Unsupervised and Supervised Approaches for Infrared Small Target Detection : Overview, Challenges and Perspectives (Part 2)

• Complex changes in target size, shape, and clutter make handcrafted features and fixed hyper-parameters difficult to use, while data-based methods learn representations through training.Source: §1 · p. 1
• CNNs are the most employed models for detection and segmentation, and the survey organizes them by architectures, losses, attention, and guidance mechanisms.Source: §1 · p. 1
• The survey also covers wavelet-based, Mamba-based, transformer-based, adversarial, and Mixture-of-Experts methods, alongside language models and graph/hypergraph learning.Source: §1 · p. 1
• Language-model approaches use textual information with image modalities but require specific datasets.Source: §1 · p. 1
• The paper introduces DNN concepts, surveys model families, and presents performance evaluation before concluding with future perspectives.It also provides overview tables of methods, including labels distinguishing dense-target and moving-target methods.Source: §1 · pp. 1–2

• The preliminaries place DNNs within a progression from earlier architectures such as RBMs, DBNs, and autoencoders to newer language, vision, Mamba, transformer, and graph-based models.Source: §2 · p. 2
• DNN architectures comprise input, hidden, and output layers whose neurons use activation functions; training learns layer weights by minimizing a loss, commonly through back-propagation and gradient descent.Source: §2 · p. 2
• The paper identifies architecture, optimization, loss functions, and training as key DNN characteristics, linking architecture to expressiveness and optimization to datasets and algorithms.Source: §2 · p. 2
• Training data should cover scene types to reduce performance discrepancy on unseen scenes, and data augmentation can address this issue.Source: §2 · p. 2
• Attention and guidance mechanisms, multi-branch structures, and plug-and-play modules can be added according to the target application.Source: §2 · p. 2

3. Application of Deep Neural Networks in Infrared Small Target Detection

• Infrared small-target detection must contend with noise, spatial and temporal constraints, real-time needs, and challenges in imaging sequences.The survey frames dedicated adaptations as changes to architectures, losses, attention, and guidance mechanisms.Source: §3 · p. 2

3.1. Conceptual Neural Networks

• CNN downsampling can erase features of tiny targets, while restricted receptive fields hinder modeling global image relationships needed to distinguish targets from similar background noise.Transformer attention can capture global context, but its smoothing may suppress targets’ high-frequency signals and its computation grows quadratically with image resolution.Source: §3.1 · p. 3
• Transformers model long-range context, but their low-pass filtering can suppress high-frequency infrared targets and their quadratic complexity burdens high-resolution processing.CNN–ViT combinations are also used to compensate for CNNs’ difficulty modeling long-distance features; the survey notes they still have issues.Source: §3.1 · p. 3

3.2. Architectures

• Nested U-shaped architectures use encoder–decoder paths and repeated multi-layer feature fusion to retain small-target representations in deep layers.Different-depth subnetworks suit targets of varying sizes; nested nodes receive features from their own and adjacent layers.Source: §3.2 · pp. 3–4

3.3. Learning/Training

• Deep models require costly annotated data and can lose performance on unseen scenes, motivating few-shot, zero-annotation, and data-augmentation strategies.The survey lists augmentation approaches including random cloning and erasing, prior-guided methods, GANs, diffusion, and hybrids.Source: §3.3 · p. 4

3.4. Loss Functions

• Soft-IoU treats all pixels equally, whereas weighted MSE emphasizes target pixels; BCE and Dice are also presented as alternative losses.The survey notes equal pixel weighting may underemphasize critical target regions such as the centroid.Source: §3.4 · p. 4

3.5. Attention and Guidance Mechanism

• Attention and guidance mechanisms target local/global, semantic, motion, and saliency features, while plug-and-play modules offer mechanisms for existing methods.The survey also reports that ablation studies are used to assess modules’ roles against the challenges they address.Source: §3.5 · p. 4

3.6. Miscellaneous Mechanisms

• Other surveyed designs include dual- and triple-branch models, synergistic frameworks, and refinement mechanisms.Figure 1 contrasts biomedical-image U-Net and UNet++ structures with infrared-target DNA-Net and iSmallNet, alongside a multi-scale nested interaction module.Source: §3.6 · p. 4, Figure 1 · p. 5

4.1. Detection based Methods

• Detection-based methods include general detectors such as Faster R-CNN, SSD, and YOLO, as well as infrared-specific detectors and Transformer-based RT-DETR methods (270).Liu et al. (269) introduced an MLP-based framework, while McIntosh et al. (132) fine-tuned general detection networks using optimized eigen-vectors as input.Source: §4.1 · p. 5

