PaperlayerView exact revision on arXiv

Source-linked AI summary

Traditional and Recent Approaches for Infrared Small Target Detection : Overview, Challenges and Perspectives (Part 1)

Thierry Bouwmans, Anastasia Zakharova, Belmar Garcia, Marwa Chendeb El Rai, Meghna Kapoor, Badri Narayan Subudhi

arXiv:2610.18276v1cs.CV

TL;DR

Infrared small target detection remains difficult because targets are dim and structurally weak, and existing surveys do not provide an overall review. This paper surveys decades of methods, challenges, datasets, and evaluation approaches; its conclusion organizes the field’s traditional and low-rank methods and presents theoretical perspectives and AI-assisted research solutions.

  • Problem

    Dim, few-pixel infrared targets are difficult to detect amid clutter, and existing surveys leave the field without an overall review.

    Source: §1 · p. 1, §1.2 · p. 3

  • Method

    The paper reviews 30 years of infrared small target detection, categorizing traditional and low-rank plus sparse decomposition methods and covering datasets, metrics, and ranking protocols.

    Source: §11 · p. 43

  • Results

    The review organizes the surveyed methods into traditional unsupervised and low-rank plus sparse decomposition unsupervised models, and presents theoretical perspectives and AI-assisted research solutions.

    Source: §11 · p. 43

  • Takeaways & Limitations

    The paper brings method reviews, evaluation resources, and future-oriented perspectives together for infrared small target detection.

    Source: §11 · p. 43

    Pixel-level metrics have size aggregation bias: global pooling lets large or numerous targets dominate while small, sparse targets contribute negligibly.

    Source: §9.1.1 · p. 38

Abstract

from arXiv · show

Infrared small target detection is often used in different civilian and military applications to detect static or moving targets in scenes like in ground, air and sea surveillance. First valuable research has emerged since 30 years ago and the two last decades witnessed very significant publications in this field following the rise of low-rank and sparse decomposition (model-based methods) and deep learning models (data-based methods). Furthermore, several surveys can be found in the literature but none of them addresses an overall review in this field. In this context, the aim of this paper is to provide a complete survey of the traditional and recent approaches in infrared small target detection. First, we categorize the different approaches found in the literature in terms of the modeling models, i.e mathematical models, machine learning models and signal processing models. Then, we have discussed them in terms of the critical situations that they claim to handle. Furthermore, we present the available resources, datasets and libraries. Then, we conclude with several promising directions for future research. For the sake of reading, this survey is divided into two parts: 1) "Traditional and Recent Approaches for Infrared Small Target Detection : Overview, Challenges and Perspectives (Part 1)" about conventional and low-rank and sparse decomposition (model-based methods), and 2) "Traditional and Recent Approaches for Infrared Small Target Detection : Overview, Challenges and Perspectives (Part 2)" about deep learning models (data-based methods).

1. Introduction

Infrared small targets are difficult to detect because they are dim, few-pixel signals with little shape or texture, often obscured by clutter. This survey addresses the lack of an overall review by organizing the literature, challenges, resources, and future directions.

1. Introduction

  • Low signal-to-clutter ratios and weak shape or texture make infrared small targets difficult to detect in complex clutter.Targets may be static or moving; the paper distinguishes these as SIRST and MIRST, respectively.Source: §1 · p. 1

1.1. Related Works

  • Earlier surveys covered subsets of infrared small target detection, while the authors identify a need for an overall review.For example, Zhang and Peng (48) review low-rank and sparse decomposition, while Zhao et al. (51) review single-frame models.Source: §1.1 · p. 2
  • The review excludes small space target detection and infrared small target tracking.Source: §1.1 · p. 2

1.2. Motivation and Contributions

  • The review aims to serve as a reference for researchers, industry developers, students, and reviewers, with resources and codes accessible through an accompanying website.The authors describe the paper and website as a one-stop source for models, algorithms, implementations, and benchmarking techniques.Source: §1.2 · pp. 2–3
  • The survey covers more than 950 papers across 30 years, organizing methods by mathematical, machine learning, and signal processing tools and by frame and target type.It presents the works in historical order, including initial studies, variants, improvements, and later developments.Source: §1.2 · p. 4
  • It also describes datasets and evaluation protocols, analyzes solved and unsolved challenges, and offers theoretical perspectives for future work.The resources discussion includes conventional and newer performance metrics.Source: §1.2 · p. 4

2. General Overview

The survey’s publication counts show growing research activity, with deep learning forming the largest reported method category. It also traces the development of modeling tools across infrared small target detection and related areas.

