• Home
  • Search
  • Enhanced Generalization Ability of Infrared Small Target Detection via Perturbed Training and Adaptive Test
  • Cite Icon3
  • https://doi.org/10.1109/tgrs.2024.3486751Copy DOI Icon

Enhanced Generalization Ability of Infrared Small Target Detection via Perturbed Training and Adaptive Test

Show More
  • Abstract
  • Literature Map
  • References
  • Citations
  • Similar Papers
Abstract

While deep learning (DL) methods based on empirical risk minimization (ERM) have achieved superior detection performance on public infrared small target detection (IRSTD) datasets, the learned models suffer severe performance degradation when dealing with out-of-distribution (OOD) data. The primary challenge in practical infrared systems lies in enhancing the generalization ability to ensure reliable detection performance in diverse scenarios. In this article, we propose a single-source domain generalization (SDG) and test-time adaptation (TTA) approach for IRSTD, aiming to train a domain-agnostic DL model in the presence of only one source domain and enabling it to perform well on any unseen target domain. During the training process, perturbed images and networks (PINs) are utilized to enhance the diversity of features extracted from a single source domain with limited samples. A Y-shaped multibranch network employs a combination of the self-supervised recovery task and the supervised detection task. Their losses jointly guide the network to converge explicitly to the flat minima in the loss landscape. During the test process, we apply domain-guided adaptation (DGA) to the learned model. The feature statistics of the dynamically inputted image batches are considered as cues to the target domain. Batch normalization (BN) layers are adaptively fine-tuned online based on the differences in feature statistics between the source model and target images. We conducted extensive ablation studies to demonstrate the rationality and effectiveness of each component of the framework. Compared to other SDG, TTA, and model-driven IRSTD methods, PIN-DGA can more effectively improve the detection performance of DL methods on unseen target domains. As a model-agnostic framework, we verified its compatibility with current state-of-the-art networks. On three OOD datasets, our method can improve the average probability of detection (PD) by 16.08 and reduce false alarm targets (FATs) by 33%. The codes are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jzchenriver/PIN-DGA</uri>.

Similar Papers
  • Research Article
  • Citations4

Phrase Grounding-Based Style Transfer for Single-Domain Generalized Object Detection

  • Jan 01, 2026
  • IEEE Transactions on Circuits and Systems for Video Technology
  • Hao Li +8
  • Research Article
  • Citations1

Single-Domain Generalization via Path Flatness-Aware Optimization of Loss Landscapes.

  • Mar 01, 2026
  • IEEE transactions on neural networks and learning systems
  • Zizhou Wang +7
  • Conference Article
  • Citations97

Generalizable Cross-modality Medical Image Segmentation via Style Augmentation and Dual Normalization

  • Jun 01, 2022
  • Ziqi Zhou +4
  • Conference Article
  • Citations22

Continual Unsupervised Domain Adaptation for Semantic Segmentation by Online Frequency Domain Style Transfer

  • Sep 19, 2021
  • Jan-Aike Termohlen +4
  • Research Article
  • Citations26

Prior-Guided Data Augmentation for Infrared Small Target Detection

  • Jan 01, 2022
  • IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
  • Ao Wang +5
  • Research Article
  • Citations43

A novel adaptive deep transfer learning method towards thermal error modeling of electric spindles under variable conditions

  • Mar 14, 2024
  • Journal of Manufacturing Systems
  • Shuai Ma +6
  • Conference Article
  • Citations140

Generalizable Person Re-identification with Relevance-aware Mixture of Experts

  • Jun 01, 2021
  • Yongxing Dai +4
  • Research Article
  • Citations32

Heterogeneous Multidomain Recommender System Through Adversarial Learning.

  • Nov 01, 2023
  • IEEE Transactions on Neural Networks and Learning Systems
  • Wenhui Liao +4
  • Research Article
  • Citations7

Adaptive Domain Generalization Via Online Disagreement Minimization.

  • Jan 01, 2023
  • IEEE Transactions on Image Processing
  • Xin Zhang +1
  • Research Article
  • Citations77

Multi-Source Contribution Learning for Domain Adaptation.

  • Oct 01, 2022
  • IEEE Transactions on Neural Networks and Learning Systems
  • Keqiuyin Li +3
  • Research Article
  • Citations3

Learning Source-Free Domain Adaptation for Infrared Small Target Detection

  • Jan 01, 2025
  • IEEE Signal Processing Letters
  • Hongxu Jin +4
  • Research Article
  • Citations277

Deep Unsupervised Domain Adaptation: A Review of Recent Advances and Perspectives

  • Aug 16, 2022
  • APSIPA Transactions on Signal and Information Processing
  • Xiaofeng Liu +6
  • Conference Article
  • Citations30

Dynamic Momentum Adaptation for Zero-Shot Cross-Domain Crowd Counting

  • Oct 17, 2021
  • Qiangqiang Wu +2
  • Research Article

Projection Reconstruction Based Domain-Adaptive Dictionary Pair Learning Method

  • Jul 01, 2025
  • Journal of Computer-Aided Design &amp; Computer Graphics
  • Guohua Zhou +5
  • Research Article

Invariant Content Representation for Generalizable Medical Image Segmentation.

  • May 17, 2024
  • Journal of imaging informatics in medicine
  • Zhiming Cheng +4
Cactus Communications logo

Copyright 2026 Cactus Communications. All rights reserved.