Abstract
Single-frame infrared small target detection (ISTD) is a critical yet challenging task characterized by extreme class imbalance, high sensitivity in bounding box regression, and the potential for target information loss in deep semantic layers. This paper presents the Enhancing Feature Learning Network (EFLNet), a framework specifically engineered to mitigate these issues. EFLNet introduces three primary innovations: an Adaptive Threshold Focal Loss (ATFL) to balance target-background attention, the Normalized Gaussian Wasserstein Distance (NWD) to stabilize regression for tiny objects, and a Dynamic Head mechanism for multi-scale feature integration. Experimental results across multiple benchmarks demonstrate that EFLNet outperforms existing state-of-the-art (SOTA) methods in detection accuracy and robustness.
Introduction
The detection of small targets in infrared (IR) imagery—such as distant aircraft or maritime vessels—is often described as finding a "needle in a haystack." These targets typically occupy a minuscule fraction of the total pixels and are frequently embedded in high-clutter environments involving clouds, sea-glint, or sensor noise.
Traditional approaches often treat ISTD as a semantic segmentation problem. However, segmentation-based methods can be computationally intensive and may not always provide the localization precision required for real-time tracking. EFLNet approaches ISTD as an object detection task, addressing the inherent mathematical instabilities that arise when applying standard detection frameworks to extremely small objects.
Core Challenges in Infrared Detection
The transition from segmentation to detection introduces three significant hurdles:
- Extreme Class Imbalance: Because targets are so small, the background dominates the feature space, causing standard models to neglect target-specific features.
- IoU Sensitivity: The standard Intersection over Union (IoU) metric is highly sensitive to small displacements. For a target only a few pixels wide, a minor prediction error can result in a zero IoU, providing no gradient for the model to learn from.
- Feature Dissipation: As data passes through successive downsampling layers in a deep neural network, the high-level semantic layers often lose the fine-grained spatial information necessary to identify a tiny target.
The EFLNet Framework
To overcome these obstacles, EFLNet incorporates three specialized components:
1. Adaptive Threshold Focal Loss (ATFL)
Standard Focal Loss treats all samples on a single continuum. ATFL improves upon this by "decoupling" the target and background. It utilizes a probability threshold (typically 0.5) to distinguish between easy background samples and difficult targets. An adaptive mechanism then adjusts the loss weights, forcing the network to allocate more attention to learning the features of dim or obscured targets that would otherwise be ignored.
2. Normalized Gaussian Wasserstein Distance (NWD)
To solve the "brittleness" of IoU, the authors replace it with NWD for bounding box regression. By modeling bounding boxes as 2D Gaussian distributions, the similarity between the predicted box and the ground truth is calculated via the Wasserstein distance. This approach creates a smooth, continuous gradient even when there is no physical overlap between boxes, significantly improving convergence during training for small targets.
3. Dynamic Head Mechanism
To prevent the loss of target information in deep layers, EFLNet employs a Dynamic Head. This component applies three levels of attention:
- Scale-aware: Determines the relative importance of different feature pyramid layers.
- Spatial-aware: Focuses the model on specific regions of interest within the frame.
- Task-aware: Dynamically weights different feature channels to optimize for detection.
Experimental Evaluation
The researchers evaluated EFLNet on three benchmark datasets: NUAA-SIRST, NUDT-SIRST, and IRSTD-1k. To facilitate this research, the authors provided bounding box annotations for these datasets, which were previously limited to pixel-level masks.
EFLNet demonstrated superior performance compared to SOTA methods such as MDvsFA, ACMNet, and DNANet. On the NUDT-SIRST dataset, EFLNet achieved an F1-score of 0.947, a notable improvement over DNANet’s 0.901. On the IRSTD-1k dataset, the model maintained high precision (0.870) and recall (0.817), proving its robustness in cluttered environments like sea-sky lines and heavy cloud cover.
Conclusion
EFLNet represents a significant advancement in infrared small target detection by shifting the paradigm toward distribution-based regression and adaptive loss weighting. While the inclusion of the Dynamic Head and NWD increases computational complexity (approximately 38M parameters), the gains in detection reliability make it a compelling solution for high-stakes infrared monitoring systems.
The source code and annotated datasets are available at: https://github.com/YangBo0411/infrared-small-target
Summary updated