2.1. Factual Overview

  • The main journal venues include IEEE Transactions on Geoscience and Remote Sensing (152 papers) and Infrared Physics and Technology (72 papers).The survey also identifies IEEE IGARSS, IAPR ICPR, and AAAI as major conference venues.Source: §2.1 · p. 4
  • The field’s annual publication output grew substantially over the decade, from 23 papers in 2017 to 320 in 2026, after reaching 255 papers in 2025.Source: §2.1 · p. 4
  • Deep learning accounts for 553 papers, approximately 51% of the publications; mathematical unsupervised models account for 201 and low-rank and sparse decomposition for 154.Source: §2.1 · p. 4
  • At least 70% of publications come from China laboratories, and Wei An is identified as the most active researcher with more than 30 papers.Source: §2.1 · p. 4

2.2. Conceptual Overview

  • The overview places traditional methods and low-rank plus sparse decomposition in Part 1, while deep neural networks are reviewed in Part 2.The figure also situates these model families in a timeline of modeling tools, computer vision, and infrared small target detection.Source: §2.2 · p. 5

3. Applications of Infrared Small Target Detection

The survey considers infrared small target detection across ground, sea, and air environments, whose backgrounds and target appearances pose different detection challenges. It also catalogs methods used in these settings.

3. Applications of Infrared Small Target Detection

  • Applications span ground, air, and sea environments, each with distinct properties and challenges.Source: §3 · p. 5

3.1. Ground Environment

  • On the ground, geometric urban layouts, changing vegetation, and visual similarities between targets and backgrounds complicate detection.Source: §3.1 · p. 5

3.2. Sea Environment

  • At sea, weak ship edges, changing surface interference, and varied ship proportions and directions contribute to missed and false detections.Source: §3.2 · p. 5
  • Sea-environment methods include traditional unsupervised models, low-rank and sparse decomposition, and supervised detection and segmentation networks.Infrared-visible fusion is described as compensating for single-modal imaging deficiencies under low illumination and complex sea states (137).Source: §3.2 · p. 5

3.3. Air Environment

  • Airborne systems detect aircraft, missiles, airplanes, and birds at distances that can extend to several hundred kilometers, against sky radiation and cloud backgrounds.The survey notes saliency-based, low-rank plus sparse decomposition, and deep neural network approaches for air environments.Source: §3.3 · p. 5

4. Preliminaries, Challenges and Issues

Infrared small targets are few-pixel, low-contrast signals with limited structural cues, and their detectability depends on sensor conditions and cluttered backgrounds. The survey groups the challenges by sensor, target, and background, and outlines modeling tool families.

4.1. Preliminaries

  • Targets are typically smaller than 9×9 pixels, and dimness is often associated with local SCR below 3 despite no unified quantitative consensus.Zhang et al. (147) instead classify targets occupying less than 0.12 percent of image pixels as small.Source: §4.1 · p. 6

4.2.1. Sensor Challenges

  • Sensor challenges include fixed or moving platforms, four wavelength bands, low signal-to-noise ratios, vignetting, and weak-intensity noise or streaks.The wavelength categories are NIR, SWIR, MWIR, and LWIR; strong vignetting can create a brighter, nearly saturated center.Source: §4.2.1 · pp. 6–7, Figure 3 · p. 7

4.2.2. Target Challenges

  • Targets lack distinct texture and may provide insufficient structural features for reliable identification; infrared and visible imagery also differ in shape and texture characteristics.The survey distinguishes extended, spotted, and point targets.Source: §4.2.2 · p. 7

4.2.3. Background Challenges

  • Cloud edges, reflective water, vegetation, and urban geometry create varied clutter, while dynamic backgrounds and target-background similarity further hinder detection.Water lines and waves may also distort scene outlines and interfere with detection, recognition, and positioning.Source: §4.2.3 · p. 7

4.3. Issues

  • The survey identifies mathematical, machine learning, and signal processing tools as approaches to low signal-to-noise data and targets with little texture.It describes statistical, fuzzy, and Dempster-Shäfer methods for uncertainty, low-rank models for robust representation, deep learning for varied situations, and temporal pixel distributions for background estimation.Source: §4.3 · p. 7

5. Conventional Unsupervised Methods

This section reviews conventional unsupervised infrared small-target detection methods based on mathematical concepts and signal-processing filters. It covers single-frame and multi-frame approaches organized around background consistency, target saliency, dim-target enhancement, clutter or vibration handling, and feature fusion.

