SANet: Selective Attention Network for Infrared Small Target Detection

I. INTRODUCTION

• Early convolutional layers can lack sensitivity to very small targets, reducing target-background contrast and detection accuracy and reliability.This limits discrimination of weak targets in complex backgrounds.Source: §I · p. 1
• Static feature connections cannot dynamically adjust transfer in salient regions, making true targets harder to distinguish from false alarms in complex backgrounds.The introduction identifies fixed fusion rules as a limitation of current networks.Source: §I · p. 1
• DSM combines standard and pinwheel-shaped convolutions to broaden receptive fields and suppress background noise, while spatial and channel attention refine fine-grained features.The standard and pinwheel paths are described as complementary feature-extraction components.Source: §I · p. 2
• SAFM replaces conventional U-Net static skip connections with dynamic feature fusion to selectively guide information transfer and suppress false alarms.The introduction describes this as improving distinction between true targets and false detections.Source: §I · p. 2
• SANet compares favorably with existing state-of-the-art detectors across multiple performance metrics on three public datasets.The introduction presents these experiments as validation of the proposed method.Source: §I · p. 2

II. RELATED WORK

• Traditional IRSTD methods use filtering, low-rank decomposition, and local contrast to suppress backgrounds or enhance targets, but each has stated operating limits.Filtering is limited when target and background frequencies overlap; low-rank methods can confuse sparse noise with targets; fixed local windows limit scale adaptation.Source: A. Traditional Methods for IRSTD · p. 2
• Traditional approaches rely on fixed handcrafted features and parameters, limiting accuracy and robustness when targets and backgrounds vary in practical applications.The section identifies this limitation as a motivation for data-driven methods.Source: A. Traditional Methods for IRSTD · p. 2

B. Deep Learning Methods for IRSTD

• Deep-learning IRSTD methods use varied strategies, including adversarial learning, contextual modulation, multiscale fusion, geometric features, and transformer-based representations.The survey also describes frequency-aware, cross-attention, and other feature-fusion approaches.Source: B. Deep Learning Methods for IRSTD · p. 2
• Recent IRSTD work also adapts general-purpose segmentation models and uses state space models, with Mamba-based methods modeling target-detail differences between scanning positions and neighborhoods.The section identifies MiM-ISTD as the first application of Mamba to IRSTD and describes IRMamba's state-update use of pixel differences.Source: B. Deep Learning Methods for IRSTD · p. 2

A. Overall Network Architecture

• SANet uses a conventional U-Net framework with DSM and SAFM as its two key components.The framework is illustrated in Fig. 1.Source: A. Overall Network Architecture · p. 2, Figure 1 · p. 3
• Five encoder units pair DSM with max pooling, while four decoder SAFMs dynamically fuse encoder and decoder features across scales before upsampling and target prediction.A 3 × 3 convolution and sigmoid activation form the detection head.Source: A. Overall Network Architecture · p. 3

B. Dual-path Semantic-aware Module

• DSM pairs standard 3 × 3 convolutions for local spatial consistency with pinwheel-shaped convolutions using 3 × 3, 5 × 5, and 3 × 3 kernels for broader context.The paper motivates DSM by the limited receptive fields of conventional convolutional layers.Source: B. Dual-path Semantic-aware Module · p. 3
• Asymmetric padding creates direction-sensitive horizontal and vertical pinwheel kernels; compared with standard dilated convolution, they require fewer parameters and less computation.The paper links this directional structure to capturing target shapes and adapting to shape variations.Source: B. Dual-path Semantic-aware Module · p. 3, Figure 3 · p. 3
• A 3 × 3 convolution fuses the two DSM branches into a unified feature representation.Source: B. Dual-path Semantic-aware Module · p. 4
• CBAM cascades spatial and channel attention to enhance spatial responses and emphasize globally relevant semantic channels, improving feature discrimination.Source: Figure 4 · p. 3, B. Dual-path Semantic-aware Module · p. 4

• SAFM introduces dynamic spatial selection into U-Net, using multiscale features to emphasize informative, target-related spatial regions.Source: C. Selective Attention Fusion Module · p. 4
• At each stage, SAFM fuses encoder and decoder feature maps after concatenating them along the channel dimension.Source: C. Selective Attention Fusion Module · p. 4
• Average and max pooling of the fused features produce spatial representations that are concatenated into a joint representation for complementary attention.Source: C. Selective Attention Fusion Module · p. 4
• The resulting spatial attention map weights encoder and decoder features through element-wise multiplication to form the fused representation.Source: C. Selective Attention Fusion Module · p. 4
• A 1 × 1 convolution transforms the fused representation before element-wise multiplication with the original features, with a learnable scalar controlling fusion strength.The scalar preserves original feature structure while strengthening discriminative responses in key regions.Source: C. Selective Attention Fusion Module · p. 4

