Source-linked AI summary
EFLNet: Enhancing Feature Learning for Infrared Small Target Detection
Bo Yang, Xinyu Zhang, Jian Zhang, Jun Luo, Mingliang Zhou, Yangjun Pi
TL;DR
Infrared small-target detection is challenged by severe target-background imbalance, sensitivity of box regression to small position changes, and loss of target information in deep layers. EFLNet combines adaptive threshold focal loss, NWD, and a dynamic head; the authors report the highest performance across all evaluation metrics on three datasets versus SOTA methods.
Problem
Small infrared targets are easily overshadowed by background features, are highly sensitive to box-position changes under IoU, and can lose information during downsampling.
Source: §I · pp. 1–2Method
EFLNet combines ATFL for target-background imbalance, NWD for bounding-box regression, and a dynamic head that learns semantic-layer importance.
Source: §I · p. 2, C. Normalized Gaussian Wasserstein… · p. 5, D. Dynamic head · p. 6Results
EFLNet achieved the highest performance across all evaluation metrics on NUAA-SIRST, NUDT-SIRST, and IRSTD-1k compared with SOTA methods.
Source: B. Quantitative Results · pp. 6–7Takeaways & Limitations
The paper supplies bounding-box annotations for public infrared small-target datasets, supporting detection-based evaluation beyond mask-only datasets.
Source: §I · p. 2
Abstract
from arXiv · showhide
Single-frame infrared small target detection is considered to be a challenging task, due to the extreme imbalance between target and background, bounding box regression is extremely sensitive to infrared small target, and target information is easy to lose in the high-level semantic layer. In this article, we propose an enhancing feature learning network (EFLNet) to address these problems. First, we notice that there is an extremely imbalance between the target and the background in the infrared image, which makes the model pay more attention to the background features rather than target features. To address this problem, we propose a new adaptive threshold focal loss (ATFL) function that decouples the target and the background, and utilizes the adaptive mechanism to adjust the loss weight to force the model to allocate more attention to target features. Second, we introduce the normalized Gaussian Wasserstein distance (NWD) to alleviate the difficulty of convergence caused by the extreme sensitivity of the bounding box regression to infrared small target. Finally, we incorporate a dynamic head mechanism into the network to enable adaptive learning of the relative importance of each semantic layer. Experimental results demonstrate our method can achieve better performance in the detection performance of infrared small target compared to the state-of-the-art (SOTA) deep-learning-based methods. The source codes and bounding box annotated datasets are available at https://github.com/YangBo0411/infrared-small-target.
I. INTRODUCTION
Infrared small targets are difficult to detect because they occupy very little image area, are easily obscured by complex backgrounds, and lose information during downsampling. EFLNet addresses these challenges with tailored losses and a dynamic head, and the paper also provides bounding-box annotations for public datasets.
Motivation
- Infrared targets may occupy only a few pixels amid sparse instances and complex, noisy backgrounds, making them easy to overshadow.The paper notes that the low signal-to-clutter ratio compounds these challenges.Source: §I · p. 1
- Conventional filter-, local-contrast-, and low-rank-based methods have limitations in complex backgrounds, dim-target detection, or changing target shapes.The paper describes these limitations as relevant to practical infrared scenes.Source: §I · p. 1
- Many deep-learning approaches frame the task as segmentation, which requires pixel-level processing and does not directly evaluate target-level detection performance.The paper notes that segmentation can entail longer training and inference times.Source: §I · p. 1
Contributions
- EFLNet combines adaptive threshold focal loss, normalized Gaussian Wasserstein distance, and a dynamic head to address target-background imbalance, box-regression sensitivity, and loss of target information in deep layers.The paper also provides bounding-box annotations for public infrared small-target datasets.Source: §I · p. 2
II. RELATED WORK
Related work spans model-based methods and deep-learning approaches, the latter divided into segmentation-based and detection-based methods. The paper’s architecture then combines a feature backbone, multiscale fusion, a dynamic head, and two losses.
