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

TL;DR

Anomaly-Aware YOLO (AA-YOLO), which integrates a statistical anomaly detection test into its detection head, which achieves competitive performance on several IRSTD benchmarks, but also demonstrates remarkable robustness in scenarios with limited training data, noise, and domain shifts.

Abstract

Infrared Small Target Detection (IRSTD) is a challenging task in defense applications, where complex backgrounds and tiny target sizes often result in numerous false alarms using conventional object detectors. To overcome this limitation, we propose Anomaly-Aware YOLO (AA-YOLO), which integrates a statistical anomaly detection test into its detection head. By treating small targets as unexpected patterns against the background, AA-YOLO effectively controls the false alarm rate. Our approach not only achieves competitive performance on several IRSTD benchmarks, but also demonstrates remarkable robustness in scenarios with limited training data, noise, and domain shifts. Furthermore, since only the detection head is modified, our design is highly generic and has been successfully applied across various YOLO backbones, including lightweight models. It also provides promising results when integrated into an instance segmentation YOLO. This versatility makes AA-YOLO an attractive solution for real-world deployments where resources are constrained. The code will be publicly released.

Figures (10)

• Figure 1: Illustration of the annotation subjectivity in IRSTD. The top-righ corners show in blue the annotation masks provided by dai2021asymmetric for the SIRST dataset.
• Figure 2: Variations of $F_{\mu_2}$ and $-\ln{F_{\mu_2}}$.
• Figure 3: AA-YOLO architecture. The input first passes through a generic YOLO backbone to extract high-level features. Then, the conventional YOLO detection head is replaced by our Anomaly-Aware Detection Head (AADH), which enhances small object features by integrating statistical anomaly testing to estimate objectness scores. For simplicity, we illustrate the application of our AADH on $F_2$-level feature maps only, although in practice AADH is applied to all detection scales: $F_1$, $F_2$ and $F_3$.
• Figure 4: Objectness score maps for different methods.
• Figure 5: Precision and recall curves obtained on the SIRST dataset.