DGNet: Dual-knowledge Guided Network for Infrared Small Target Detection

Chenglong Yu; Mingzhu Xu; Jing Wang; Tongtong Wang; Pingping Miao; Liqiang Nie

Source-linked AI summary

DGNet: Dual-knowledge Guided Network for Infrared Small Target Detection

Chenglong Yu, Mingzhu Xu, Jing Wang, Tongtong Wang, Pingping Miao, Liqiang Nie

arXiv:2609.00666v1Sep 1, 2026cs.CV

TL;DR

IRSTD text-guided methods face semantic entanglement from jointly modeling backgrounds and targets, as well as deployment constraints from image-specific prompts. DGNet uses dual generalizable textual knowledge through frequency-domain PWM and consensus-based CDA optimization. Experiments on three public datasets demonstrate superior performance and component effectiveness.

Problem
Text-guided IRSTD methods jointly describe backgrounds and targets with one text and rely on image-specific prompts requiring external inference models, creating semantic and deployment limitations.
Method
DGNet combines a PWM module that separately modulates background and target frequency components with a CDA loss that guides optimization from complex background to bright targets.
Results
Experiments on three public datasets demonstrate DGNet’s superior performance and the effectiveness of its key components.
Takeaways & Limitations
Multiple generalizable texts provide DGNet with decoupled frequency-domain semantics and a unified cross-sample optimization trajectory without image-specific textual guidance.

Abstract

from arXiv · show

InfRared Small Target Detection (IRSTD) is a prominent and challenging task in computer vision. In recent years, text-guided methods have significantly improved detection performance. However, they still suffer from two key limitations. First, a single text description simultaneously modeling both background and target leads to semantic entanglement, which contradicts the objective of background suppression and target enhancement. Second, reliance on image-specific textual prompts (requiring additional external models such as CLIP during inference) results in deployment constraints. To address these issues, we propose a novel Dual-knowledge Guided Network (DGNet) based on multiple generalizable texts. Specifically, we design a Prior-knowledge Wavelet Modulation (PWM) module, which leverages dual textual priors that separately characterize large-scale backgrounds and sparse targets to effectively disentangle and modulate entangled semantics in the frequency domain. Furthermore, we introduce a Consensus-knowledge Directional Alignment (CDA) loss, which models the initial state and the ideal target across samples as `complex background' and `bright target', respectively, thereby constructing a clear and unified directional optimization trajectory for the model. Extensive experiments on three public datasets demonstrate the superior performance of DGNet and the effectiveness of each component. The source code is available at https://github.com/iLearn-Lab/MM26-DGNet.

1 Introduction

IRSTD must localize tiny, low-signal targets amid complex infrared backgrounds, while text-guided methods introduce semantic entanglement and deployment constraints. DGNet addresses these challenges with dual generalizable texts, frequency-domain modulation, and consensus-guided optimization.

IRSTD targets occupy few pixels and lack distinct color, shape, or texture, while background clutter makes robust segmentation challenging.
Purely visual methods rely on single-modal information, making discriminative feature extraction difficult and causing false alarms.
Single textual guidance entangles large-scale backgrounds with sparse targets, conflicting with separate background suppression and target enhancement.
Image-specific prompts require external models such as CLIP during inference, increasing computational overhead and limiting practical deployment.
DGNet uses multiple generalizable texts, with PWM decoupling background and target semantics in the frequency domain and CDA defining a cross-sample path from complex background to bright targets.
Ablation and comparative experiments on three public datasets demonstrate DGNet’s superiority and the effectiveness of its key components.

2 Related Work

IRSTD methods have progressed from handcrafted filters, contrast, and low-rank representations to data-driven and text-guided approaches. DGNet targets remaining semantic-entanglement and optimization challenges through dual-frequency modulation and consensus-based directional alignment.

Filter-based, local-contrast, and low-rank methods use handcrafted priors but struggle with false alarms, missed targets, or complex backgrounds.
Deep learning methods learn target features from data and have achieved significant progress in IRSTD.
Single textual prompts jointly model backgrounds and targets, confounding heterogeneous semantics and preventing explicit disentanglement.
Pixel-wise BCE can miss sparse targets, while spatial geometric losses remain susceptible to gradient domination by large background regions.
The network combines a four-stage encoder-decoder, PWM skip connections, and joint CDA and IoU optimization.
DGNet’s CDA loss uses cross-sample semantic consensus to construct a directed optimization path from the source state to the ideal state.

3 Method

DGNet uses dual textual knowledge to separate background and target semantics in feature extraction and to align visual optimization with a consensus semantic trajectory. Its PWM module operates in the frequency domain, while CDA training uses CLIP-space visual and textual states to guide predictions toward ideal targets.

Overall Architecture: DGNet is an end-to-end multimodal encoder-decoder whose PWM modules connect corresponding encoder and decoder layers.The architecture uses four encoder-decoder stages with skip connections, progressive downsampling, upsampling, and feature refinement.
Prior-knowledge Wavelet Modulation: PWM decomposes features into low- and high-frequency subbands, then uses background and target textual priors to modulate them separately.DWT produces one low-frequency and three high-frequency subbands; background guidance acts on low frequencies, while target guidance acts on high frequencies.
Prior-knowledge Wavelet Modulation: Background-guided modulation suppresses large smooth regions, whereas target-guided modulation enhances sparse, discriminative high-frequency details.The B-KGM uses a suppression gate, while T-KGM uses globally salient responses and enhancement gating before inverse wavelet transformation and residual fusion.
Consensus-knowledge Directional Alignment: Consensus knowledge models all samples as transitioning from complex background to bright targets, avoiding image-specific textual guidance during the optimization objective.The ultimate loss combines CDA with IoU supervision to retain pixel-level learning while providing high-level semantic direction.
Consensus-knowledge Directional Alignment: Mask-guided fusion creates CLIP-encodable predicted and ideal visual states from the infrared image, predicted map, and ground-truth map.The fused states and original image are passed through the frozen CLIP image encoder to obtain Vp, Vs, and Vg; the fusion ratio is r=0.8.
Consensus-knowledge Directional Alignment: CDA aligns the visual optimization direction ΔV with the semantic direction ΔT defined by fixed consensus texts.The loss constrains the visual transition from source feature Vs toward predicted feature Vp to follow the trajectory from background clutter toward salient small targets.