4.2. Segmentation based Methods

• Segmentation-based methods produce both pixel-level classification and localization, using varied mechanisms and architectures including attention, diffusion, Mamba, Transformers, and wavelets.SAM-based approaches face domain discrepancy from infrared imagery’s noise, weak texture, low-contrast targets, and foreground-background imbalance; handcrafted priors may limit generalizability and end-to-end learning (168; 50).Source: §4.2 · p. 6

4.3. Dedicated Architectures

• Dedicated architectures exploit shape-aware representations, multiscale feature fusion, or frequency-domain processing; SBR and CESBR outperform prior methods, while ID-UNet reduces conventional U-Net parameters by 81% and improves segmentation metrics (50; 51; 56).ISTDU-Net reports an AUC of 0.9977 (159); STNet outperforms previous methods on NUDT-SIRST and IRSTD-1K, while LASNet outperforms previous methods in sparse and dense target scenarios (43; 161).Source: §4.3 · pp. 7–8
• SP-KAN outperforms previous state-of-the-art methods in accuracy, robustness, and generalization through sparse nonlinear modeling (25).Its compressed vision-transformer encoder captures long-range dependencies, while pattern-complementarity and sparse-sine perception modules model nonlinear interactions.Source: §4.3 · pp. 8–9

4.4. Learning Mechanims

• Point-supervised methods address costly pixel annotations: LESPS progressively expands point labels, and GMVG-DIR refines structurally coherent pseudo-labels to preserve irregular target shapes (110; 188).LESPS reports gains over fully supervised performance of over 70% in pixel-level IoU and 95% in object-level probability of detection; TGPL reaches 99% of detection-probability performance and outperforms PAL and LESPS under point supervision (113).Source: §4.4 · pp. 9–11
• Weakly supervised methods generate pseudo-masks from point annotations, while source-free and active domain adaptation methods transfer or select target-domain samples to improve adaptation (191; 190; 194; 221; 300).The learning-free pseudo-mask method reports average IoU 4.3% above its second-best competitor, and its hybrid raises average IoU by a further 3.4%; WSSISTD reports performance similar to fully supervised methods (191; 190).Source: §4.4 · p. 11
• Other learning strategies combine edge-target feature fusion (197), contrastive frequency, gradient, or motion cues (89; 302), and data augmentation tailored to image or sequence properties (63; 68; 62; 198).AugTarget improves mIoU but does not preserve temporal correlations; Diff-Mosaic uses pixel and diffusion priors, while HQDA generates infrared backgrounds using image-intrinsic properties (63; 68; 198).Source: §4.4 · p. 12
• The GAN uses infrared-image nonlocal self-correlation to improve generated-data realism and robust random-walker target extraction to preserve the real distribution during target enhancement.Source: §4.4 · p. 13

4.5. Attention Mechanisms

• Top-k and pyramid attention select task-relevant features or fuse multiscale features, while hierarchical and spatial approaches integrate semantic, local, and global information.Examples include LGTS’s global and local sparse attention, PAU2-Net’s pyramid attention encoder, HAFNet’s hierarchical feature-fusion skip connections, and SANet’s global spatial semantics.Source: §4.5 · p. 13
• Direction-induced, contextual, and energy-field attention methods use target shape, multiscale context, or energy changes to enhance targets and suppress clutter.ISNet aggregates edge information and attends across rows and columns; ACM exchanges high-level semantics with low-level details; CEFANet amplifies target energy mutations and guides low-level features with high-level semantic states.Source: §4.5 · pp. 13, 15
• MSC-HA reports average IoU of 75.89%, Pd of 96.70%, and Fa of 8.67 × 10−6 across four datasets.It combines a multiscale convolutional decoder, learnable soft hyperedges for high-order spatial and channel dependencies, and a distribution-aware Wasserstein loss.Source: §4.5 · p. 15