A. Datasets and Experimental Settings

• The three benchmarks contain 427, 1001, and 1327 annotated images, with training-to-test ratios of 1:1, 4:1, and 1:1, respectively.Training images are resized to 320 × 320 for NUAA-SIRST and IRSTD-1K and 256 × 256 for NUDT-SIRST.Source: A. Datasets and Experimental Settings · p. 4

B. Evaluation Metrics

• Evaluation uses IoU, nIoU, detection probability (Pd), and false alarm rate (Fa); metric definitions specify ground-truth, predicted, and true/false positive and negative pixel counts.Training runs for 600 epochs with Soft-IoU loss and AdamW; the saliency-map threshold is 0.5.Source: B. Evaluation Metrics · p. 4, A. Datasets and Experimental Settings · p. 4

C. Quantitative Evaluation

• SANet is compared with 14 existing IRSTD methods—five traditional and nine deep-learning approaches—and, for fairness, all deep-learning models are retrained on the same training benchmark using settings from their original papers.Source: C. Quantitative Evaluation · pp. 4, 6
• SANet achieves the best IoU, nIoU, and false alarm rate, and best or joint-best detection probability across three datasets.On IRSTD-1K, its detection probability of 93.27% equals SCTransNet’s.Source: Table I · p. 5, C. Quantitative Evaluation · p. 6

D. Qualitative Evaluation

• In the complex IRSTD-1K scene in Fig. 6(c), all compared deep-learning methods except UIUNet and SANet produce false alarms, while UIUNet generates many false detections in Fig. 6(d).SANet maintains accurate and robust detection across the illustrated scenes.Source: D. Qualitative Evaluation · p. 6
• In Fig. 7’s multiple-target complex scene, competing methods have difficulty maintaining stable detection, and UIUNet generates many false alarms in the second scene.These visualizations support SANet’s robustness in complex environments.Source: D. Qualitative Evaluation · p. 6

1) Effectiveness of the Core Components of SANet

• The ablation uses conventional U-Net as the baseline and progressively adds DSM and SAFM to assess their individual and combined effects.The module ablations are reported in Tables III and IV.Source: 1) Effectiveness of the Core… · p. 6, Table III · p. 7, Table IV · p. 7
• Adding DSM alone raises IoU by 3.32, 6.71, and 9.63 percentage points on NUAA-SIRST, IRSTD-1K, and NUDT-SIRST, respectively.The study attributes these gains to improved extraction and modeling of small-target details.Source: 1) Effectiveness of the Core… · pp. 6–7
• Adding SAFM further improves IoU, detection probability, and false alarm rate on all three datasets, while nIoU decreases on IRSTD-1K from 67.08% to 66.59%.nIoU improves on NUAA-SIRST and NUDT-SIRST.Source: 1) Effectiveness of the Core… · p. 7

2) Effectiveness of the Dual-path Semantic-aware Module

• Pinwheel-shaped convolution expands the receptive field and improves sensitivity to dim targets, but can increase background interference before attention refinement.On NUDT-SIRST, adding it raises Fa from 4.71 × 10⁻⁶ to 6.96 × 10⁻⁶; spatial and channel attention adapt responses to strengthen target features and suppress background noise.Source: 2) Effectiveness of the Dual-path… · p. 7

3) Effectiveness of the Selective Attention Fusion Module

• Removing either SAFM’s residual connection or learnable scalar λ reduces IoU; the two components make complementary contributions to segmentation accuracy.The learnable scalar improves IoU on all three datasets, while the residual connection helps preserve target information and raises Pd from 89.56% to 93.27% on IRSTD-1K; some ablated variants nevertheless achieve lower Fa or higher nIoU on individual datasets.Source: 3) Effectiveness of the Selective… · p. 7

F. Model Complexity Analysis

• 82.18% mean IoU across three benchmarks, exceeding the other methods by 3.84–4.56 percentage points; SANet uses 12.72 million parameters and 19.75 GFLOPs.The comparison covers DNANet, UIUNet, SCTransNet, and L2SKNet.Source: F. Model Complexity Analysis · p. 7
• Although SANet is not the smallest model, its computational cost is moderate relative to UIUNet and SCTransNet, supporting a favorable accuracy–complexity trade-off.The paper notes its detection performance supports potential use in practical infrared imaging applications.Source: F. Model Complexity Analysis · pp. 7–8

V. CONCLUSION

• SANet’s DSM combines standard and pinwheel-shaped convolutions, while SAFM replaces static U-Net skip connections with dynamic fusion to strengthen discrimination between targets and false alarms.The conclusion associates these modules with small-target perception, discriminative feature representation, and background suppression.Source: §V · p. 8
• Experiments on multiple public datasets show favorable detection accuracy and robustness across datasets; future work targets lower model complexity and better detection in challenging scenes.The stated aim is to extend applicability to different infrared imaging tasks.Source: §V · p. 8