5.2. Single-Frame Detection

  • Single-frame methods either exploit targets disrupting local background correlations or use low-rank decomposition when nonadjacent pixels share similarities.Local-prior methods detect local contrast, gradients, or abnormal points, while low-rank approaches formulate detection as iterative sparse and low-rank optimization without labels.Source: §5.2 · p. 8

5.2.1. Classification /Clustering Methods

  • PA-AHDPC improves signal-to-clutter ratio gain by approximately 5 to 10 dB over the second-best SOTA algorithms, while IDPC enhances clustered targets and multitarget detection.PA-AHDPC uses hyperbolic geometry and a gradient-based soft decision to reject pseudo-targets; IDPC fuses local and gray-level global features.Source: §5.2.1 · pp. 8–9

5.2.2. Discriminative Methods

  • FICA-TD combines frequency-data sphering, independent component analysis, and sparsity filtering, outperforming seven state-of-the-art methods on five real infrared sequences.The method constructs a spatial-temporal tensor from adjacent video frames and suppresses low-frequency and residual low-response information.Source: §5.2.2 · p. 9

5.3. Consistency of Backgrounds

  • Background-consistency filters predict background from pixel correlations and subtract it; Top-Hat methods depend on structuring-element selection, and adaptive QGA optimization is more stable than traditional filtering.Wavelet and contourlet methods decompose images and threshold fused or energy-based representations; classical Top-Hat methods can lose targets or produce false alarms amid noise and clutter.Source: §5.3 · p. 10, §5.3.1 · p. 10, §5.3.2 · p. 10

5.3.3. Least Mean Square Filters

  • LMS filters predict pixel intensity from neighboring pixels, producing larger prediction errors near targets; Zhao et al. (174) report better detection and competitive efficiency with bilateral filtering.Bilateral filtering fuses leftward and rightward prediction-error images and uses a fixed step size and simplified structure to improve speed over prior TDLMS algorithms.Source: §5.3.3 · p. 11

5.4.1. Local Contrast Methods

  • IDoGb produces fewer false alarms than ICLM at the same detection rate, taking about 0.1 s per frame; local-contrast methods can be costly across multiple window scales.Local-contrast methods enhance targets relative to their surroundings, but absolute contrast may fail amid noise or bright clutter; multiscale windows increase computation and processing time.Source: §5.4.1 · p. 12

5.4.2. Global Contrast Methods

  • MSGCM improves on GCM by 2.95 times in background suppression factor, 6.17 times in signal-to-clutter ratio gain for single-target detection, and 10.49 times for multitarget detection.Global contrast methods address cases where targets are not locally salient; TGCM uses a globally decomposed baseline, while HCM combines relative global and local-patch contrast.Source: §5.4.2 · p. 12

5.4.3. Transform Domain Methods

  • Transform-domain methods magnify frequency differences between target and background, using phase information for detection, while HVS methods in spatial or frequency domains perform better in complex scenes than filter-based methods.Phase-spectrum methods risk information loss and missed detections; QDCT supports four channels, whereas traditional DCT uses grayscale information.Source: §5.4.3 · p. 13

5.6. Multi-Frame Cumulation Algorithms

  • The earlier cumulation methods (213; 214; 215) enhance both target signals and streak noise, while method (214) combines spatial and temporal information to preserve targets in the difference image.Method (213) estimates background from temporal pixel information, method (215) from spatial information, and each produces a noisy difference image.Source: §5.6 · p. 13
  • Ring-perturbation correction and directional multi-frame cumulation (216) enhance dim-target local SNR and detect targets effectively in strong clutter, outperforming methods (213; 214; 215).The method corrects image non-uniformity, uses Poisson sampling and high-order cumulants to extract candidates, then estimates motion to accumulate along the target direction.Source: §5.6 · p. 13
  • IABP with high-order cumulant target enhancement and sequential energy-center motion detection (217) outperforms previous background-prediction methods and increases image signal-noise ratio.The detection integrates neighborhood, continuity, area, energy, and other target motion characteristics.Source: §5.6 · p. 13

5.7. Miscellaneous Algorithms

  • The maritime pipeline-filtering method (219) corrects inter-frame vibration using block matching with normalized cross-correlation and mutual information, enabling accurate small-target detection.It corrects coordinates at the single-frame level before completing pipeline filtering.Source: §5.7 · p. 13
  • Peng et al. (218) fuse nonlinear dim-target features into an integrated feature using a Karhunen-Loeve transform.Source: §5.7 · p. 13
  • IGVF (220) enhances targets with multiscale flux density, suppresses residual clutter noise with gradient direction diversity, and outperforms previous SOTA methods in complex backgrounds.It transforms the image into an infrared gradient vector field and uses adaptive thresholding to separate targets.Source: §5.7 · pp. 13–14

6. Low-rank and Sparse Decomposition Unsupervised Models

Low-rank plus Sparse matrix Decomposition (LRSD), also known as Robust Principal Component Analysis (RPCA), is a dominant model-based approach for infrared small target detection. Its variants extend the decomposition to model noise and spatio-temporal structure.