4.6. Adaptive and Multi-Cue Mechanisms

• Adaptive feature-fusion methods reuse and refine features across layers, while AER-Net combines CNN local correlation with Transformer global correlation and TAMNet adapts feature extraction, context, and prediction.MDAFNet is reported to provide performance similar to previous state-of-the-art methods while meeting real-time requirements on an RTX3080 GPU; AER-Net reports significant mean IoU improvements.Source: §4.6 · pp. 15, 17
• RRCA-Net recurrently reuses convolution without extra parameters to refine high-level target features, and is reported to outperform prior state-of-the-art methods with few parameters.Its dual interactive attention aggregation fuses refined information; the network is also described as a plug-and-play module that improves existing IRSTD methods.Source: §4.6 · p. 17
• Diffusion-based methods use continuous high-resolution representations or target priors to capture small-target contours and reduce background interference.DCFR-Net combines implicit neural representation with conditional denoising diffusion; EDPrior uses thermal-radiation diffusion properties, while APCD uses adaptive blocking and multiscale attention to suppress clutter.Source: §4.6 · p. 17
• Multi-branch methods assign distinct branches to target detection, clutter suppression, contrast enhancement, or edge and position cues, then fuse branch information.DB-DCNet is reported to outperform previous state-of-the-art methods on four datasets; DBCE U-Net reports IoU of 94.60% and Pd of 99.05% on NUDT-SIRST.Source: §4.6 · pp. 17–18
• Multiscale and multilevel approaches combine scale-aware feature extraction, fusion, and losses to preserve target detail, improve localization, or handle varying target sizes.STDMANet uses temporal and spatial multiscale refinement with center-point annotations and reports F1 of 0.9744, surpassing state-of-the-art algorithms by 0.1682; HNU-Net reports over 20% fewer parameters while increasing average IoU.Source: §4.6 · pp. 18–19
• SAFARI-Net reports 95.77% IoU on NUDT-SIRST and 70.17% on IRSTD-1, while MFAGNet reports 74.12% mIoU, 99.33% Precision, and 0.0002477 Fa on ISDD.ECMSFNet reports a real-time detection speed of 74.63 frames/s.Source: §4.6 · pp. 20–21
• AGMFNet outperforms ISTDU-Net (159), MTU-Net (284), AGPCNet, DNA-Net (53), UIU-Net (155), DMFNet (434), and TODENet (435), while HPFNet performs similarly to ISNet and UIU-Net.AGMFNet combines receptive-field expansion, dynamic channel weighting, three-level feature fusion, and an imbalance-mitigating loss; HPFNet uses patch-wise global-local extraction, sparse cross-level fusion, and Top-K background filtering to reduce occasional false alarms and missed detections under challenging conditions.Source: §4.6 · p. 20
• MEETNet outperforms previous SOTA methods on NUAA-SIRST and NUDT-SIRST.Its triple-cascaded architecture preserves high resolution, while edge-detail enhancement and multireceptive-field fusion combine target details with deep semantic information.Source: §4.6 · pp. 20–21
• CSAFNet outperforms previous SOTA methods on NUAA-SIRST and IRSTD-1K, while MSSLNet provides performance similar to SOTA algorithms in accuracy and efficiency.CSAFNet combines physical-prior self-attention and wavelets with semantic-guided filtering and a multiscale spatial pyramid; MSSLNet uses sparse large-kernel blocks and bidirectional guided fusion and has 0.293 M parameters and 2.99 GFLOPs.Source: §4.6 · p. 21
• MLIGNet balances detection accuracy and computational efficiency, while MFEFNet reports superior detection accuracy to several SOTA methods on SIRST, IRSTD-1K, and NUDT-SIRST.MLIGNet encodes local intensity and gradient cues and aggregates cross-scale features; MFEFNet combines multishape receptive fields, local-global spatial-channel context, and cross-attention fusion.Source: §4.6 · p. 21
• FGOM-RTDETR improves precision from 0.921 to 0.942, mAP50 from 0.938 to 0.958, and recall from 0.923 to 0.934; MSDA-Net and MAR-UNet outperform their listed comparison methods.MSDA-Net injects parameter-free high-frequency directional priors and calibrates multiscale feature fusion; MAR-UNet combines multigranularity features, linear-complexity channel attention, and boundary refinement.Source: §4.6 · pp. 21–22

Causal Guidance; Memory Guidance; Feedback Guidance

• Causal, memory, and feedback guidance treat targets as background anomalies, preserve spatio-temporal information, or iteratively use prior prediction masks to improve detection robustness [436], [217], [518], [218].CHAL reports outperforming single- and multi-frame models; Tridos and DCPNet report gains over prior methods, with DCPNet reporting a 5% IoU gain in complex scenes.Source: §4.7 · pp. 22, 29

Prior Guidance; Features Guidance

• Prior-guided PSGF combines depth-aware and directional-saliency priors, cross-scale calibration, and Pixel-Selective Gated Fusion to suppress clutter and preserve target integrity, outperforming several previous methods on SIRST, NUDT-SIRST, and IRSTD-1K.Source: §4.7 · p. 22

Features Guidance

• Feature-guidance approaches use handcrafted priors, directional attention, paired inputs, or contrast-invariant features to improve target representation, while AGDFE reports higher scores on NUDT-SIRST and SIRST Aug [55], [449], [450], [453], [452].MDS-DWGANet reports 98.30% Pd and 93.04% mIoU with 1.4 M parameters; AGDFE reports 99.81% vs. 87.58% Precision, 98.96% vs. 88.12% Recall, 99.38% vs. 87.85% F-measure, and 98.77% vs. 78.33% mIoU on NUDT-SIRST and SIRST Aug, respectively.Source: §4.7 · pp. 24–25
• GCFLNet outperforms filter-, local-contrast-, model-, and data-based approaches in heterogeneous backgrounds on NUDT and IRSTD-1K, while ESE-Net outperforms UIU-Net, DNA-Net, and RDIAN on NUAA-SIRST and IRSTD-1K [455], [457].Source: §4.7 · p. 26

