You are currently viewing a new version of our website. To view the old version click .
Entropy
  • This is an early access version, the complete PDF, HTML, and XML versions will be available soon.
  • Article
  • Open Access

28 November 2025

Pseudo-Sample Generation and Self-Supervised Framework for Infrared Dim and Small Target Detection

,
,
,
,
,
and
School of Electronic and Information Engineering, Changchun University of Science and Technology, Changchun 130022, China
*
Author to whom correspondence should be addressed.
This article belongs to the Section Information Theory, Probability and Statistics

Abstract

Infrared dim and small target detection is crucial for long-range sensing. However, its deep representation learning is severely constrained by the scarcity of accurately annotated real data, and related research remains underdeveloped. Existing data generation methods based on patch synthesis or geometric transformations fail to incorporate the physical degradation mechanisms of infrared imaging systems and reasonable environmental constraints, leading to significant discrepancies between synthetic data and real-world scenarios. To address this issue, this paper proposes a novel pseudo-sample generation paradigm based on physics-informed degradation modeling and high-order constraints. First, we construct an infrared image degradation model that decouples the degradation processes of targets and backgrounds at the signal level, achieving accurate modeling of real infrared imaging while ensuring the reliability of the degradation process through information fidelity optimization. Second, an online grid-based high-order constraint strategy is designed, which synergistically integrates global semantic, local structural, and grayscale constraints based on statistical distribution consistency to generate a high-fidelity infrared simulation dataset. Finally, we build a complete self-supervised detection framework incorporating classical neural networks, customized loss functions, and two-dimensional information evaluation metrics. Extensive experiments demonstrate that the synthetic data generated by our method significantly outperforms existing simulated datasets on authenticity metrics. It also effectively enhances the generalization performance of various detectors in real-world scenarios, achieving detection accuracy superior to baseline models trained on traditional simulated data.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.