6. Low-rank and Sparse Decomposition Unsupervised Models

  • LRSD jointly optimizes nuclear-norm background regularization and the l1-norm, providing a mathematical foundation distinct from filtering and human-visual contrast enhancement.The passage describes robustness to complex dynamic infrared backgrounds and variable-sized targets across heterogeneous environments.Source: §6 · p. 17
  • Candès et al. (67) introduced RPCA in 2009 for robustly handling sparse outliers, and Zhou et al. (316) extended it in 2010 to include a noise matrix.Source: §6 · p. 17
  • Stable RPCA names the three-term low-rank, sparse, and noise decomposition, while tensor-domain extensions (317) explicitly account for spatio-temporal constraints.Source: §6 · p. 17
  • In moving-object detection, LRSD supports spatio-temporal, incremental, and deep unrolled RPCA algorithms, with incremental methods identified as crucial for real-time applications (318–327).Source: §6 · p. 17

6.1. Mathematical Preliminaries

This section frames low-rank and sparse decomposition (LRSD) as separating a correlated background from sparse target outliers, then surveys its objective terms, solvers, and infrared-specific adaptations. Its suitability depends on the decomposition assumptions and choices such as the rank estimate and loss function.

6.1. Mathematical Preliminaries

  • Under minimal assumptions, Principal Component Pursuit (PCP) perfectly recovers the low-rank and sparse matrices using nuclear- and l1-norm terms [67].The balance parameter is usually λ = 1/√max(m,n).Source: §6.1 · p. 17
  • LRSD models the background as a low-rank matrix and targets as sparse outliers; thresholding the sparse component produces a target mask.A rank set too high can absorb targets, while one set too low can miss multimodal backgrounds; separation is best after optimization converges.Source: §6.1 · p. 17, Figure 4 · p. 18
  • The nuclear norm is a convex surrogate for rank, but equal penalization of singular values can over-shrink targets; weighted and partial-sum variants address this limitation.The Frobenius norm is a nuclear-norm proxy but fails to recover the low-rank matrix without a rank estimate and poorly suppresses strong background edges.Source: §6.1 · p. 18
  • Solver choices trade off iteration count and SVD efficiency, with regularization-based methods and statistical-inference methods forming two broad categories.Regularization-based examples include SVT, APG, and ALM; statistical approaches use priors and Bayesian inference to approximate posterior distributions.Source: §6.1 · p. 19

6.1.1. Definition and Classifications of Outliers

  • Outliers in infrared small-target detection are element-wise target values, while background values are inliers; RPCA also distinguishes element, row or column, and missing-data corruption.Missing data can result from an acquisition problem.Source: §6.1.1 · p. 19

6.1.2. Application to Infrared Small Target Detection

  • Infrared LRSD formulations extend basic decomposition to account for noise, spatial and temporal constraints, real-time needs, and imaging challenges.They add background and target regularizers alongside low-rank, sparse, and noise terms; the background term may use a mixed norm, while the target term may use gradients or total variation.Source: §6.1.2 · pp. 19–20

6.2. Single-Frame Detection

This section reviews single-frame infrared target detection using low-rank and sparse decomposition (LRSD), including methods that separate targets from backgrounds and later variants that modify the original model’s priors, constraints, and solvers. It also covers deep-unfolded LRSD networks and their approaches to detection performance, efficiency, robustness, and temporal-feature limitations.

6.2. Single-Frame Detection; 6.2.1. Computation Time

  • The IPI model uses patch-image low-rank and sparse decomposition, while Sobolev-norm and compressive variants address the computation cost of nuclear-norm minimization.IPI relies on background patch self-correlation and the small size of targets; the Sobolev-norm approach reduces computation time, while the compressive method uses a Walsh Hadamard measurement matrix.Source: §6.2 · p. 20, §6.2.1 · p. 20

6.2.2. l1-norm Defects

  • WIPI adds steering-kernel weights to suppress strong edges while retaining dim targets, and DGRLA combines spatial-feature graph regularization with an l1/2-norm constraint to reduce false positives.WIPI formulates separation as column-wise weighted RPCA, though computing structural priors adds time; DGRLA is reported as more robust than IPI and GRLA in complex scenes with cloudy clutter, sea glitter, and man-made interference.Source: §6.2.2 · pp. 20, 22

6.2.3. Nuclear-norm Defects

  • NIPPS, TV-PCP, NRAM, and NOLC modify low-rank or sparse constraints to improve background estimation, clutter suppression, or target recovery over IPI.NIPPS uses a partial sum of singular values and a non-negative target constraint; TV-PCP adds total variation to the low-rank matrix. NOLC uses Schatten q-norm and Lp-norm constraints, and is reported to detect more targets and suppress background better than IPI, RIPT, and NRAM, with faster convergence than IPI and NRAM.Source: §6.2.3 · p. 22