Reconstruction Guidance; Semantic Guidance

• Reconstruction and semantic guidance address lost information and semantic degradation through invertible target-aware reconstruction, frequency-separated target/background reconstruction, or joint target detection and background-scene understanding [222], [484], [486], [531].BAFE-Net exchanges target and background semantics across tasks, while SMPISD-MTPNet uses scene perception and expert-informed semantic priors to suppress background false alarms.Source: §4.7 · pp. 26–27

Semantic Guidance

• Semantic-guided methods preserve localization and detail through temporal-semantic interaction, upsampling-free spatial-semantic fusion, detail compensation, frequency-aware guidance, and edge-semantic consistency [77], [104], [481], [482], [487], [106].TSNet, SemDetNet, FSGNet, and ET-DSC report outperforming prior methods; CCDNet uses contrastive distractor discrimination to reduce false alarms, and SDG-ResNet injects semantic-density predictions into intermediate layers.Source: §4.7 · pp. 27–28

Motion Guidance

• Motion-guided detectors encode motion maps or temporal features and fuse them with spatial cues; MI-DETR uses no extra motion labels or alignment, while MOCID targets a throughput-performance balance [146], [274], [275].MI-DETR, MIST-Net, and MOCID report advantages over comparison methods on their evaluated datasets.Source: §4.7 · p. 28

Saliency Guidance; Physics-Informed Guidance

• Saliency guidance uses a learnable center-subtracts-neighbors kernel, while physics-informed PI-SAM embeds infrared contrast and SNR priors and is reported to improve the SAM baseline [214], [85].PI-SAM converts physical cues into prompt heatmaps and applies contrast- and SNR-aware regularization with sparsity and total-variation constraints.Source: §4.7 · pp. 28–29

Guidance Mechanisms

• Experiments report DFINet outperforms several SOTA methods, while Huang et al. (219) describe FeedbackSTS-Det 47’s paired forward/backward semantic-refinement loop.Its Sparse Semantic Module models long-range temporal dependencies at low computational cost, and a consistent training-inference pipeline supports performance transfer and robustness.Source: §4.7 · p. 29

Noise-Suppression Guidance

• NSDSAM combines noise-suppressing adapters with spatial-frequency processing, and experiments report SOTA performance.Its self-attention adapter uses entropy-aware feature reconstruction and gated soft-attention fusion, while SFHM addresses bias toward low-frequency components.Source: §4.7 · p. 29

Geometry Guidance

• GeoRankNet aligns convolutional sampling to irregular target contours and dynamically ranks features; experiments report improved accuracy and robustness on diverse SIRST datasets.Its modules are GAD2C, IoU- and loss-feedback-driven Top-K attention, and gated fusion of channel, spatial, and global context.Source: §4.7 · p. 29
• RPR-Net+PAMG bridges point supervision and mask-level detection through pseudo-mask generation and reciprocal feedback between geometric predictions and pixel-level mask recovery.Experiments report that RPR-Net+PAMG outperforms the listed methods, including ACM (37), DNA-Net (53), and SCTransNet (285).Source: §4.7 · p. 30

Ordinary Differential Equation Guidance

• OFSPNet combines CNN, transformer, and Mamba features with ODE-inspired adaptive fusion, while FreqODEs use frequency interaction and stacked neural ODE modules.OFSPNet’s optimization strategies increased inference speed by 15.28%; FreqODEs were reported superior on NUAA-SIRST and IRSTD-1k.Source: §4.7 · p. 30

Fuzzy-Logic Guidance

• FLINet uses local and global fuzzy reasoning plus discrepancy-guided feature fusion; experiments report superior performance across multiple metrics on three SIRST datasets.Its fuzzy branches address target contrast and clutter, while fusion prioritizes shallow-layer details and deep-layer semantic cues.Source: §4.7 · p. 30

High Resolution Guidance

• H2D-Net uses hierarchical high-resolution guidance for bidirectional interaction between spatial detail and multiscale representations, and a bottleneck filter interrupts false feedforward responses.Its two components are the High Resolution Guidance module and the Holistic Distractor Filter.Source: §4.7 · p. 30

