Zero-Shot Infrared Object Detection via Frequency-Domain Feature Decoupling of Visible Images

LI Jiale; HUI Bingwei; WANG Kaiyue; ZHANG Jun

doi:10.3788/gzxb20265508.0810001

Chinese

您当前的位置：

首页 >

文章列表页 >

Zero-Shot Infrared Object Detection via Frequency-Domain Feature Decoupling of Visible Images

更新时间：2026-07-15

- Zero-Shot Infrared Object Detection via Frequency-Domain Feature Decoupling of Visible Images
- Acta Photonica Sinica Pages: 1-18(2026)
- 作者机构：
  
  中国人民解放军国防科技大学电子科学学院，长沙 410028
- 作者简介：
- 基金信息：
- DOI：10.3788/gzxb20265508.0810001
  CLC： TP391.4
- CSTR：32255.14.gzxb20265508.0810001
- Revised：2026-06-25，
  
  Accepted：30 June 2026，
- 稿件说明：
移动端阅览
LI Jiale, HUI Bingwei, WANG Kaiyue, et al. Zero-Shot Infrared Object Detection via Frequency-Domain Feature Decoupling of Visible Images[J/OL]. Acta Photonica Sinica, 2026, gz26-0175
DOI：

LI Jiale, HUI Bingwei, WANG Kaiyue, et al. Zero-Shot Infrared Object Detection via Frequency-Domain Feature Decoupling of Visible Images[J/OL]. Acta Photonica Sinica, 2026, gz26-0175 DOI： 10.3788/gzxb20265508.0810001. CSTR： 32255.14.gzxb20265508.0810001.

摘要

针对典型应用场景下红外图像训练样本匮乏导致的零样本检测难题，提出一种基于可见光图像频域特征解耦的零样本红外目标检测方法。首先，设计频域处理模块，通过削弱高频纹理以降低源域特有风格干扰，并对低频分量实施扰动以提升源域样本多样性，从而为跨模态域不变特征学习提供先验约束。其次，构建权重共享的双分支特征提取框架，并在浅层嵌入特征解耦模块，以显式分离域不变特征与域私有特征，增强模型对跨波段共性表征的建模能力。最后，在颈部网络中引入注意力引导的特征拼接策略，优化多尺度信息融合过程，提升小目标感知与定位性能。在DroneVehicle和FLIR公开数据集上的实验结果表明，本方法与最新的单域泛化目标检测算法相比，mAP@50分别提升了2.9%和5%，验证了所提方法在跨波段零目标域样本条件下的有效性与鲁棒性，为单域泛化目标检测提供了新的解决思路。

Abstract

Infrared imaging systems are widely used in battlefield reconnaissance， disaster warning， autonomous driving， and industrial security because of their all-weather sensing capability and robustness to illumination variations. However， in non-cooperative environments， infrared samples and annotations are often difficult or costly to obtain， making zero-shot infrared object detection highly challenging. Traditional domain adaptation and fine-tuning methods usually require target-domain samples， which limits their applicability in infrared scenarios. Single-domain generalization （SDG）， which learns from only one source domain and does not rely on target-domain data， provides a feasible solution. Existing SDG methods mainly focus on generic visual feature generalization or cross-modal semantic alignment， but they rarely exploit the frequency characteristics of infrared imaging and are often insufficient for small-target detection under cross-modal domain shifts. Moreover， few SDG methods are specifically designed for the lightweight YOLO architecture required by real-time applications.To address the above challenges， this paper proposes a zero-shot infrared object detection method based on visible-image frequency-domain feature decoupling. First， a High-Frequency Suppression and Low-Frequency Enhancement （HSLE） module is designed to process visible images in the frequency domain. High-frequency components are adaptively suppressed to reduce source-domain-specific texture interference， while low-frequency statistics are perturbed to increase source-domain diversity and to narrow the visible-to-infrared gap without severely damaging semantic structure. During training， a weight-shared dual-branch backbone is adopted， where the original visible image is used as the primary branch and the HSLE-enhanced image is used as the auxiliary branch. A Feature Decoupling （FD） module is embedded in shallow backbone stages to separate domain-invariant features from domain-private features， so that deeper layers focus more on cross-spectral common semantics rather than source-style cues. To reduce the loss of weak target information， conventional downsampling is replaced by Space-to-Depth Convolution （SPD_Conv）， which preserves more fine-grained details during feature transformation. In addition， a Guided Attention Concatenation （GA_Concat） module is introduced in the neck to enhance multi-scale feature fusion by using shallow detail-rich features to guide deep semantic aggregation. During inference， the auxiliary branch is removed， and the detector remains a single-branch structure， thereby preserving deployment efficiency.Experiments are conducted on the DroneVehicle dataset under a strict zero-shot visible-to-infrared protocol， where only labeled visible images are used for training and infrared images are used only for validation. To ensure a rigorous evaluation of scene-level generalization， the visible training set and infrared validation set are constructed with scene-level non-overlap. On the infrared validation set， the proposed method achieves an mAP@50 of 87.2%， which is 17.4 percentage points higher than YOLOv8n. Compared with SW， IBN-Net， ISW， CDSD， G-NAS， OA-DG， and PDOC， the proposed method improves mAP@50 by 6.8， 10.7， 13.7， 4.6， 8.6， 2.9， and 7.2 percentage points， respectively. Meanwhile， the detector remains lightweight， with only 3.7 M parameters and 14.1 GFLOPs.To further evaluate multi-view infrared generalization， experiments are also conducted on the FLIR dataset. The proposed method achieves an mAP@50 of 72.1%， outperforming YOLOv8n by 17.5 percentage points and surpassing SW， IBN-Net， ISW， CDSD， G-NAS， OA-DG， and PDOC by 16.1， 13.6， 12.5， 11.0， 6.3， 5.0， and 5.8 percentage points， respectively. Ablation studies verify the effectiveness of HSLE， FD， and the small-object detection components. Qualitative results further show that the proposed method reduces missed detections and false alarms in cluttered infrared scenes. These results demonstrate that the proposed method improves zero-shot visible-to-infrared object detection and small infrared target detection while maintaining the efficiency of a lightweight one-stage detector.