A. Model-based method
- Model-based filters, local-contrast methods, and low-rank decomposition rely on assumptions that limit performance with complex backgrounds, dim targets, or clutter.The cited approaches use background estimation, neighborhood intensity differences, or low-rank-background and sparse-target structure.Source: A. Model-based method · p. 2
B. Deep-learning based method
- Deep-learning methods adaptively extract image features and are categorized here as segmentation-based or detection-based approaches.The paper describes their performance as superior to traditional approaches in varied complex environments.Source: B. Deep-learning based method · p. 2
1) Segmentation-based methods:
- Segmentation-based work includes methods for balancing miss detection and false alarms, combining semantic and spatial cues, and mitigating information loss from pooling.Other cited methods address shape information, handcrafted-feature integration, multiscale features, and long-range dependencies.Source: 1) Segmentation-based methods · pp. 2–3
2) Detection-based methods:
- Detection-based methods directly predict target position and scale, with prior work using image enhancement, YOLO-based designs, cascade refinement, filtering, or temporal energy accumulation.The cited approaches also address bounding-box regression and sample misidentification.Source: 2) Detection-based methods · p. 3
A. Overall Architecture
- The EFLNet workflow extracts backbone features, fuses multiscale information with FPN and PAN, and sends the result to a dynamic head before NWD and ATFL compute training losses.The dynamic head learns the relative significance of semantic layers.Source: A. Overall Architecture · p. 3
B. Adaptive threshold focal loss
ATFL separates easy background samples from harder targets with a probability threshold, then adjusts loss weights to emphasize target features. Its adaptive hyperparameters are intended to reduce repeated tuning while retaining attention to difficult samples.
Threshold focal loss
- ATFL uses a prediction-probability threshold to separate easy background samples from hard target samples and increase target-focused learning.The paper describes intensifying target loss while mitigating background loss to address the imbalance.Source: B. Adaptive threshold focal loss · pp. 3–4
- Unlike focal loss, which can reduce both easy- and hard-sample losses, threshold focal loss reduces easy-sample weight while increasing hard-sample weight.The threshold classifies prediction probabilities above 0.5 as easy and below it as hard.Source: B. Adaptive threshold focal loss · pp. 3–4
Adaptive hyperparameters
- ATFL adaptively adjusts its hyperparameters to reduce the repeated tuning required across datasets and models.The paper links its adaptive design to lower time costs from multiple parameter adjustments.Source: B. Adaptive threshold focal loss · pp. 3–4
- The adaptive design estimates training progress from predicted probabilities and adjusts γ and η, while the figure compares loss changes across γ values.The paper describes exponential smoothing for the predicted probability and notes that late-training probabilities can reduce the proportion of difficult samples.Source: B. Adaptive threshold focal loss · p. 4, Figure 3 · p. 4
C. Normalized Gaussian Wasserstein distance
IoU is highly sensitive to small positional deviations for tiny infrared objects, complicating bounding-box regression. The section models boxes as 2D Gaussian distributions and uses normalized Wasserstein distance instead.
C. Normalized Gaussian Wasserstein distance
- A same position deviation changes tiny-object IoU from 0.47 to 0.08, versus 0.80 to 0.51 for normal-sized objects.The figure compares sensitivity for tiny and normal scale objects.Source: C. Normalized Gaussian Wasserstein… · pp. 4–5, Figure 4 · p. 5
- NWD replaces IoU-based similarity with Wasserstein distance, which can compare distributions with minimal or no overlap and is insensitive to object scale.The paper links this choice to positive-negative sample similarity and sparse positive samples during training.Source: C. Normalized Gaussian Wasserstein… · p. 5
- Each horizontal bounding box is modeled as a 2D Gaussian, and the Wasserstein distance between box distributions is simplified before exponential normalization to a 0–1 range.The normalization constant is related to the dataset.Source: C. Normalized Gaussian Wasserstein… · p. 5
D. Dynamic head
The dynamic head addresses information loss and differing feature semantics by adaptively attending to scale, spatial position, and task channels. Its three attention functions are designed to weight feature levels and spatial information according to their relevance.
D. Dynamic head
- The dynamic head adaptively focuses on scale-space-task information to learn the relative importance of semantic levels and spatial information.This addresses the concern that down-sampling can lose small-target information carried by shallow features.Source: D. Dynamic head · p. 5
- Scale-aware attention dynamically fuses features according to the importance of each pyramid layer.The figure distinguishes scale-aware, spatial-aware, and task-aware attention components.Source: D. Dynamic head · p. 6, Figure 5 · p. 6
- Spatial-aware attention uses deformable convolution to fuse features from different levels at the same spatial position.Task-aware attention dynamically switches channels on and off to support different tasks.Source: D. Dynamic head · p. 6
- The task-aware threshold controller reduces the layer-by-position dimensions through average pooling, then applies two fully connected layers, normalization, and sigmoid activation.Source: D. Dynamic head · p. 6
A. Dataset
Experiments use three public infrared small-target datasets with both bounding-box and semantic-segmentation annotations. Precision, recall, and F1 are used for evaluation.