Robust Algorithms

• RISTDnet combines handcrafted features and CNNs with a likelihood map and threshold to detect varied-size, low-SNR targets in complex backgrounds.The survey reports better robustness than previous algorithms.Source: §4.8 · p. 30
• SANet consistently outperforms DNA-Net (53), UIU-Net (155), SCTransNet (285), and L2SKNet (214) on NUAA-SIRST, IRSTD-1K, and NUDT-SIRST.Its spatially adaptive fusion replaces static skip connections for context-aware, cross-scale feature fusion.Source: §4.8 · pp. 30–31
• MISTNet outperforms 31 previous methods on the MIST dataset, with a 6% gain in IoU.Its shifted neighborhood compensation block models multi-scale correspondences for motion compensation, while progressive distillation filters redundancy.Source: §4.8 · p. 31

Sparsity Algorithms

• SAGCP uses sparse target-region modeling to identify redundant channels; on DNA-Net, pruning reduces parameters by 72.34% and increases IoU by 2.02%.ACSNet addresses sparse-data inefficiency with adaptive differentiable binarization and sparse convolution for lightweight feature extraction.Source: §4.8 · p. 31

Aliasing Rectification Algorithms

• EAA-Net filters channels affected by aliasing and uses PDE-based channel pruning; experiments on IRSTD-1k and SIRST report better results than state-of-the-art methods.The two modules target aliasing effects and model efficiency, respectively.Source: §4.8 · p. 31
• FGARNet outperforms the second-best method by 2.74% on IRSTD-1k.Its Bidirectional Frequency Filter isolates frequency bands, while Wavelet Rectifying Downsampling corrects aliasing using wavelet-derived high-frequency features.Source: §4.8 · p. 32

Low Illumination Algorithms

• Nv-Net provides the best reported detection performance on the LII dataset in low illumination, while FDI-Net and CFD-Net outperform previous state-of-the-art methods under low illumination.FDI-Net combines convolution and Transformer features with a Navier–Stokes refinement module; CFD-Net uses forward pixel flow and dynamically selected paths for targets of varying shapes and orientations.Source: §4.8 · p. 32

Lightweight Algorithms

• LR-Net achieves state-of-the-art performance and ranks third in the ICPR 2024 Resource-Limited Infrared Small Target Detection Challenge Track 2.Its low-level feature distribution strategy addresses small-target loss in high-level feature maps.Source: §4.9 · p. 32
• TBC-Net uses a Target Extraction Module for inference and a Semantic Constraint Module during training; the survey reports reduced false alarms and real-time detection on an NVIDIA Jetson AGX board.LR-Net combines lightweight feature extraction and refinement with low-level feature guidance and cropping during training and inference.Source: §4.9 · p. 32
• MDPNet outperforms previous networks on SIRST and NUDT-SIRST with 0.23 M parameters and an inference speed of 137 FPS for 320 × 320 images.A multi-scale dilated pyramid enlarges the receptive field, and semantic distillation transfers information from a teacher Transformer to a student convolution network.Source: §4.9 · p. 32
• ILNet integrates important shallow-layer features into deep layers, while IR-WSANet combines wavelet-based frequency enhancement with space-channel dual attention to improve detection performance in complex scenes on SIDD and HIT-UAV.Source: §4.9 · pp. 32–33
• LLCEFA separates local-contrast prior processing and edge-weight extraction into two branches, making it a plug-and-play enhancer with minimal additional burden.TFSP-IR adapts pruning to spatial relationships, while STMG-PCNet combines texture retention and shape-aware global context for accuracy and efficiency.Source: §4.9 · pp. 33–34
• UAMF reduces computational complexity by two orders of magnitude compared with previous state-of-the-art models, while WSNet is reported as the field’s most lightweight model and fastest inference speed to date.UAMF adaptively fuses multi-frame inputs using a memory-fusion module; WSNet has 0.054 M parameters and requires 1.050 G FLOPs.Source: §4.9 · p. 34
• BiisNet retains full-precision features within binarized convolutions and is reported to outperform other binary architectures while competing with state-of-the-art full-precision models.The survey notes that existing pruning methods are challenging to apply directly because infrared small targets have weak texture, shape, and color distribution.Source: §4.9 · p. 34
• SGDP protects important layers, groups coupled structures, and decorrelates substructures to balance compression with small-target detection performance.The surveyed pruning challenges include removing essential modules and ignoring cross-layer dependencies or substructure correlations, which can degrade performance or recovery and generalization.Source: §4.9 · p. 34

