Anomaly-Aware YOLO: A Frugal yet Robust Approach to Infrared Small Target Detection

Alina Ciocarlan, Sylvie Le Hégarat-Mascle, Sidonie Lefebvre

Abstract

Infrared Small Target Detection (IRSTD) is a critical task for applications like drone tracking and remote sensing, characterized by targets that occupy minimal pixels against complex backgrounds. Traditional segmentation-based approaches are often computationally heavy and sensitive to inconsistent pixel-level annotations. We present AA-YOLO, which redefines IRSTD as a statistical anomaly detection problem. By replacing the standard YOLO objectness head with an Anomaly-Aware Detection Head (AADH), we achieve state-of-the-art results with a fraction of the computational cost. AA-YOLO is remarkably data-frugal and robust to sensor noise, offering a practical solution for resource-constrained deployment.

Introduction

Detecting small targets in infrared (IR) imagery is a "needle in a haystack" challenge. These targets often lack distinct shape or texture, appearing as faint blobs buried in cluttered backgrounds.

Examples of small targets in infrared imagery, illustrating the low contrast and minimal pixel footprint.

Current state-of-the-art (SOTA) methods typically employ heavy segmentation networks. However, these models face significant hurdles: annotation subjectivity leads to inconsistent training signals, high model complexity hinders deployment on edge devices like drones, and they often suffer from object fragmentation or adjacency issues.

The AA-YOLO Framework

AA-YOLO (Anomaly-Aware YOLO) addresses these challenges by treating the background as a null hypothesis ($H_0$) and targets as statistical anomalies. Instead of a bespoke architecture, AA-YOLO utilizes standard YOLO backbones (v5, v7, or v9) and introduces a mathematically grounded detection head.

The AA-YOLO framework, highlighting the integration of the Anomaly-Aware Detection Head (AADH) into the standard YOLO pipeline.

Anomaly-Aware Detection Head (AADH)

The core innovation is the AADH, which replaces the standard objectness head. It models background activations in the latent feature space using an exponential distribution—the most unbiased choice for non-negative variables with a fixed mean, according to the maximum entropy principle.

The head performs a statistical test where the sum of activations across feature channels ($\mu_2$) is evaluated. Under the null hypothesis, this sum follows an Erlang distribution. The model calculates a significance score based on the p-value, effectively flagging pixels that deviate significantly from the background distribution. This head is trained end-to-end using Mean Squared Error (MSE) loss, encouraging the network to learn features that maximize this statistical separation.

Experimental Evaluation

The approach was validated on three major benchmarks: SIRST, IRSTD-1k, and VEDAI. These datasets cover a range of scenarios from classic IR small targets to aerial remote sensing of vehicles.

Key Results and Insights

  • Efficiency and Performance: AA-YOLOv7-tiny outperformed the heavy-duty EFLNet while using 6x fewer parameters and 7x fewer GFLOPs. On IRSTD-1k, it achieved an F1 score of 92.5 and an Average Precision (AP) of 96.6.
  • Data Frugality: The model is exceptionally robust to limited data. When trained on only 10% of the SIRST dataset (approx. 25 images), it retained 90% of its full-scale performance.
  • Noise Robustness: In the presence of Gaussian noise, AA-YOLO maintained superior stability, outperforming baseline models by over 4 points in F1 score.
  • Background-Centric Learning: By focusing on modeling the background rather than specific target appearances, the system suppresses false alarms more effectively in noisy environments.

Conclusion

AA-YOLO demonstrates that IR small target detection does not require massive segmentation models. By leveraging statistical anomaly detection within a lightweight YOLO framework, we achieve a solution that is both high-performing and computationally efficient. Future work will explore more complex probabilistic modeling of channel dependencies and the application of this "anomaly-aware" paradigm to other domains like medical imaging.

Summary updated

BibTeX
@inproceedings{ciocarlan2026anomaly,
  author = {Alina Ciocarlan and Sylvie Le Hégarat-Mascle and Sidonie Lefebvre},
  title = {{Anomaly-Aware YOLO: A Frugal yet Robust Approach to Infrared Small Target Detection}},
  booktitle = {Engineering},
  year = {2026},
  doi = {10.1016/j.engappai.2026.114186},
}
Figure