch tokens Q <ref:2608.20870#pg5> Then, it computes the similarity between each query token qi and a memory token mj using a frequency-aware cosine measure based on log-transformed magnitude spectra extracted through cropped two dimensional Fourier transforms <ref:2608.20870#pg5>

Tom: So they’re not just looking for anything similar; they are using spectral information from the patches to find prototypes that match both semantics and spatial compatibility <ref:2608.20870#pg5> That sounds like a sophisticated way to pull in context

Jane: The experiment results show that RDANet achieves the best performance on most evaluation metrics across three public benchmarks <ref:2608.20870#pg5> And it specifically shows more stable behavior across different target sizes and scene complexity when you look at scale- and background-stratified evaluations <ref:2608.20870#pg5>

Meng: But if I’m trying to deploy this, I need to know how much computational overhead this adds compared to a baseline detector. Is the cost of running RDANet manageable for production?

Lu: The paper mentions that RDANet achieves the highest IoU on IRSTD-1k at seventy-three point eight two percent and NUDT-SIRST at ninety-five point four three percent <ref:2608.20870#pg5> On NUAASIRST, even though its IoU is slightly lower than MSDANet, RDANet reaches a prediction accuracy of one hundred percent and the lowest False Alarm rate of one point nine one <ref:2608.20870#pg5>

Tom: I’m looking at the scale-stratified results for IRSTD-1k, and RDANet ranks first in three of the four scale intervals and gets a total IoU of ninety-five point four three percent <ref:2608.20870#pg5> That suggests it handles different target sizes really well

Jane: And for background robustness, they look at local complexity in an annular neighborhood Omega n, where the complexity Cn is defined as Var(IOmega n) plus lambda Mean(nabla IOmega n) <ref:2608.20870#pg5> RDANet achieves the highest mean IoU of seventy-three point eight two percent and the smallest Background Gap of three point nine seven, ranking first in both the low- and high-complexity groups

Meng: So, it’s not just good on average; it’s good when things get really messy or when targets are small, which is where we usually fail <ref:2608.20870#pg5> That level of stability is what I need for deployment

Tom: The main contribution they highlight is analyzing that relative degradation in IRSTD and proposing RDANet to improve robustness under those conditions <ref:2608.20870#pg4>

Paper summary: Jane: So, to wrap up this part, the authors are claiming that by using MSAD and PGSM, you get better structural preservation during downsampling and more stable local contrast cues across different scenes <ref:two thousand six hundred eight point two zero eight seven zero#pg1 That means the detection performance is more consistent no matter what's happening in the scene

Lu: And they also point out that existing methods still have difficulty achieving stable performance across both target scales and background conditions <ref:2608.20870#pg3> This paper shows how to build a system that handles both issues simultaneously

Tom: So, we’ve seen the architecture, the modules MSAD and PGSM, and the fact that it outperforms others on scale stability and background robustness across those specific stratified tests

Jane: It really means for remote sensing applications where you're dealing with tiny targets in huge, unpredictable scenes that this approach gives a much more reliable result <ref:two thousand six hundred eight point two zero eight seven zero#pg5 The next thing we have to consider is how these ideas translate into actual deployment challenges and what the authors themselves admit are still limitations

Meng: What do they admit? Where does RDANet fall short in practice, especially when you’re looking at real-world data that doesn't perfectly match their training set?

Tom: They flag a few things. First, as shown in tabs two and five, it doesn't always get the best results in the smallest target interval <ref:2608.20870#pg5> Second, even though RDANet is better at handling scene variation and transferring to other datasets, its effectiveness still really depends on how much scale coverage and scene diversity is in their training data

Jane: And third, they note that while it has a good trade-off between accuracy and efficiency as shown in table nine, it’s not the fastest method available <ref:two thousand six hundred eight point two zero eight seven zero#pg5 So the system works well, but maybe it’s not the absolute quickest solution

Lu: Future work is mentioned exploring finer-grained scale-adaptive enhancement and more efficient memory interaction mechanisms for those tougher infrared scenes

Tom: So we've covered the RDANet paper today. It’s a network with MSAD for anti-aliasing and PGSM for memory guided feature retrieval that aims to stabilize detection across changing target scales and backgrounds

Jane: That’s the essence of it. It gives us a robust tool when we are looking at difficult infrared small targets in complex remote sensing images <ref:two thousand six hundred eight point two zero eight seven zero#pg5

Conclusion: Tom: So, to wrap up, RDANet is this network designed specifically to handle how infrared small target detection gets worse when the targets get smaller or when the background changes way around.

Jane: Exactly. The title itself tells us it’s all about that relative degradation—how performance degrades relative to a baseline as things change.

Lu: The authors, Rui Liu and his team, they are focusing on fixing two specific issues in existing detectors that cause this instability across different target scales and scene complexities.

Meng: From my side, the big question is what this actually means for deployment. It’s not just about getting a higher number on a test set; it’s about whether a model can handle real-world variability without constantly needing retraining.

Lalam: I see that the core contribution is proposing two modules, MSAD and PGSM, which together make the network robust to those specific types of degradation.