关键词

Keywords

references

ZHENG Guangtao ， HUAI Mengdi ， ZHANG Aidong ， et al . AdvST： revisiting data augmentations for single domain generalization ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2024 ， 38 （ 19 ）： 21832 - 21840 .

SU Zixian ， YAO Kai ， YANG Xi ， et al . Rethinking data augmentation for single-source domain generalization in medical image segmentation ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2023 ， 37 （ 2 ）： 2366 - 2374 .

SHI Caijuan ， ZHENG Yuanfan ， REN Bijuan ， et al . Single-domain generalized breast tumor detection in X-ray images ［J］. Journal of Image and Graphics ， 2024 ， 29 （ 3 ）： 725 - 740 .

史彩娟，郑远帆，任弼娟，等 . 单域泛化X-ray乳腺肿瘤检测［J］. 中国图象图形学报， 2024 ， 29 （ 3 ）： 725 - 740 .

ZHANG Chen ， LI Yunping ， TANG Xin ， et al . Review of domain generalization person re-identification based on deep learning ［J］. Journal of Kunming University of Science and Technology （Natural Science）， 2024 ， 49 （ 6 ）： 86 - 99 .

张臣，李云平，唐鑫，等 . 基于深度学习的域泛化行人重识别综述［J］. 昆明理工大学学报（自然科学版）， 2024 ， 49 （ 6 ）： 86 - 99 .

JIN Xin ， LAN Cuiling ， ZENG Wenjun ， et al . Style normalization and restitution for generalizable person re-identification ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2020 ： 3143 - 3152 .

ZHENG Kecheng ， LIU Jiawei ， WU Wei ， et al . Calibrated feature decomposition for generalizable person re-identification ［J/OL］. arXiv ： 2111 . 13945 ，（ 2021-11-27 ）［ 2026-06-24 ］. https：//doi.org/10.48550/arXiv.2111.13945 https://doi.org/10.48550/arXiv.2111.13945 .

LIU Jiawei ， HUANG Zhipeng ， LI Liang ， et al . Debiased batch normalization via gaussian process for generalizable person re-identification ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2022 ， 36 （ 2 ）： 1729 - 1737 .

WU Aming ， DENG Cheng . Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2022 ： 847 - 856 .

LIN Chuang ， YUAN Zehuan ， ZHAO Sicheng ， et al . Domain-invariant disentangled network for generalizable object detection ［C］. Proceedings of the IEEE International Conference on Computer Vision . Piscataway ： IEEE ， 2021 ： 8771 - 8780 .

YU Shijie ， ZHU Feng ， CHEN Dapeng ， et al . Multiple domain experts collaborative learning： multi-source domain generalization for person re-identification ［J/OL］. arXiv ： 2105 . 12355 ，（ 2021-05-26 ）［ 2026-06-24 ］. https：//doi.org/10.48550/arXiv.2105.12355 https://doi.org/10.48550/arXiv.2105.12355 .

LI Da ， YANG Yongxin ， SONG Yizhe ， et al . Learning to generalize： meta-learning for domain generalization ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2018 ， 32 （ 1 ）： 3490 - 3497 .

ZHAO Yuyang ， ZHONG Zhun ， YANG Fengxiang ， et al . Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2021 ： 6273 - 6282 .

ZHAO Yuyang ， ZHONG Zhun ， ZHAO Na ， et al . Style-hallucinated dual consistency learning： a unified framework for visual domain generalization ［J］. International Journal of Computer Vision ， 2024 ， 132 （ 3 ）： 837 - 853 .

QI Lei ， DONG Peng ， XIONG Tan ， et al . DoubleAUG： single-domain generalized object detector in urban via color perturbation and dual-style memory ［J］. ACM Transactions on Multimedia Computing， Communications， and Applications ， 2024 ， 20 （ 5 ）： 1 - 20 .