A. Dataset
- Experiments use NUAA-SIRST, NUDT-SIRST, and IRSTD-1k, with each dataset divided into training, validation, and test sets at a 6:2:2 ratio.The experiments use both bounding-box and semantic-segmentation annotations.Source: A. Dataset · p. 6
- The public datasets provide image, semantic-segmentation, and bounding-box annotation forms.Fig. 6 depicts these three forms.Source: Figure 6 · p. 6
- Precision, recall, and F1 are used to compare the proposed method with state-of-the-art methods.Precision uses true positives and false positives; recall uses true positives and false negatives; F1 is their harmonic mean.Source: A. Dataset · p. 6
B. Quantitative Results
EFLNet achieves the highest reported performance across the evaluation metrics on all three datasets compared with the state-of-the-art methods. The paper relates this result to its focus on target-level detection challenges.
B. Quantitative Results
- EFLNet achieves the highest performance across all evaluation metrics on NUAA-SIRST, NUDT-SIRST, and IRSTD-1k compared with the state-of-the-art method.Table I compares precision, recall, and F1 across the three datasets; the supplied passages do not give the individual values.Source: B. Quantitative Results · p. 6, Table I · p. 7
- Pixel-level segmentation methods can produce false alarms or missed detections from slight inadequacies, lowering target-level precision, recall, and F1.The paper says many existing methods pay less attention to target-background imbalance and bounding-box sensitivity.Source: B. Quantitative Results · p. 7
- The authors attribute EFLNet’s stronger target-level detection to ATFL, NWD, and dynamic head learning.ATFL addresses target-background imbalance, while NWD and dynamic head are used to improve infrared small-target feature learning.Source: B. Quantitative Results · p. 7
C. Visual Results
Visual comparisons cover three datasets and distinguish correct detections, false alarms, and missed detections. EFLNet is reported to detect and localize targets despite local bright interference and dim appearances.
C. Visual Results
- EFLNet detects and localizes targets even with local highlight interference or dim target appearances.Fig. 7 presents partial results on NUAA-SIRST, NUDT-SIRST, and IRSTD-1k.Source: C. Visual Results · p. 7
- Other deep-learning methods tend to produce false alarms near local highlight interference and miss dim targets.The cited examples are Fig. 7 (1), (4), and (5) for interference, and (5), (7), and (8) for dim targets.Source: C. Visual Results · p. 7
- Fig. 7 marks correctly detected targets in red, false detections in orange, and missed detections in purple.More red circles and fewer orange and purple circles indicate superior detection performance in the visualization.Source: C. Visual Results · p. 7
D. Ablation Study
Ablations assess ATFL, NWD, dynamic head, and their combination, using IRSTD-1k. The reported results favor adaptive ATFL settings, NWD, and four dynamic-head modules.
D. Ablation Study
- On IRSTD-1k, adaptive ATFL hyperparameters yield better results with one tuning operation, while fixed settings require repeated η and γ adjustments.ATFL raises recall from the baseline 0.743 to as high as 0.790; overall performance is optimal at λ = 3.5 and threshold 0.5.Source: D. Ablation Study · p. 7, Table III · p. 7, Table IV · p. 7
- NWD performs best at C = 11, with the training-process comparison indicating alleviated convergence difficulty.The paper notes that C is closely tied to the dataset; Fig. 8 compares precision, recall, and AP0.5 during training.Source: D. Ablation Study · p. 9, Figure 8 · p. 9, Table V · p. 9
- With NWD in use, dynamic head improves performance, with four modules giving the optimal reported result.Increasing the number of dynamic-head modules brings slight performance gains.Source: D. Ablation Study · p. 9, Table VI · p. 9
- Adding each component improves performance, while only dynamic head increases the number of network parameters.The loss-function design does not add model complexity, and the authors report that it integrates with the network structure.Source: D. Ablation Study · p. 9
V. CONCLUSION
EFLNet combines adaptive loss weighting, NWD, and dynamic head learning to address target-background imbalance, bounding-box sensitivity, and loss of shallow target features. The authors report outperforming state-of-the-art methods on public datasets and provide bounding-box annotations for detection-based evaluation.
V. CONCLUSION
- EFLNet outperforms state-of-the-art methods on public infrared small-target datasets.The paper also provides bounding-box annotations for existing datasets, enabling detection-based treatment of the task.Source: §V · p. 9
- ATFL automatically adjusts loss weights to treat target and background differently, addressing their imbalance.This is intended to reduce the detector's tendency to focus on background information.Source: §V · p. 9
- NWD improves positive- and negative-sample quality to address the sensitivity of infrared small-target bounding-box regression to IoU.Source: §V · p. 9
- Dynamic head learns the relative importance of semantic layers and directs more attention to shallow target features.Source: §V · p. 9
- The ablation study evaluates ATFL, NWD, and dynamic head using precision, recall, AP0.5, F1, and GFLOPS.Source: Table VII · p. 10