Tom: So, we're looking at MSAD for keeping target shape information intact during resolution reduction, and PGSM for stabilizing local contrast even when the scene background shifts.

Jane: That’s a really concrete way of putting it. It moves beyond just getting a better overall score on one benchmark to making the detector reliable across a whole range of conditions.

Lu: What this implies is that we can build infrared detection systems that are less sensitive to the exact scale or background noise they encounter in the field.

Meng: If it works as well as they claim, then for production, we might be able to use models that are more adaptable without needing a whole new pipeline every time the environment looks different.

Lalam: I'm confident this approach addresses those structural preservation and contrast modeling gaps directly.

Tom: It sounds like the authors are showing that by focusing on these two mechanisms—MSAD and PGSM—they can achieve a much more stable detection performance overall.

Jane: And while they show strong results on benchmarks, we need to keep an eye on their limitations, especially how much training data diversity is still required for that stability to hold up in the wild.

Lu: That’s fair. The authors themselves noted that effectiveness really depends on the scale coverage and scene diversity of the training data they use.

Meng: So, while it’s a strong architectural fix, we still have to worry about getting enough diverse real-world data into those modules for it to truly shine everywhere.

Lalam: I'm certain this paper provides a solid framework for tackling that specific type of instability in infrared imagery.

Tom: It sets up a clear direction for how we should think about making these small target detectors more resilient in complex remote sensing scenes.

Rui Liu, Jing Nie, Ying Fu

Beijing Institute of Technology

cs.CV

Submitted: 2026-08-21

Updated: 2026-08-21

Comments: Accept by TGRS 2026

DOI: 10.1109/TGRS.2026.3726938

Code: https://github.com/BIT-RuiLiu/RDANet

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: The RDANet proposes a new network architecture, consisting of two dedicated modules, to improve infrared small target detection by addressing relative degradation caused by target scale changes and

Key concepts

Multi-Scale Anti-Alias Downsampling (MSAD)
MSAD is a module that handles image resolution reduction. It uses multi-kernel low-pass filtering to suppress aliasing and background noise while simultaneously using pixel-fold aggregation. This combination helps preserve the shape of small targets during downscaling, ensuring their structure isn't lost.
Prototype-Guided Skip Memory (PGSM)
PGSM enhances skip features by retrieving representative feature patches, or prototypes, from a shared memory bank. These prototypes summarize local structures across different scenes. By adaptively fusing these retrieved prototypes into the current feature map, PGSM helps maintain stable local contrast cues even when the scene background changes.
Relative Degradation Aware Network (RDANet)
RDANet is the proposed architecture designed to handle relative degradation in infrared small target detection. It specifically tackles performance instability that occurs when targets change scale or backgrounds shift. The network's design focuses on structural preservation and contrast maintenance to ensure consistent detection accuracy.
Pixel-Fold Aggregation
This technique, used within MSAD, reorganizes fine-grained local structures during resolution reduction. Instead of simply discarding information, it reconfigures the data into a channel dimension. This process is crucial for better preserving the detailed shape information of small targets while effectively suppressing irrelevant background responses.

Summary

The RDANet proposes a new network architecture, consisting of two dedicated modules, to improve infrared small target detection by addressing relative degradation caused by target scale changes and scene variations. This work matters because existing detectors often show unstable performance when the target scale changes or when the scene background varies, which results in unbalanced detection performance across different conditions <ref:2608.20870#pg2>

How it works

The RDANet consists of two dedicated modules: Multi-Scale Anti-Alias Downsampling (MSAD) and Prototype-Guided Skip Memory (PGSM) <ref:2608.20870#pg5>. 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 <ref:2608.20870#pg5> 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 <ref:2608.20870#pg5>

Multi-Scale Anti-Alias Downsampling (MSAD)

MSAD performs resolution transformation through a multikernel low-pass filtering path together with a pixel-fold aggregation branch <ref:2608.20870#pg5>. The low-pass path suppresses aliasing effects and reduces background leakage during sampling, while the pixel-fold branch reorganizes fine-grained structures to better preserve target shape across scales <ref:2608.20870#pg5> Specifically, MSAD involves applying three fixed binomial kernels with kernel sizes 3, 5, and 7 as depthwise convolutions to obtain three scale-specific responses <ref:2608.20870#pg6>. A channelwise attention gate is then used to predict the fusion weights α through a softmax activation, and the filtered responses are aggregated as Y = α3 ⊙ Y3 + α5 ⊙ Y5 + α7 ⊙ Y7 <ref:2608.20870#pg6>. After anti-alias filtering, a pixel-unshuffle operator with stride 2 is applied to reorganize each local (2 × 2) neighborhood into the channel dimension <ref:2608.20870#pg6>. This operation retains local sub-pixel information during resolution reduction and avoids discarding fine-grained target structures <ref:2608.20870#pg6>.