FAN Qi ， SEGU M ， TAI Y W ， et al . Towards robust object detection invariant to real-world domain shifts ［C］. The Eleventh International Conference on Learning Representations . Kigali ： OpenReview ， 2023 .

LIAO Shengcai ， SHAO Ling . Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting ［J/OL］. arXiv ： 1904 . 10424 ，（ 2019-04-23 ）［ 2026-06-24 ］. https：//doi.org/10.48550/arXiv.1904.10424 https://doi.org/10.48550/arXiv.1904.10424 .

VIDIT V ， ENGILBERGE M ， SALZMANN M . CLIP the gap： a single domain generalization approach for object detection ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2023 ： 3219 - 3229 .

DENG Li ， WU Aming ， WANG Yaowei ， et al . Prompt-driven dynamic object-centric learning for single domain generalization ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2024 ： 17606 - 17615 .

LI Hao ， WANG Wei ， WANG Cong ， et al . Phrase grounding-based style transfer for single-domain generalized object detection ［J］. IEEE Transactions on Circuits and Systems for Video Technology ， 2026 ， 36 （ 1 ）： 106 - 118 .

SUNKARA R ， LUO Tie . No more strided convolutions or pooling： a new CNN building block for low-resolution images and small objects ［C］. Machine Learning and Knowledge Discovery in Databases . Cham ： Springer Nature Switzerland ， 2023 ： 443 - 459 .

TONG Yunfei ， LIU Jing ， FU Zhiling ， et al . Guided attention and joint loss for infrared dim small target detection ［J］. IEEE Transactions on Geoscience and Remote Sensing ， 2024 ， 62 ： 1 - 14 .

HWANG S ， HAN D ， JEON M . DG-DETR： toward domain generalized detection transformer ［J］. Pattern Recognition Letters ， 2026 ， 199 ： 128 - 134 .

JEON S ， HONG K ， LEE P ， et al . Feature stylization and domain-aware contrastive learning for domain generalization ［C］. Proceedings of the 29th ACM International Conference on Multimedia . New York ： Association for Computing Machinery ， 2021 ： 22 - 31 .

GUO Jintao ， WANG Na ， QI Lei ， et al . ALOFT： a lightweight MLP-like architecture with dynamic low-frequency transform for domain generalization ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2023 ： 24132 - 24141 .

YOO J ， UH Y ， CHUN S ， et al . Photorealistic style transfer via wavelet transforms ［C］. Proceedings of the IEEE/CVF International Conference on Computer Vision . Piscataway ： IEEE ， 2019 ： 9035 - 9044 .

YUAN Yuxuan ， TANG Luyao ， XU Ying ， et al . Filling and disentanglement： toward low- and high-order parallel single-domain generalization for SAR ship detection ［J］. IEEE Transactions on Aerospace and Electronic Systems ， 2025 ， 61 （ 2 ）： 3668 - 3682 .

HOU Qibin ， ZHOU Daquan ， FENG Jiashi . Coordinate attention for efficient mobile network design ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2021 ： 13713 - 13722 .

HU Jie ， SHEN Li ， SUN Gang . Squeeze-and-excitation networks ［C］. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2018 ： 7132 - 7141 .

PAN Xingang ， ZHAN Xiaohang ， SHI Jianping ， et al . Switchable whitening for deep representation learning ［C］. Proceedings of the IEEE/CVF International Conference on Computer Vision . Piscataway ： IEEE ， 2019 ： 1863 - 1871 .

PAN Xingang ， LUO Ping ， SHI Jianping ， et al . Two at once： enhancing learning and generalization capacities via IBN-Net ［C］. Proceedings of the European Conference on Computer Vision . Cham ： Springer ， 2018 ： 464 - 479 .

CHOI S ， JUNG S ， YUN H ， et al . RobustNet： improving domain generalization in urban-scene segmentation via instance selective whitening ［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway ： IEEE ， 2021 ： 11575 - 11585 .

WU Fan ， GAO Jinling ， HONG Lanqing ， et al . G-NAS： generalizable neural architecture search for single domain generalization object detection ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2024 ， 38 （ 6 ）： 5958 - 5966 .

LEE W ， HONG D ， LIM H ， et al . Object-aware domain generalization for object detection ［C］. Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto ： AAAI Press ， 2024 ， 38 （ 4 ）： 2947 - 2955 .

Views

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

Inverse-designed Polarization-insensitive 2×2 Waveguide Crossings

A Quality-gating-based Method for Optimizing Coherent Peak Localization in White-light Interferometry

Broadband Wavelength Trimming of Silicon Nitride Microring Resonators Based on Non-volatile Sb₂S₃ Phase Change

Design of Airglow Two-dimensional Temperature Imager and Influence of the Filter Parameters

Liquid Crystal Metamaterial Absorber Based on Grating Coding Regulation

Related Author

No data

Related Institution

No data

AI问答

Postal code：100079
Tel：（010）53879206 Email：tmw@bjxintong.com.cn
Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09082226号-64 京公网安备11010602201714号
It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.

⁰