6.2.4. Adaptivity

  • Adaptive LRSD methods vary weights or parameters across patches or singular values, while image entropy and patch statistics guide parameter regulation in complex backgrounds.Wang and Qin's patch-level weighting allows parallel computing; Gao et al. use local weighted entropy, and Yuan and Dong's method is reported more accurate with a lower false-alarm rate than IPI in extremely variant and complex backgrounds.Source: §6.2.4 · p. 23

6.2.5. Heterogeneous Backgrounds

  • For heterogeneous backgrounds, GreB approximates the low-rank component by bilateral factorization, while GPCP is reported to outperform IPI and NRAM on ground, sea, and sky backgrounds.SCPLBMA partitions images by statistics to preserve natural structure; the passage also reports that LRR-based modeling does not suppress clutter and noise as well as IPI.Source: §6.2.5 · p. 24

6.2.6. Robustness to noise; 6.2.7. Temporal Constraints

  • Noise and temporal constraints are addressed with a Markov-random-field-guided mixture-of-Gaussians model and with interframe similarity and total-variation regularization, respectively.The Bayesian MoG approach separates targets from complex background noise; experiments report that TV-IIPI outperforms IPI, ECA-STT, and PSTNN in sea and ground scenes. TNLRS also combines non-local patches and entropy-based saliency filtering to preserve target and surrounding context.Source: §6.2.2 · p. 22, §6.2.6 · p. 24, §6.2.7 · p. 24

6.2.8. Deep Unfold LRSD Networks

  • RPCANet++ unfolds relaxed RPCA into background, object-extraction, and image-restoration modules, and its memory and contrast-prior modules support background preservation and target extraction.Experiments report that RPCANet++ outperforms NRAM and RPCANet; the earlier RPCANet also uses an unfolded relaxed-RPCA model for interpretable target extraction and image reconstruction.Source: §6.2.8 · p. 25
  • Deep-unfolding variants address patch limitations, local structure, and scene adaptivity; FIRST-Net's 5-frame configuration achieves 74.6% detection accuracy on high-velocity targets, 76.8% IoU, and 7.8 × 10−5 false alarm suppression.LRRNet applies low-rank constraints in the image domain rather than to patch images, and DUSRNet is reported to outperform RPCANet on NUDT-SIRST, SIRST, and IRSTD-1k in highly complex and variable scenes. DRPCA-Net generates iteration-wise parameters conditioned on the input scene.Source: §6.2.8 · p. 25
  • L-RPCANet matches or surpasses RPCANet (383), DRPCANet (397), and RPCANet++ (385), while using fewer parameters and faster GPU inference.Liu et al. (385) use hierarchical bottlenecks for channel-wise feature refinement, noise reduction, and SENet channel attention to support robustness and real-time detection.Source: §6.2.8 · p. 26
  • ESTNet outperforms IPI (336) and RPCANet (383), using learnable CUR decompositions to replace SVT and CLSTM modules to capture multi-frame target motion.Zhu and Feng (399) formulate detection as bilevel-optimized RPCA and describe ESTNet as the first unfolding network integrating CUR decomposition.Source: §6.2.8 · p. 26
  • Across SIRST V1, NUDT-SIRST, and IRSTD-1K, LST-RPCANet outperforms listed data-based and prior unfolding methods with lightweight parameters.Wang et al. (388) introduce MCASTM to purify cross-stage background features, LSTEM to learn sparse target representations, and DAGFEM to enhance reconstruction fidelity.Source: §6.2.8 · p. 26
  • RPCA-unfolded networks require high-dimensional SVD that slows inference, and unfolding networks neglect temporal motion features while focusing on spatial structure.Existing methods also face cascading background-estimation errors and insufficient adaptability of fixed handcrafted sparse priors, which LST-RPCANet is designed to address.Source: §6.2.8 · p. 26

6.3. Sparse Representation

Sparse-representation methods build dictionaries to distinguish target signals from background clutter. Across the described approaches, dictionaries are adapted to image content or learned online, and one method outperforms several contrast-based methods.

6.3. Sparse Representation

  • In 2014, Li et al. (417) trained an adaptive morphological dictionary online and split it into target and background dictionaries for sparse decomposition.The approach reports stronger clutter and dim-target feature capture than structural dictionaries, and a larger sparse-feature difference than offline-learned, manually classified discriminative dictionaries.Source: §6.3 · p. 26
  • In 2016, Wan et al. (418) used frequency saliency to select candidate regions, then applied non-negativity-constrained sparse representation and Sparse Rate segmentation.Their target dictionary uses a two-dimensional Gaussian model with extent-feature constraint and a background term; candidate-region optimization uses APG.Source: §6.3 · p. 26
  • The 2019 online-learned double-sparse background dictionary decomposes images into background, target, and noise, and outperforms six listed methods.Lu et al. (419) report better results than Top-hat, ILCM (191), RLCM (199), LIGP (300), DECM (274), and SRvDOCD (417).Source: §6.3 · pp. 26–27

6.4. Multi-Frame Detection

Multi-frame detection uses temporal information that efficient single-frame methods omit, with robust low-rank plus sparse tensor decomposition as a prominent approach. The survey traces tensor models and their adaptations to challenging backgrounds, computation, and multiple subspaces.

