$\textit{A Contrario}$ Paradigm for YOLO-based Infrared Small Target Detection

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

2024-02-04

Abstract:Detecting small to tiny targets in infrared images is a challenging task in computer vision, especially when it comes to differentiating these targets from noisy or textured backgrounds. Traditional object detection methods such as YOLO struggle to detect tiny objects compared to segmentation neural networks, resulting in weaker performance when detecting small targets. To reduce the number of false alarms while maintaining a high detection rate, we introduce an $\textit{a contrario}$ decision criterion into the training of a YOLO detector. The latter takes advantage of the $\textit{unexpectedness}$ of small targets to discriminate them from complex backgrounds. Adding this statistical criterion to a YOLOv7-tiny bridges the performance gap between state-of-the-art segmentation methods for infrared small target detection and object detection networks. It also significantly increases the robustness of YOLO towards few-shot settings.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

This paper aims to address the problem of small target detection in infrared images (IRSTD), particularly the challenge of distinguishing these small targets from complex backgrounds. Traditional object detection methods (such as YOLO) perform poorly when detecting extremely small objects, especially when faced with noisy or texture-rich backgrounds, leading to false positives. To reduce the number of false positives while maintaining a high detection rate, the paper proposes a new YOLO detection head—OL-NFA (Object-Level Number of False Alarms). This method introduces an anomaly decision criterion to train the YOLO detector, distinguishing small targets from complex backgrounds by leveraging their unexpectedness. This approach not only improves YOLO's performance in small target detection but also significantly enhances its robustness in few-shot scenarios.

Specifically, the OL-NFA module achieves this goal by re-estimating the objectness scores generated by the YOLO backbone network and allowing backpropagation during training. Experimental results show that adding OL-NFA to YOLOv7-tiny significantly improves YOLO's performance on the NUAA-SIRST dataset and bridges the gap with current state-of-the-art segmentation networks. Moreover, this method performs exceptionally well in few-shot settings, maintaining high detection accuracy even when the number of training samples is significantly reduced.