Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation

Infrared small target detection (IRSTD) in high-resolution images remains challenging due to targets' small size, weak features, and severe interference from complex dynamic backgrounds. This paper proposes an efficient coarse-to-fine infrared small target detection framework called ECFNet, which addresses these issues by integrating a denoising-assisted training strategy and an attention prior-guided knowledge distillation mechanism to enhance performance while maintaining high real-time processing efficiency.

The gist: The proposed ECFNet framework significantly enhances the detection performance of IRSTs under complex backgrounds while efficiently reducing computational costs by leveraging a coarse-to-fine structure with context region proposals, denoising-assisted training, and attention prior-guided knowledge distillation.

How it works

The ECFNet framework is structured into two main stages: a coarse stage and a fine stage, designed to balance accuracy and efficiency. In the coarse stage, the framework utilizes a Region Binary Classification Network (RBCN) to reformulate pixel-wise prediction as a grid-level binary classification task, which significantly reduces the number of predictions required compared to dense prediction methods. This network employs a four-stage hierarchical architecture where spatial resolution is progressively reduced while channel width increases across stages to capture higher-level semantic representations. The output of the RBCN is then processed by an ExpSlicer module, which dynamically expands around predicted target regions to extract complete targets from multi-scale feature maps, ensuring that the target is fully contained within a single region patch while preserving critical contextual cues.

Denoising-Assisted Training (DAT)

To improve the ability of the RBCN to distinguish target proposals from complex backgrounds, a novel denoising-assisted training (DAT) strategy is introduced. This strategy involves injecting target-like noise to corrupt the ground-truth (GT) masks and integrating them into the RBCN feature space. The network is then trained through a denoising task where it must reconstruct the original GT masks through a denoising task. This process encourages the network to explicitly model contextual target–background relationships, which significantly improves its ability to discriminate target region proposals from complex backgrounds. The reconstruction loss, denoted as LDAT, combines binary cross-entropy (BCE) and IoU loss, with a learnable weighting coefficient to balance the objectives.

Attention Prior-Guided Knowledge Distillation (APKD)

In the fine stage, a lightweight target detector is customized for the region proposals generated by the coarse stage. To enhance this detector's discriminative feature representation, an Attention Prior-Guided Knowledge Distillation (APKD) strategy is proposed. This mechanism leverages cross-attention to propagate the teacher’s spatial attention priors over critical target regions to the student. The process involves calculating a global cross-attention between teacher and student features to model semantic dependencies, and then modulating the student’s features with these cross-attention weights, guiding it to focus on critical target regions emphasized by the teacher. This active modulation makes the student aware of semantically meaningful regions, which is especially effective for lightweight detectors.

Key Components and Enhancements

The framework incorporates several key innovations designed to address specific challenges in IRSTD. The RBCN simplifies detection by focusing on grid-level classification, while the ExpSlicer ensures complete target coverage during the coarse stage refinement. The DAT module specifically targets background interference by simulating realistic distractors, forcing the network to exploit surrounding context rather than isolated responses. Furthermore, the APKD module guides the fine detector using teacher attention priors to enhance target awareness and suppress background interference without increasing inference cost. The overall architecture is optimized by customizing a lightweight detector in the fine stage and employing channel-adaptive scaling in the APKD mechanism.

Experimental Validation

Extensive experiments on three real infrared datasets—UAV, Car, and IRSTD-1k—demonstrate that ECFNet outperforms existing single-stage and two-stage approaches while maintaining high real-time processing efficiency. The quantitative results show that ECFNet achieves the highest values for Recall (R) and AP50 compared to other state-of-the-art (SOTA) methods. Specifically, the method achieves an approximately 60% reduction in computational cost under comparable detection performance and yields significant improvements in P and R, confirming its suitability for embedded deployment. The ablation studies confirm that incorporating both DAT and APKD leads to the best overall performance gains, highlighting their complementary roles in enhancing target discrimination across the coarse and fine stages. Additionally, experiments show that ECFNet maintains high precision even in challenging backgrounds where other methods suffer from missed detections or false alarms.

Conclusion

ECFNet successfully proposes an efficient coarse-to-fine infrared small target detection framework by combining RBCN for context region proposals, DAT for target–background discrimination, and APKD for guiding the fine detector with attention priors.

Here are the specific improvements and capabilities that an AI system, powered by the ECFNet framework described in this paper, can achieve:

1. The ECFNet framework significantly enhances performance in Infrared Small Target Detection (IRSTD) under challenging conditions by adopting a highly efficient coarse-to-fine strategy.

2. The system can effectively detect small targets (like UAVs or ground vehicles) even when they have weak features and are heavily obscured by complex, dynamic infrared backgrounds, which is the primary challenge in this domain.

3. By employing the Region Binary Classification Network (RBCN) during the coarse stage, the AI system can efficiently filter vast amounts of irrelevant background data into a manageable set of context region proposals containing potential targets, drastically reducing redundant computations compared to dense prediction methods.

4. The Denoising-Assisted Training (DAT) strategy allows the model to learn robust target-background discrimination by injecting synthetic target-like noise into the ground truth masks during training. This forces the network to explicitly model the contextual relationship between targets and clutter, leading to superior proposal quality and suppression of false alarms.

5. The Coarse-to-Fine pipeline ensures high accuracy:

late The system refines these context region proposals in a lightweight fine stage using an Attention Prior-Guided Knowledge Distillation (APKD) mechanism. This mechanism directs the lightweight detector to focus precisely on the most critical target regions emphasized by a powerful teacher model, leading to highly discriminative feature representations for IRSTs.

6. The resulting AI system achieves state-of-the-art performance across multiple real-world infrared datasets (UAV, Car, IRSTD-1k) while maintaining high real-time processing efficiency (high FPS), making it suitable for deployment on embedded hardware like edge devices or UAV platforms.

7. The system can achieve measurable improvements in detection metrics:

8. It yields significant gains in Recall (R%) and AP50% compared to existing single-stage and two-stage methods, while achieving substantial reductions in computational cost (FLOPs) for comparable performance levels.

9. For pixel-level tasks, the system improves both Pixel Detection Rate (Pd%) and Normalized Intersection over Union (nIoU%), providing highly accurate localization of targets within their detected regions.

Infrared small target detection (IRSTD) in high-resolution images is crucial for unmanned aerial vehicle (UAV) surveillance and UAV-based ground monitoring. However, small target size, weak features, and interference from complex dynamic backgrounds make IRSTD challenging. Existing methods incur redundant computation in non-target background regions and insufficiently exploit target context, limiting detection performance. To address these issues, we propose ECFNet, an efficient coarse-to-fine IRSTD framework with attention prior-guided knowledge distillation. In the coarse stage, we design a region binary classification network (RBCN) on grid-based multi-scale feature maps to efficiently identify target-containing context region proposals. A new denoising-assisted training strategy incorporates noisy ground-truth (GT) masks into RBCN feature maps and trains the network to reconstruct the original GT masks. This auxiliary task encourages explicit learning of target-background context to better distinguish target proposals from background regions. In the fine stage, we customize a lightweight target detector to the coarse-stage region proposals to balance accuracy and efficiency. Furthermore, we introduce a knowledge distillation strategy guided by a teacher-student cross-attention prior. This strategy directs the student to focus on critical target regions, enhancing discriminative feature representations for infrared small targets. Extensive experiments on three real infrared datasets demonstrate that ECFNet outperforms existing single-stage and two-stage approaches while maintaining high real-time processing efficiency. Code: https://github.com/IVPLabs/ECFNet.