PaperlayerView exact revision on arXiv

Source-linked AI summary

ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection

Tongtong Wang, Mingzhu Xu, Chenglong Yu, Jing Wang, Xiaohui Lin, Weili Guan

arXiv:2609.00853v1cs.CVcs.AI

TL;DR

IRSTD methods struggle to distinguish weak infrared targets from clutter, and single-prompt multimodal methods inadequately represent asymmetric target and background semantics. ADGNet introduces asymmetric prompts, separate text-guided branches, adaptive fusion, and the AITIR dataset; it reports consistent superiority over 21 existing SOTA methods.

  • Problem

    Vision-only methods struggle to distinguish weak targets from clutter, while single textual prompts inadequately represent infrared targets and backgrounds with asymmetric semantics.

  • Method

    ADGNet combines an abstract fixed target prompt, a detailed background prompt, independent target-localization and background-suppression guidance, adaptive feature aggregation, and asymmetric annotations for AITIR.

  • Results

    ADGNet consistently outperforms 21 existing state-of-the-art methods in complex scenarios.

  • Takeaways & Limitations

    The proposed asymmetric design separately guides target enhancement and clutter suppression for accurate segmentation in complex infrared scenes.

  • Takeaways & Limitations

    Vision-only methods remain limited by infrared images’ asymmetric semantics and pixel-level reliance when distinguishing true targets from similar background clutter.

Abstract

from arXiv · show

InfRared Small Target Detection (IRSTD) is a challenging task. Relying solely on pixel-level information, vision-only methods struggle to distinguish targets from clutter. Current multimodal methods typically describe both targets and backgrounds with a single textual prompt. Such an approach lacks dedicated regional guidance and ignores infrared semantic asymmetry. Consequently, it provides insufficient background suppression information and introduces severe feature optimization conflicts, overwhelming small targets with noise. To address these issues, we propose a novel Asymmetric Dual-text Guided Network (ADGNet). Specifically, accounting for the infrared semantic asymmetry, we first design the Asymmetric Dual-text Prompt (ADP), comprising an image-agnostic abstract target prompt and an image-specific detailed background prompt. To leverage these prompts, we introduce an Asymmetric Dual-Branch Interaction (ADBI) module to separately guide visual features with their respective text priors, protecting targets from noise while suppressing background clutter. Subsequently, we introduce an Adaptive Feature Aggregation (AFA) module to dynamically fuse features from the two branches. Furthermore, we construct a multimodal Asymmetric Image-Text Infrared (AITIR) dataset by providing asymmetric text annotations for three public datasets (IRSTD-1K, NUDT-SIRST, and SIRST). Extensive experiments demonstrate that ADGNet outperforms 21 state-of-the-art (SOTA) methods. Code is available at https://github.com/iLearn-Lab/MM26-ADGNet.

1 Introduction

IRSTD must extract weak, semantically sparse targets from cluttered, structured backgrounds, but existing vision-only and single-prompt multimodal methods provide insufficient discrimination and suppression guidance. ADGNet addresses this asymmetry with separate target and background prompts, independent feature guidance, and dynamic fusion.

  • Infrared targets have weak, singular semantics and little texture, whereas backgrounds contain rich structure and clutter, making target extraction difficult.
  • Vision-only methods rely on pixel-level information and therefore struggle to distinguish true targets from morphologically similar background clutter.
  • Single-prompt multimodal methods over-describe simple targets, under-specify complex backgrounds, and merge distinct semantics into conflicting feature optimization.
  • ADGNet uses an abstract target prompt and detailed background prompt, then independently guides target localization and background suppression before dynamically fusing both branches.
  • The paper constructs AITIR by adding asymmetric text annotations to three public datasets, and reports performance exceeding 21 existing SOTA methods.

2 Related Work

Traditional and vision-only IRSTD methods face semantic and robustness limitations, while text-guided methods add explicit guidance but commonly retain holistic target-background prompting. ADGNet instead separates these semantics through asymmetric prompts and branch-specific interaction.

  • Traditional IRSTD methods use hand-crafted features, with trade-offs among speed, noise sensitivity, false alarms, computational cost, and complex-background robustness.
  • Vision-only networks learn hierarchical features but struggle to distinguish weak targets from morphologically similar clutter because infrared semantics are asymmetric.
  • Text-guided methods introduce explicit semantic guidance through vision-language models, but existing approaches typically use holistic descriptions for targets and backgrounds.
  • ADGNet separates target and background semantics with abstract target prompts and detailed background prompts, followed by independent text-vision guidance and adaptive fusion.

3 Asymmetric Image-Text Infrared Dataset

ADGNet’s AITIR dataset extends existing infrared benchmarks with asymmetric textual descriptions that keep target semantics abstract while detailing scene context and local clutter.

  • Holistic target-background descriptions can add semantic noise to tiny targets while providing insufficient priors for suppressing complex background clutter.
  • The Asymmetric Dual-text Prompt uses an image-independent Fixed Target Prompt and an image-dependent Detailed Background Prompt.
  • The Fixed Target Prompt is the abstract phrase ‘an infrared image featuring one or multiple target’, while the Detailed Background Prompt specifies macro scene and local thermal clutter.
  • Template-guided background descriptions are refined collaboratively, filtered for meaningless characters, limited to 18 words, and paired with the fixed target prompt.
  • This asymmetric annotation process extends three existing datasets into the multimodal AITIR dataset while leaving original images and masks unchanged.