Wavelets based Methods

• HaarConv improves detection accuracy for dim and small infrared ship targets by more than the passage’s stated threshold.Its Haar-based high-frequency features and HaarUp-HaarDown modules are paired with pseudo-label-based semi-supervised learning; the supplied passage does not state the threshold value.Source: §5 · p. 34
• IRPruneDet reduces parameters by 64.13% and FLOPS by 51.19%, while IoU improves from 73.31% to 75.12% and nIoU from 70.92% to 74.30%.Wavelet-domain regularization induces structural sparsity, and soft channel reconstruction retains key target information against premature pruning.Source: §5 · p. 35
• WMRNet outperforms previous state-of-the-art algorithms on IRSTD-1k and SIRST, using wavelet subbands in state-space equations and reversible edge refinement.FSCNet addresses downsampling information loss by enhancing high-frequency details and interacting across multi-scale features; WaveTD is reported to outperform the listed methods including MSHNet (40), WTAPNet (559), and DNA-Net (53).Source: §5 · p. 35
• SWAN outperforms ACM (37), ALCNet (312), DNA-Net (53), RDIAN (202), UIU-Net (155), SCTransNet (285), and DATransNet.It combines wavelet convolution with spatial attention and residual dual-channel attention; WRSSNet uses synthetic pseudo-infrared data and semi-supervised learning and outperforms DNANet (53) and UIU-Net (155).Source: §5 · pp. 35–36
• ARFC-WAHNet outperforms most of the time the listed model-driven and data-driven methods on SIRST, NUDT-SIRST, and IRSTD-1k.Its modules combine multi-receptive-field feature interaction, wavelet-enhanced downsampling, high-low feature fusion, and global attention.Source: §5 · p. 36
• WT-HMFF provides better results than MDvsFA (243), ACM (37), DNA-Net (53), Pconv (7), ALCNet (312), IRPruneDet (318), UIU-Net (155), and SCTransNet (285) most of the time.Its wavelet convolution captures global context while HMFF fuses shallow and deep features to preserve high-resolution target details.Source: §5 · p. 36
• WCDMF-Net reports performance similar to state-of-the-art methods on NUDT-SIRST, NUAA-SIRST, and IRSTD-1k; DFAwareNet reports competitive or superior performance on four benchmark datasets.WCDMF-Net fuses spatial and spectral representations; DFAwareNet combines wavelet-guided refinement, a linear-complexity vision state space model, and multi-dilated aggregation.Source: §5 · pp. 36–37
• PMNet outperforms previous state-of-the-art methods in accuracy and robustness, while IRKAN reports 68.89% mean normalized IoU and 8.33 × 106 False Alarm Rate on WideIRSTD-Full.IRKAN also reports cross-resolution generalization in zero-shot evaluations on unseen NUDT-SIRSTD and IRSTD-1K datasets.Source: §5 · p. 37

6. Mamba-based State Space Models

• MiM-ISTD captures global visual-sentence and local visual-word features; on NUAA-SIRST and IRSTD-1k it is 7 times faster than several SOTA methods and uses 62.2% less GPU memory on 2048 × 2048 images.The nested Outer and Inner Mamba blocks model the two feature scales.Source: §6 · p. 37
• Other designs target dense, small, or weak targets with multiscale feature integration, edge-aware modules, or reconstruction refinement.FANM is reported to outperform several algorithms in dense scenarios; CPMamba provides better results than several named methods most of the time.Source: §6 · pp. 37, 39, 40
• IRMamba uses pixel-difference information and layer restoration, while EAMNet uses parallel feature processing and adaptive filtering; both are reported to improve on prior methods.MRNet targets missed detections and false alarms in complex scenes; SBMambaNet is reported to outperform most of the time a broad set of model-based, data-based, hybrid, and Mamba methods.Source: §6 · p. 38
• Mamba-based approaches combine local and global features, including CNN–Mamba hybrids, nested structures, and scan modules designed to reduce redundancy.HMCNet combines CNN local features with global context; MOU-Mamba uses multi-order and local-guided selective scans. MOU-Mamba outperforms SOTA methods with an effective accuracy–complexity balance.Source: §6 · pp. 37–39
• For heterogeneous backgrounds, SMILE combines spatial and spectral Mamba with dual-path aggregation and local-feature refinement, and outperforms 14 previous algorithms.MRL-Mamba uses nested global–local feature learning and outperforms 12 previous methods.Source: §6 · p. 39
• Video-oriented methods model temporal relationships or target motion: SAMamba aligns feature streams, while STLMamba uses frequency cues and local-motion modeling.TAPM-Net models target-induced feature propagation along trajectories and achieves SOTA performance on NUAA-SIRST and IRSTD-1K.Source: §6 · pp. 39–40