4.1 Datasets and Evaluation Metrics

Experiments use three public IRSTD datasets with dataset-specific train/test splits and evaluate both pixel-level segmentation and object-level detection.

Experiments cover IRSTD-1K, SIRST, and NUDT-SIRST, containing 1001, 427, and 1327 infrared images, respectively.
IRSTD-1K and SIRST use 4:1 training/testing splits, whereas NUDT-SIRST uses equal halves for training and testing.
IoU evaluates pixel-level performance, while Pd and Fa evaluate object-level detection.
ROC curves are plotted using different true-positive-rate and false-positive-rate settings.

4.2 Implementation Details

DGNet is trained in PyTorch with CLIP-ViT-B/32 providing text and image encoders during training, but CLIP is omitted during inference.

DGNet uses PyTorch on one NVIDIA GeForce RTX 4090 GPU and trains for 600 epochs with batch size 16.
Training uses Adam with an initial learning rate of 5e-4, reduced by 0.9 at epochs 300 and 450.
Input images are resized to 256 × 256, and CLIP-ViT-B/32 serves as the text and image encoder during training.
CLIP is not used during inference, so the deployed model incurs no additional CLIP overhead.

4.3 Quantitative Comparison

DGNet achieves strong quantitative performance across IRSTD-1K, SIRST, and NUDT-SIRST, while ROC results and the reported design emphasize competitiveness without inference-time image-specific text.

DGNet reaches 72.72% IoU and 4.25 Fa on IRSTD-1K while maintaining high Pd.
On SIRST, DGNet achieves 100% Pd, 82.68% IoU, and 1.24 Fa.
On NUDT-SIRST, DGNet achieves 95.78% IoU and 99.37% Pd.
The model uses dual knowledge to modulate visual features and does not require complex image-specific text design.
DGNet achieves higher TPR at lower FPR than other compared methods on the IRSTD-1K ROC curves.

4.4 Qualitative Comparison

Qualitative comparisons examine DGNet and seven representative methods across challenging infrared scenarios where clutter, occlusion, and low signal-to-noise ratios impair detection.

Purely vision-based methods can produce false alarms under complex backgrounds.The cited example identifies FGARNet as affected in rows 1–3, column 4.
Dense cloud occlusion and extremely low signal-to-noise ratios make discriminative target-background feature extraction difficult for most methods.

4.5 Ablation Study

Ablation studies show that combining PWM and CDA produces the strongest detection behavior, while wavelet decomposition, directional alignment, and ratio selection each contribute to performance in complex scenes.

PWM and CDA integration: The full DGNet achieves the best overall performance and significantly outperforms the purely visual baseline.The baseline shows extremely high false alarms and the worst IoU and Pd on IRSTD-1K and SIRST.
PWM and CDA integration: Adding PWM and CDA suppresses background interference while improving small-target detection, reducing false alarms and missed targets.The results attribute feature modulation against complex backgrounds to PWM and the cross-sample semantic optimization trajectory to CDA.
PWM module: Removing wavelet decomposition noticeably reduces performance because DWT separates high-frequency edges from low-frequency smooth components for text-guided modulation.The comparison uses the spatial-only variant, ‘w/o wave’, against the PWM design.
CDA loss: The full CDA loss gives the best detection performance on both datasets by providing a high-level optimization path based on cross-sample semantic consensus.The CDA ablation compares constraint terms using the PWM-equipped network as the baseline.
Ratio slider: The ratio slider affects background prominence and CLIP image encoding, with r=0.4 and r=0.6 producing false alarms while r=1.0 produces missed targets.The paper reports the best results on both datasets under an appropriate intermediate configuration that retains weak global structural information.

4.6 Computational Efficiency

DGNet uses CLIP only during training for feature modulation and alignment, excluding it during inference and avoiding additional inference computation.

Inference design: CLIP is introduced during training but omitted during inference, so DGNet adds no external encoder overhead at deployment.The model therefore maintains a relatively efficient inference time.

5 Conclusion

DGNet addresses semantic entanglement and deployment limitations in text-guided IRSTD through multiple generalizable texts, PWM, and CDA, with effectiveness demonstrated on three public datasets.

Conclusion: DGNet uses multiple generalizable texts to address semantic entanglement from single prompts and deployment limitations from image-specific prompts.The approach combines prior knowledge and consensus knowledge rather than relying on image-specific text during inference.
Conclusion: PWM disentangles background and target semantics in the frequency domain, while CDA guides optimization along a cross-sample path from complex background to salient target.The conclusion describes these as the two main components of the proposed network.
Conclusion: Extensive experiments on three public datasets demonstrate the effectiveness and superiority of DGNet and its designed loss function.The supplied conclusion states the overall experimental validation without specifying dataset-wise metrics.