6.4.1. Mathematical Preliminaries

  • Tensor RPCA separates an observed tensor into a low-rank background and a sparse target component, extending matrix decomposition to several possible tensor representations.The Image-Patch Tensor model stacks a single-frame image into a patch tensor and separates background and target before post-filtering.Source: §6.4.1 · p. 27, §6.4.2 · p. 27

6.4.2. RPCA Tensor based Decomposition

  • Weighted, nonconvex, and contrast-based tensor constraints address shortcomings of nuclear- and l1-norm formulations, with reported gains over earlier methods.Examples include STIPT-WNSM, PSTNN, NTRS, and STFFT; the latter uses a partial tubal nuclear norm.Source: §6.4.2 · pp. 27–28
  • Spatial-temporal TV methods incorporate inter-frame information to model target continuity and suppress dynamic-background or motion interference.STTV-WNIPT combines weighted tensor nuclear norm, spatial-temporal TV, l1-norm, and Frobenius-norm terms; later methods add 3D-TV and motion-saliency priors.Source: §6.4.2 · pp. 28, 31
  • ASTIT-WPTAR adapts temporal steps using SSIM to represent background changes, and is reported to outperform five comparison methods.The 5D-STFC model instead uses a whitened spatial-temporal patch tensor and joint spatial-temporal norms; it reports better results than FAMSIS, SSTTC, 4D-TR, and AGPCNet.Source: §6.4.2 · p. 31
  • Facet-derived directional features provide target priors: FDMDEA-STT incorporates multidirectional edge awareness and reports better results than RIPT and ECA-SST.Facet-LRSD combines directional derivatives with a structure-tensor eigenvalue matrix to enhance targets and suppress background clutter.Source: §6.4.2 · p. 31
  • Graph-regularized tensor methods combine low-rank background estimation with sparse-target regularization; HGLAPR and FGLR-MCP report gains over their stated baselines.HGLAPR uses hypergraph Laplacian regularization and an asymmetric penalty; FGLR-MCP uses fast graph regularization and MCP instead of nuclear and l1 norms, respectively.Source: §6.4.2 · p. 31
  • For heterogeneous backgrounds, RIPT uses local and non-local priors, while ANLPT adapts weights using image-block entropy and reports better results than RIPT, PSTNN, and NTFRA.RIPT replaces globally constant weighting with entry-wise local structure-adaptive and sparsity-enhancing weights.Source: §6.4.2 · pp. 31–32

6.4.3. RTC Tensor based Decomposition

  • RTC tensor completion incorporates structural target priors and background self-correlation, while multi-subspace learning extends subspace modeling to heterogeneous scenes and spatial-temporal tensors.NRLTC reports better results than IPI and NARM; MSLSTIPT uses spatial-temporal tensor correlation and a multi-subspace property to separate target and background.Source: §6.4.3 · p. 32, §6.4.4 · p. 32

7. Resources and Codes

The survey catalogs online resources and code repositories for infrared small target detection, including toolboxes for training, testing, benchmarking, and optimization-based research.

7.1. Target Detection WebSite

  • The Target Detection Website lists 610 references, 19 datasets, and 161 codes, classified by mathematical, machine-learning, and signal-processing models.The website regularly updates and classifies its listings.Source: §7.1 · p. 34

7.2. Codes

  • BasicIRSTD offers a PyTorch pipeline and benchmark across method categories, while TensorISTD provides a MATLAB evaluation and prototyping pipeline for optimization-based detection.TensorISTD covers single- and multi-frame detection, matrix and tensor representations, and several evaluation metrics.Source: §7.2 · p. 34

8. Datasets

The survey organizes datasets by modality, scale, target density, and environment. It describes a progression from sparse-target visual datasets toward dense, large-scale, and text-paired resources, while noting dataset-specific constraints.