7. Transformers-based Neural Networks

• RKformer combines convolution and Transformer encoding with random-connection sparse attention; the review reports SOTA performance.MTU-Net uses a ViT–CNN hybrid encoder and Copy–Rotate–Resize–Paste augmentation for tiny-ship detection on NUDT-SIRST-Sea.Source: §7 · pp. 40–41
• Hybrid designs pair CNN local-feature extraction with Transformer global-context modeling, while attention-guided fusion combines their features.FATC-Net outperforms previous SOTA algorithms on BSIRST-v2 for probability of detection, F1 score, and intersection over union; STASPPNet reports 96.9% average accuracy for mAP50 and 49.7% for mAP50-95 on NUAA-SIRST.Source: §7 · pp. 41–42
• Several architectures fuse spatial, temporal, and channel information or model motion across frames, including STC, SCTransNet, and LMAFormer.LMAFormer is reported to provide results similar to previous SOTA methods on NUDT-MIRSDT, IRDST, and TSIRMT.Source: §7 · pp. 41–42
• Other surveyed models use semantic tokens, sparse spatial-channel attention, or reconstructed backgrounds to handle small-target features and clutter.STPSA-Net reports superior performance to previous SOTA approaches; HST-Net reports SOTA performance, while STUTD reports satisfactory overall and real-time performance.Source: §7 · p. 42
• The section also covers adaptive or scale-aware methods, including IrisNet, KECS-Net, GMTNet, DTFE-Net, and SGMP-UNet, with reported improvements over prior methods.These approaches include adaptive decoding, multiscale feature enhancement, spatial-frequency fusion, and motion-aware representations.Source: §7 · pp. 42–43
• 86.7%, 88.1%, and 99.5% mAP@50 are reported for LCR-RTDETR on SIRST, IRSID-1 K, and NUDT-SIRST, respectively.The same method reports 14.4% fewer parameters, 10.9% lower computational load, and inference speed gains of 5.8 FPS, 3.4 FPS, and 5.7 FPS on Jetson Orin NX.Source: §7 · p. 43

9. Adversarial Networks

• MDvsFA-cGAN divides its generator into miss-detection and false-alarm subtasks.The surveyed methods also include a GAN that predicts target intensity through a U-Net generator.Source: §8 · pp. 43–44
• A GAN approach formulates detection as image-to-image translation, reconstructing targets with a U-Net generator and a five-layer discriminator, and reports better IoU than SOTA methods across various backgrounds and targets.The model combines adversarial loss with an l2 localization loss.Source: §8 · p. 44
• DAENet simulates infrared imaging degradation during adversarial pretraining and combines wavelet–Gabor frequency features with spatial gradient enhancement for weak targets.The review reports that DAENet outperforms SOTA algorithms in complex environments.Source: §8 · p. 44

9. Mixture-of-Experts (MoE)

• CoMoE targets cross-domain joint learning with domain-specific and global prototypes, Top-K domain-aware expert routing, and noise-guided contrastive feature alignment.Its stated aim is to move beyond using one universal detector simultaneously across multiple datasets.Source: §9 · p. 44
• FAME-Net models high-frequency structure and low-frequency thermal context with cross-frequency experts, mutual-guided routing, and semantic-guided gating.The review reports SOTA detection accuracy and boundary recovery across multiple benchmarks.Source: §9 · p. 44

10.1. Large Language Models (LLMs)

• SAIST combines scene-recognition contrastive language-image pretraining with a CLIP-guided Segment Anything Model and outperforms seven named methods on MIRSTD (249).Its prompts guide background-feature learning in the mask decoder, alongside infrared imaging equations.Source: §10.1 · p. 44
• MIRSAM adapts SAM with contourlet denoising and CLIP text prompts, outperforming MDvsFA, ACM, ALCNet, and ISNet on IR-TXPair (250).The denoising adapter targets noise during feature extraction and encoding.Source: §10.1 · p. 44
• Text-IRSTD uses fuzzy semantic prompts and a progressive cross-modal decoder, with better detection and contour recovery than five named methods on FZDT (251).Source: §10.1 · p. 44
• TIPP uses text-derived modulation factors to guide spatiotemporal-feature polymerization for incremental few-shot learning, with reported effectiveness on IRDST, DAUB, and ITSDT (252).Source: §10.1 · p. 45

10.2. Vision-Language Models (VFMs)

• Conventional CNN, GAN, and transformer methods are described as limited in global context, efficiency, annotation demands, multimodal learning, spatial features, temporal appearance matching, and motion representation.The section connects these stated weaknesses to approaches that incorporate textual and visual information.Source: §10.2 · p. 45
• FDEP adds frozen vision-foundation-model representations to encoder-decoder methods; experiments report improved HSE-series accuracy for eight named methods (258).Its SAMF module fuses semantic priors with task-specific features, while CO-ISD transfers semantics between main and lightweight branches.Source: §10.2 · p. 45
• SeViL combines adaptive text prompts, pseudo-label filtering, and cross-modal masking, matching fully supervised methods on three datasets with 10% labeled training samples (257).Source: §10.2 · p. 45
• SPIRIT uses physical constraints in spatial and temporal modules, while MoPKL and iMoPKL use language-guided motion priors and report gains over prior methods (259; 255; 256).SPIRIT can operate as a spatial detector when temporal information is unavailable; iMoPKL is reported to outperform MoPKL.Source: §10.2 · pp. 45–46

