Towards Robust Infrared Small Target Detection: A Feature-Enhanced and Sensitivity-Tunable Framework

TL;DR

This work tackles infrared small target detection by introducing FEST, a framework that enhances robustness without changing backbones. It combines a feature-enhanced multi-scale fusion with an edge-focused EEDM loss and an adjustable-sensitivity post-processing to adapt to diverse scenes, achieving notable gains across multiple SIRST networks and datasets. The approach delivers both accuracy benefits (via EEDM and multi-scale fusion) and improved recall in challenging conditions (via AS), while offering a lightweight variant (LW-FEST) for efficiency-constrained applications. Overall, FEST advances practical SIRST detection by marrying edge-aware training, multi-scale integration, and probabilistic mask exploitation, with demonstrated cross-dataset generalization.

Abstract

Recently, single-frame infrared small target (SIRST) detection technology has attracted widespread attention. Different from most existing deep learning-based methods that focus on improving network architectures, we propose a feature-enhanced and sensitivity-tunable (FEST) framework, which is compatible with existing SIRST detection networks and further enhances their detection performance. The FEST framework improves the model's robustness from two aspects: feature enhancement and target confidence regulation. For feature enhancement, we employ a multi-scale fusion strategy to improve the model's perception to multi-scale features of multi-size targets, and design an edge enhancement difficulty mining (EEDM) loss to guide the network to continuously focus on challenging target regions and edge features during training. For target confidence regulation, an adjustable sensitivity (AS) strategy is proposed for network post-processing. This strategy enhances the model's adaptability in complex scenarios and significantly improves the detection rate of infrared small targets while maintaining segmentation accuracy. Extensive experimental results show that our FEST framework can effectively enhance the performance of existing SIRST detection networks. The code is available at https://github.com/YuChuang1205/FEST-Framework

Figures (12)

• Figure 1: Overall structure of the FEST framework. "SIRST Network" denotes the selected existing SIRST networks. Size $O_1$, ..., Size $O_{n-1}$, and Size $O_n$ denote the actual sizes of the original training images. Size 1, Size 2, and Size 3 denote the input resolutions corresponding to each model. Size $O_{t1}$, ..., Size $O_{t(m-1)}$, and Size $O_{tm}$ denote the actual sizes of the original test images.
• Figure 2: Illustration of Edge enhancement difficulty mining (EEDM) loss.
• Figure 3: Adjustable sensitivity strategy. Each connected component in the figure denotes an infrared small target. Green denotes correctly included weak targets, while red denotes incorrectly included weak targets.
• Figure 4: Some samples from multiple datasets.
• Figure 5: ROC curves of models trained with different loss functions.