Prototype-Guided Skip Memory (PGSM)

PGSM enhances skip features by retrieving patch-wise prototypes from a shared memory bank and adaptively fusing them with learnable weights <ref:2608.20870#pg5>. Given an input skip feature map F, PGSM first constructs a query representation for patch-wise retrieval by projecting F into a reduced channel space and unfolding it into a set of patch tokens Q. A global learnable memory bank M stores prototypical feature patches that summarize representative local structures observed across scenes. PGSM computes the similarity between a query token qi and a memory token mj using a frequency-aware cosine measure based on their log-transformed magnitude spectra extracted through cropped 2D FFT The aligned memory features are then aggregated through attention weighting, allowing PGSM to retrieve not only semantically relevant prototypes but also spatially compatible local patterns.

Experimental Results and Contributions

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 <ref:2608.20870#pg5> The main contributions are summarized as follows: 1) We analyze a relative degradation phenomenon in IRSTD, where the performance of existing detectors becomes unstable across target scales and background changes, and accordingly propose RDANet to improve robustness under such conditions <ref:2608.20870#pg4>. 2) We design MSAD, which combines multi-kernel antialias filtering with pixel-fold aggregation to preserve target structure and suppress background leakage during resolution transformation. 3) We introduce PGSM, which retrieves patch-wise prototypes from a shared memory and adaptively fuses them into skip features to maintain more stable local contrast cues across scenes.

Performance Analysis

RDANet achieves the highest IoU on IRSTD-1k (73.82%) and NUDT-SIRST (95.43%). On NUAASIRST, although its IoU is slightly lower than that of MSDANet, RDANet reaches Pd = 100% and the lowest Fa of 1.91. Scale-stratified results on IRSTD-1k show RDANet ranks first in three of the four scale intervals and achieves the highest overall IoU of 95.43%. For background robustness, local complexity is computed in an annular neighborhood omegan as Cn = Var(Iomegan) + λ Mean(∇Iomegan). RDANet achieves the highest mean IoU of 73.82% and the smallest BG-Gap of 3.97, while ranking first in both the low- and high-complexity groups.

Conclusion

In this paper, we revisit the problem of relative degradation in infrared small target detection, where detector performance becomes unstable as target scales expand or background conditions vary. We attribute this issue to insufficient structural preservation during resolution transformation and fragile local contrast modeling under scene variation. To address this issue, we propose RDANet, a simple encoder–decoder framework equipped with MSAD and PGSM. MSAD reduces aliasing and background leakage while preserving target morphology, and PGSM refines skip features to provide more stable local contrast cues across different scenes. Together, these two components improve feature robustness across both target scales and background conditions. Extensive experiments show that RDANet achieves competitive quantitative performance and more stable qualitative results across challenging scenes. Further analyses verify its advantages in scale stability, scene robustness, and cross-dataset generalization.

Limitations

Despite its strong overall performance, RDANet still has several limitations. First, as shown in Tabs. II and V, it does not always achieve the best results in the smallest target interval. Second, although RDANet improves robustness under scene variation and cross-dataset transfer, its effectiveness still depends on the scale coverage and scene diversity of the training data. Third, as indicated in Table IX, RDANet maintains a favorable accuracy–efficiency trade-off, but it is not the fastest method. Future work will explore finer-grained scale-adaptive enhancement and more efficient memory interaction mechanisms for more challenging infrared scenes.

References

[1] T. X. B. Nguyen, K. Rosser, and J. Chahl, “A review of modern thermal imaging sensor technology and applications for autonomous aerial navigation,” MDPI J. Imaging, vol. 7, no. 10, p. 217, 2021.

[3] J. Yao, S. Xiao, Q. Deng, G. Wen, H. Tao, and J. Du, “An infrared maritime small target detection algorithm based on semantic, detail, and edge multidimensional information fusion,” MDPI Remote Sens., vol. 15, no. 20, p. 4909.

[7] M. Zhang, R. Zhang, Y. Yang, H. Bai, J. Zhang, and J Guo “Isnet: Shape matters for infrared small target detection” in IEEE Conf Comput Vis Recog., 2022.

[11] B Li, C Xiao, L Wang, Y Wang, Z Lin, M Li, W An, and Y Guo “Dense nested attention network for infrared small target detection” IEEE Trans Image Process., vol. 32.

Improvements for AI systems

  1. Bold header: MSAD for Structure Preservation

The Multi-Scale Anti-Alias Downsampling (MSAD) module will be improved by explicitly using a pixel-fold aggregation branch in addition to the multi-kernel low-pass filtering path, as described in Section III.B. This ensures that target shape information can be better preserved while irrelevant background responses are suppressed, directly addressing the limitation where deep target semantics are not sufficiently preserved during resolution changes.

  1. Bold header: PGSM for Scene Robustness

The Prototype-Guided Skip Memory (PGSM) will be enhanced by refining the similarity measure in Equation 8, moving beyond simple frequency-aware cosine similarity to incorporate more sophisticated structural matching that is less sensitive to small positional deviations. This will improve the retrieval of spatially compatible local patterns and maintain a stable local contrast cue under scene changes.

  1. Bold header: Joint Module Effectiveness

The system will benefit from the complementary nature of MSAD and PGSM, as demonstrated by the ablation study showing that removing either MSAD or PGSM consistently degrades performance. This synergy allows the network to simultaneously achieve scale-consistent structure preservation and scene-robust local contrast modeling, leading to a final performance boost where both modules are present.

Related papers