11.1. Preliminaries

• The section distinguishes spectral GNNs, which learn node representations through graph spectral theory, from spatial GNNs, which aggregate local information to describe pairwise relations.It also describes transductive prediction on partially labeled data and inductive prediction on unseen data.Source: §11.1 · p. 46
• Hypergraphs connect multiple vertices through a hyperedge to represent higher-order relations that a simple graph may lose.Source: §11.1 · p. 46

11.2. Application to Infrared Small Target Detection

• CS-ViG-UNet builds a sparse graph with cyclic-shift graph attention and reports F1 scores of 0.8561 on SIRST AUG and 0.745 on IRSTD-1K (630).On an RTX3090 with TensorRT acceleration, it processes approximately 357 256×256 images per second.Source: §11.2 · p. 46
• Graph-based context learning in GCLNet combines local features and multiple graph-reasoning modules, with IoU of 80.26% on NUAA-SIRST and 94.84% on NUDT-SIRST (263).Source: §11.2 · p. 46
• HyperTea integrates CNNs, RNNs, and hypergraph neural networks to model high-order spatiotemporal correlations for moving infrared small-target detection (105).Its global and local temporal modules enhance context and adjacent-frame motion patterns, while temporal alignment addresses cross-scale feature misalignment; it outperforms ISNet, RPCANet, and SSTNet on DAUB and IRDST.Source: §11.2 · p. 46

• These protocols complement conventional protocols because supervised deep-learning methods have specific characteristics that require evaluation for fair comparison.Source: §12 · p. 46
• The proposed evaluation considers the compromise between detection performance and the number of parameters, illustrated by comparing LW-DSTransNet with earlier algorithms.Source: §12 · p. 46
• A second protocol examines the impact of plug-and-play single-point-supervised modules PAL, LESPS, and DtoD across several SOTA algorithms.The cited figures compare their impact on ISNet, DNA-Net, UIU-Net, GGLNet, and MS-DANet.Source: §12 · p. 47

13. Prospective Research

• The section frames future work around analysis of current methods and protocols, review of solved and unsolved challenges, and theoretical perspectives for method categories.Source: §13 · p. 47

13.1. Analysis of Current Methods

• Deep-learning methods improve detection in complex images through automatic feature extraction, but face annotation costs, possible loss of critical semantics, computational demands, and limitations with unseen data.Single-frame methods also lack temporal information, which makes high detection accuracy difficult to ensure.Source: §13.1 · p. 47

13.2. Evaluation Protocols

• Conventional evaluation protocols use fragmented metrics, offer limited error analysis, and rely on dataset-specific splits that do not reveal robustness across infrared scenarios.Source: §13.2 · p. 47

13.3. Prospective Solutions

• Proposed directions include graph-based learning with less labeled data, alternative learning strategies, and hypergraph methods for unsolved challenges.Source: §13.3.1 · p. 47

13.3.1. Perspectives for Researcher Solutions

• Inductive learning is presented as a way to predict on unseen data, while transductive learning uses labeled and unlabeled data together without a clear training-testing distinction.The text states that inductive learning is more robust in unseen cases.Source: §13.3.1 · p. 47
• Contrastive-aided Distractor Discriminator uses adaptive local and global similarity between targets and backgrounds to discriminate distractors and reduce false alarms (483).Source: §13.3.1 · p. 47
• The section presents performance-versus-parameter comparison for LW-DSTransNet and figures on plug-and-play module impacts as evaluation perspectives.The figures are described as comparisons, without reported outcomes in these passages.Source: Figure 2 · p. 48, Figure 3 · p. 48, Figure 4 · p. 48
• Hypergraph learning represents higher-order relations by connecting multiple vertices in a hyperedge, relations that simple graphs may lose.Source: §13.3.1 · p. 48

14. Conclusion

• The review covers 30 years of infrared small target detection research and categorizes supervised deep neural network models historically.For each category, it surveys first works, related research, variants and improvements, and recent developments.Source: §14 · p. 48
• The paper presents the specific metrics and ranking protocols used to evaluate performance.Source: §14 · p. 48
• The paper concludes with theoretical perspectives and AI-assisted research solutions.Source: §14 · p. 48