4 Method

ADGNet addresses infrared semantic asymmetry by separately guiding target localization and background suppression with asymmetric text priors, then dynamically fusing the resulting features. Its ADBI module uses word-level target guidance and sentence-level background semantics, while AFA produces the fused representation for decoding.

  • Asymmetric Dual-text Prompt: ADGNet uses an asymmetric dual-text prompt with abstract target semantics and detailed background semantics to provide distinct visual guidance.The target prompt supplies word-level features, while the background prompt supplies sentence-level features.
  • Asymmetric Dual-Branch Interaction: ADBI separates target localization from background suppression into independent branches, avoiding interference between target enhancement and clutter filtering.The TL branch uses target word-level priors, whereas the BS branch uses global sentence-level background semantics.
  • Target Localization Branch: The TL branch flattens visual features and applies dense cross-attention with target semantics to preserve high-resolution details during weak-target localization.The resulting target-localization feature is modulated by pixel-level spatial responses intended to highlight targets without introducing noise.
  • Target Localization Branch: The TL projection uses GELU and dropout before cross-modal interaction, with GELU preserving near-zero feature variations associated with faint thermal signatures.The projected visual representation is formed from the flattened feature through convolution, GELU, and dropout.
  • Target Localization Branch: The target-localization operation activates target pixels in feature heatmaps, preventing weak targets from being erased and reducing noise interference.The final target-localization feature is reshaped to C×H×W after element-wise modulation.
  • Background Suppression Branch: The BS branch projects sentence-level background semantics into channel-wise gates that suppress thermal-clutter channels while preserving spatial context.The gated feature is reported to suppress complex clutter responses and reduce false alarms in the corresponding heatmaps.
  • Adaptive Feature Aggregation: AFA aligns and concatenates the two branch features, extracts pooled spatial statistics, and generates dynamic spatial weights for adaptive fusion.A 7×7 convolution produces two-channel spatial attention weights that separately modulate the aligned target-localization and background-suppressed features.

5 Experiments

Experiments on the AITIR benchmark show that ADGNet delivers strong quantitative and qualitative IRSTD performance, while ablations validate its asymmetric prompts, dedicated interaction branches, and adaptive fusion.

  • Quantitative Comparison: ADGNet achieves the highest IoU on IRSTD-1K, NUDT-SIRST, and SIRST, reaching 72.38%, 95.53%, and 83.08%, respectively.It also attains the highest Pd on NUDT-SIRST and SIRST: 99.47% and 100%.
  • Quantitative Comparison: ADGNet balances detection quality and false-alarm control more effectively than the compared methods on challenging infrared datasets.On IRSTD-1K, it records the lowest Fa of 4.10; ROC curves on SIRST show higher TPR while bounding FPR.
  • Qualitative Comparison: Visual comparisons show that ADGNet reduces false alarms in cluttered scenes and missed detections for dim or multiple targets.The compared vision-only methods lack semantic priors for clutter suppression and explicit target guidance for sparse features.
  • Ablation Study: Adding ADBI and AFA improves the vision-only baseline, with the full model increasing IRSTD-1K IoU by 7.86% and decreasing Fa by 17.99.The full model also raises Pd to 99.47% on NUDT-SIRST.
  • Ablation Study: The complete image-specific background prompt achieves the best reported IoU and Fa values across the evaluated datasets.The full template with scene and clutter descriptions reaches IoU of 72.38% and 95.53%, with Fa of 4.10 and 2.64.
  • Ablation Study: Removing or reversing ADBI branches degrades either clutter suppression or target extraction, while removing AFA or either pooling branch causes accuracy losses and artifacts.The complete ADBI assigns text features to dedicated branches, and AFA combines global-context and salient-local features.
  • Efficiency Analysis: ADGNet uses fewer core network parameters than SAIST, with 4.61 M versus 6.29 M, and fewer overall parameters, with 42.44 M versus 389.57 M.The comparison separately reports pre-trained model and core network parameters for multimodal methods.

6 Conclusion

The paper concludes that ADGNet addresses infrared semantic asymmetry through asymmetric prompts, dedicated branch interactions, adaptive fusion, and multimodal annotations. Experiments report state-of-the-art performance with fewer false alarms and missed detections.

  • Conclusion: ADGNet separates fixed target and detailed background descriptions to address the semantic asymmetry of infrared images.The design aims to avoid target noise while supplying priors for background suppression.
  • Conclusion: ADBI assigns corresponding text priors to independent branches, while AFA dynamically fuses their features for target localization and clutter suppression.The conclusion describes these modules as resolving feature optimization conflicts and filtering clutter.
  • Conclusion: The multimodal AITIR dataset adds asymmetric text annotations to three public infrared datasets.The datasets are IRSTD-1K, NUDT-SIRST, and SIRST.
  • Conclusion: Extensive experiments report that ADGNet achieves state-of-the-art performance while reducing severe false alarms and missed detections.
Loading 2609.00853v1…