8.1. Visual Modality Datasets

  • Visual datasets include small- and large-scale resources for sparse and dense targets, with newer datasets also categorized by target density.The survey distinguishes visual-only datasets from datasets containing image and text modalities.Source: §8 · p. 34, §8.1 · p. 34, §8.1.1 · p. 34

8.1.2. Small-Scale Datasets

  • Early sparse-target datasets do not represent frames containing densely clustered targets, motivating increased attention to dense infrared small target detection.NUST-SIRST remains a historical baseline, but its synthesized targets and inflated masks violate the SPIE small-target definition.Source: §8 · p. 34, §8.1.2 · p. 34
  • Dense-target resources include DenseSIRST and datasets for moving targets and drone swarms, addressing clustered targets and dense moving-target scenarios.DMIST-60 and DMIST-100 each contain more than 50 small targets per frame; DSISTD integrates real drone targets into simulated infrared backgrounds.Source: §8.1.2 · pp. 35–36
  • Small-scale datasets have varied limitations: SIRST-V2 size and variable resolution constrain training and evaluation, while other real datasets have narrow coverage or annotation issues.NUAA-SIRST has 427 images and many inaccurately labeled pixels; MWIRSTD's fixed resolution limits generalization to other sensor configurations.Source: §8.1.2 · p. 35

8.1.3. Large-Scale Datasets

  • Large-scale datasets provide diverse scenes and many images or videos for training data-based models, including IRDST and the multi-source WideIRSTD collection.IRDST contains 142,727 frames; WideIRSTD combines public datasets and NUDT-developed simulated and real space-based data.Source: §8.1.3 · p. 36

8.1.4. Air Environments

  • Air-environment datasets cover bird monitoring, varied authentic target shapes and clutter, and synthetic moving-target scenarios blended into real backgrounds.MIST models variations in target pose, size, intensity, and motion.Source: §8.1.4 · pp. 36–37

8.1.5. Maritime Environmments

  • Maritime datasets include ship imagery, but some have limited complex clutter: NUDT-SIRST-Sea lacks wave-clutter images and IRSDSS has about 10% wave-clutter and complex-cloud backgrounds.Maritime-SIRST contains 1,131 images and 2,647 targets from Landsat-8 infrared imagery.Source: §8.1.5 · p. 37

8.2. Visual and Text Modalities Datasets

  • Since 2025, image-text datasets have been introduced for evaluating language and vision-language models in infrared small target detection.IR-TXPair provides textual descriptions alongside images, while FZDT pairs images with fuzzy semantic text prompts.Source: §8.2 · p. 37

9. Performance Evaluation

Conventional IRSTD evaluation uses pixel-level and target-level metrics for different formulations of detection, alongside complexity measures for deep models. These measures capture distinct aspects of performance, and several have important blind spots.

9.1.1. Pixel-Level Metrics

  • Pixel-level metrics assess mask agreement, but Pixel Accuracy is uninformative when targets occupy less than 0.1% of image pixels, and PSNR and SSIM measure background fidelity rather than detection.Normalised IoU averages per-sample ratios so each sample contributes equally, regardless of target size.Source: §9.1.1 · pp. 38–39
  • IoU is the primary segmentation metric, but global pooling lets large or numerous targets dominate while small, sparse targets contribute negligibly.IoU and F1 remain pixel-wise measures and cannot capture target-level spatial localisation.Source: §9.1.1 · p. 38

9.1.2. Target-Level Metrics

  • Target-level metrics include Pd and false-alarm measures, but Pd ignores false positives and Fa can undercount small false positives while over-penalising large ones.SCRG and BSF jointly describe target-contrast enhancement and clutter flattening; high SCRG alone does not guarantee detection without suitable thresholding.Source: §9.1.2 · pp. 39–40

9.1.3. Complexity Metrics

  • Conventional protocols also use FLOPs to indicate a deep model’s computational complexity and parameters to indicate its size and computational burden.Pixel-level metrics treat detection as binary segmentation, while target-level metrics treat it as target-instance localisation.Source: §9.1 · p. 37, §9.1.3 · p. 40

9.2. Common Ranking Protocols

A common ranking protocol linearly weights IoU and Pd to select a best method. The paper argues this ranking does not reflect real respective performance and calls for fairer, application-aware recommendations.

9.2. Common Ranking Protocols

  • The ICPR 2024 Resource-Limited Infrared Small Target Detection Challenge ranks methods using a linear weighted sum of IoU and Pd.The authors report that this protocol is unsuitable and does not reflect methods’ real respective performance.Source: §9.2 · p. 41
  • A method’s performance depends on its environmental application, so one method need not be best in every challenge to suit a particular application.The authors call for alternative ranking protocols that provide engineers with fair recommendations.Source: §9.2 · p. 41
  • The protocol’s single best-method ranking is not, according to the authors, a reliable reflection of comparative performance.They propose designing alternative ranking protocols rather than relying on that ranking for recommendations.Source: §9.2 · p. 41

9.3. Disruptive Protocols

Conventional protocols use fragmented metrics and overall scores that limit comprehensive capability assessment, error analysis, and understanding of cross-scenario generalization. The section reviews hybrid metrics that combine pixel- and target-level evaluation.

