Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust infrared small target detection using self-supervised and a contrario paradigms

Alina Ciocarlan, Sylvie Le Hégarat-Mascle, Sidonie Lefebvre, Arnaud Woiselle

TL;DR

This work tackles infrared small-target detection by jointly leveraging an a contrario paradigm and self-supervised learning (SSL) within a YOLO-based detector. It introduces a pixel-level NFA head, $NFA_{\mathcal{N}}$, integrated into YOLO to suppress background-driven false alarms while preserving small-target responses, and systematically evaluates SSL pretraining strategies for IRSTD. The results show that instance-discrimination SSL methods, particularly when combined with the NFA head and attention mechanisms, yield substantial gains, enabling near state-of-the-art performance on challenging datasets like SIRST and strong gains in frugal settings, although IRSTD-850 with complex backgrounds remains challenging. Overall, the proposed two-pronged approach narrows the gap with segmentation-based SOTA and offers a robust, data-efficient solution for IRSTD with potential extensions to Vision Transformer architectures.

Abstract

Detecting small targets in infrared images poses significant challenges in defense applications due to the presence of complex backgrounds and the small size of the targets. Traditional object detection methods often struggle to balance high detection rates with low false alarm rates, especially when dealing with small objects. In this paper, we introduce a novel approach that combines a contrario paradigm with Self-Supervised Learning (SSL) to improve Infrared Small Target Detection (IRSTD). On the one hand, the integration of an a contrario criterion into a YOLO detection head enhances feature map responses for small and unexpected objects while effectively controlling false alarms. On the other hand, we explore SSL techniques to overcome the challenges of limited annotated data, common in IRSTD tasks. Specifically, we benchmark several representative SSL strategies for their effectiveness in improving small object detection performance. Our findings show that instance discrimination methods outperform masked image modeling strategies when applied to YOLO-based small object detection. Moreover, the combination of the a contrario and SSL paradigms leads to significant performance improvements, narrowing the gap with state-of-the-art segmentation methods and even outperforming them in frugal settings. This two-pronged approach offers a robust solution for improving IRSTD performance, particularly under challenging conditions.

Robust infrared small target detection using self-supervised and a contrario paradigms

TL;DR

This work tackles infrared small-target detection by jointly leveraging an a contrario paradigm and self-supervised learning (SSL) within a YOLO-based detector. It introduces a pixel-level NFA head, , integrated into YOLO to suppress background-driven false alarms while preserving small-target responses, and systematically evaluates SSL pretraining strategies for IRSTD. The results show that instance-discrimination SSL methods, particularly when combined with the NFA head and attention mechanisms, yield substantial gains, enabling near state-of-the-art performance on challenging datasets like SIRST and strong gains in frugal settings, although IRSTD-850 with complex backgrounds remains challenging. Overall, the proposed two-pronged approach narrows the gap with segmentation-based SOTA and offers a robust, data-efficient solution for IRSTD with potential extensions to Vision Transformer architectures.

Abstract

Detecting small targets in infrared images poses significant challenges in defense applications due to the presence of complex backgrounds and the small size of the targets. Traditional object detection methods often struggle to balance high detection rates with low false alarm rates, especially when dealing with small objects. In this paper, we introduce a novel approach that combines a contrario paradigm with Self-Supervised Learning (SSL) to improve Infrared Small Target Detection (IRSTD). On the one hand, the integration of an a contrario criterion into a YOLO detection head enhances feature map responses for small and unexpected objects while effectively controlling false alarms. On the other hand, we explore SSL techniques to overcome the challenges of limited annotated data, common in IRSTD tasks. Specifically, we benchmark several representative SSL strategies for their effectiveness in improving small object detection performance. Our findings show that instance discrimination methods outperform masked image modeling strategies when applied to YOLO-based small object detection. Moreover, the combination of the a contrario and SSL paradigms leads to significant performance improvements, narrowing the gap with state-of-the-art segmentation methods and even outperforming them in frugal settings. This two-pronged approach offers a robust solution for improving IRSTD performance, particularly under challenging conditions.

Paper Structure

This paper contains 15 sections, 1 equation, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. A contrario paradigm
  4. Self-supervised learning and object detection
  5. Proposed method
  6. Integrating an a contrario criterion into YOLO
  7. Choosing appropriate SSL initialisation
  8. Experiments
  9. Experimental set-up
  10. Datasets and evaluation metrics
  11. Baselines
  12. Mixing both SSL and a contrario paradigms improves the baselines
  13. What about frugal setting?
  14. Comparison with SOTA segmentation networks
  15. Conclusion

Figures (2)

  • Figure 1: Integration of our pixel-level criterion into a YOLO framework, through the $\textsc{NFA}_{\mathcal{N}}$ detection head. This module can be added on top of any YOLO.
  • Figure 2: Qualitative results obtained with YOLO-R50 and YOLO-R50+$\textsc{NFA}_{\mathcal{N}}$ initialized with ReSim weights on challenging scenes from SIRST and IRSTD-850 datasets. True positives, false positives and missed detections are circled in green, red and yellow lines, respectively.