9.3. Disruptive Protocols

  • Conventional evaluation separates pixel- and target-level metrics, emphasizes overall scores, and predominantly uses dataset-specific training and testing, limiting error analysis and cross-scenario generalization.The paper identifies changing metrics or ranking protocols as ways to revise evaluation.Source: §9.3 · p. 41

9.3.1. Hybrid-Level Metric

  • Hierarchical IoU combines target-level localisation and pixel-level segmentation through Overlap Priority with Distance Compensation matching.Its multiplicative form prevents weakness in either component from being masked by the other.Source: §9.3.1 · p. 41

9.3.2. Holistic SIRST Evaluation Metric

  • Holistic SIRST Evaluation combines pixel-level confidence quality, measured by HSE-P, with target-level detection robustness, measured by HSE-T, into a unified score.The two sub-metrics are fused multiplicatively so neither pixel-level precision nor target-level detectability masks deficiencies in the other.Source: §9.3.2 · p. 41

9.4. Protocols for Data-based Methods

The passage points readers to Part 2 for evaluation protocols specific to supervised deep neural network methods.

9.4. Protocols for Data-based Methods

  • Protocols for supervised deep neural network methods are treated in Part 2 as complementary to the conventional protocols reviewed here.The passage identifies these as protocols specific to supervised models.Source: §9.4 · p. 41

10. Prospective Research

The paper assesses current methods and evaluation protocols, identifies unsolved detection challenges, and proposes research directions spanning researcher-designed methods, AI-assisted research, and performance evaluation.

10.1. Analysis of Current Methods and Evaluation Protocols

  • Filtering methods can require manual tuning in complex scenes, while saliency methods can struggle with similar distractors and introduce detection noise.The paper also notes that varying background noise can cause target loss for filtering approaches.Source: §10.1 · p. 42
  • Low-rank and sparse decomposition methods model targets as sparse components and backgrounds as low-rank structures, but can be computationally intensive on high-dimensional data.The paper describes IPI variants as researcher-designed norms and minimization problems tailored to specific challenges.Source: §10.1 · p. 42
  • Current evaluation protocols use fragmented pixel- and target-level metrics, overall scores that limit error analysis, and dataset-specific train-test setups that do not assess cross-scenario generalization.The paper argues that these weaknesses hinder fair model ranking and identification of failure modes.Source: §10.1 · p. 42

10.2. Solved and Unsolved Challenges

  • Unsolved challenges remain in both single-frame and multi-frame detection, while dense scenes expose limitations in local-saliency and sparse-target approaches.Local spatial saliency can yield target adhesion, blurred responses, and incomplete separation when adjacent targets have subtle energy differences from the background; sparse-target models can produce ambiguous structures and unstable localization.Source: §10.2 · p. 42
  • Deep-network backbones process spatial locations homogeneously and do not explicitly encode target density, even when paired with powerful backbones and transformer-based heads.The paper identifies this density-aware semantic guidance deficiency as a challenge for dense scenes.Source: §10.2 · p. 42

10.3.1. Perspectives for Researcher Solutions

  • Researcher-designed directions include exploring Dempster-Shafer theory and addressing orthogonal-complement outliers, sparse-component outliers, or simultaneous corruption of both components.The paper notes that existing LRSD works considered original-observation-space outliers, while PCP may fail on orthogonal-complement outliers.Source: §10.3.1 · pp. 42–43

10.3.2. Perspectives for AI-Assisted Research Solutions

  • The paper proposes AI-assisted research tools to explore combinations of modeling characteristics that may not yet have been investigated for infrared small target detection.It describes variation across decomposition, losses, solvers, architectures, optimization, and training.Source: §10.3.2 · p. 43

10.4. Perspectives for Performance Evolution

  • Performance evaluation should account for sensor and deployment diversity rather than treating one top-ranked method as universally suitable.The paper points to evaluation approaches from moving-object detection for ranking across environments and assessing unseen scenarios; Pang et al. (66) also proposed a hybrid-level metric and open-source benchmarking toolkit.Source: §10.4 · p. 43

11. Conclusion

The paper reviews three decades of infrared small target detection research, categorizes its methods, and surveys evaluation resources and protocols. It concludes with theoretical perspectives and AI-assisted research solutions.

11. Conclusion

  • The review covers 30 years of infrared small target detection research.It surveys traditional unsupervised models and low-rank and sparse decomposition unsupervised models in historical and methodological context.Source: §11 · p. 43
  • The paper presents datasets, metrics, and ranking protocols used to evaluate detection methods.Source: §11 · p. 43
  • The review provides theoretical perspectives and AI-assisted research solutions.Source: §11 · p. 43

Related papers

Loading 2610.18276v1…