[Paper Review] Dense Nested Attention Network for Infrared Small Target Detection

Boyang Li, Chao Xiao+6 more|arXiv (Cornell University)|Jun 1, 2021

Infrared Target Detection Methodologies22 citations

TL;DR

This paper proposes a Dense Nested Attention Network (DNA-Net) for infrared small target detection, leveraging a tri-directional dense nested interactive module (DNIM) and cascaded channel-spatial attention to enable progressive, multi-level feature fusion and enhancement, preserving small targets in deep layers. The method achieves state-of-the-art performance on both public and a newly introduced large-scale dataset (NUDT-SIRST), outperforming prior methods in detection probability, false alarm rate, and IoU.

ABSTRACT

Single-frame infrared small target (SIRST) detection aims at separating small targets from clutter backgrounds. With the advances of deep learning, CNN-based methods have yielded promising results in generic object detection due to their powerful modeling capability. However, existing CNN-based methods cannot be directly applied for infrared small targets since pooling layers in their networks could lead to the loss of targets in deep layers. To handle this problem, we propose a dense nested attention network (DNANet) in this paper. Specifically, we design a dense nested interactive module (DNIM) to achieve progressive interaction among high-level and low-level features. With the repeated interaction in DNIM, infrared small targets in deep layers can be maintained. Based on DNIM, we further propose a cascaded channel and spatial attention module (CSAM) to adaptively enhance multi-level features. With our DNANet, contextual information of small targets can be well incorporated and fully exploited by repeated fusion and enhancement. Moreover, we develop an infrared small target dataset (namely, NUDT-SIRST) and propose a set of evaluation metrics to conduct comprehensive performance evaluation. Experiments on both public and our self-developed datasets demonstrate the effectiveness of our method. Compared to other state-of-the-art methods, our method achieves better performance in terms of probability of detection (Pd), false-alarm rate (Fa), and intersection of union (IoU).

Motivation & Objective

• To address the challenge of infrared small targets being lost in deep CNN layers due to pooling operations.
• To enhance feature representation of small targets by enabling progressive, repeated fusion of high- and low-level features.
• To develop a large-scale, diverse infrared small target dataset (NUDT-SIRST) to support robust evaluation under realistic conditions.
• To introduce a comprehensive set of evaluation metrics tailored for infrared small target detection.
• To demonstrate that data augmentation with synthetic images can reduce reliance on costly real-world data while improving model generalization.

Proposed method

• Proposes a dense nested interactive module (DNIM) that connects all encoder and decoder layers in a tri-directional, fully connected nested structure to enable repeated multi-level feature interaction.
• Introduces a cascaded channel and spatial attention module (CSAM) to adaptively enhance features across multiple scales by modeling interdependencies in both channel and spatial dimensions.
• Employs a feature pyramid fusion module (FPFM) to aggregate features from multiple layers, combining low-level details and high-level semantics for robust detection.
• Designs a novel training strategy using a mixed dataset of real and synthetic infrared images to improve model generalization and reduce data collection costs.
• Utilizes a multi-scale supervision strategy by supervising detection heads at multiple decoder layers to preserve small target features during training.
• Applies a hybrid loss function combining binary cross-entropy and dice loss to optimize segmentation performance.

Experimental results

Research questions

• RQ1 Can a dense nested architecture effectively preserve small target features in deep convolutional layers, preventing their loss due to downsampling?
• RQ2 To what extent does cascaded channel and spatial attention improve feature representation for infrared small targets?
• RQ3 How effective is the proposed DNA-Net compared to state-of-the-art methods on both public and newly collected datasets?
• RQ4 Can synthetic infrared small target images effectively augment real data to improve model generalization and reduce annotation costs?
• RQ5 Does the proposed evaluation metric set better reflect the performance of infrared small target detection systems in complex, real-world scenarios?

Key findings

• DNA-Net achieves a 77.47% IoU, 98.48% detection probability, and 4.223×10⁻⁶ false alarm rate on the NUAA-SIRST test set when trained on a mixed dataset of real and synthetic images.
• The ablation study shows that removing the feature pyramid fusion module (FPFM) reduces IoU by 0.83% and increases the false alarm rate by 1.81×10⁻⁶, demonstrating its critical role in performance.
• Removing features from the deepest layer (L5) leads to a 0.23% drop in IoU and a 3.01×10⁻⁶ increase in false alarm rate, indicating the importance of deep-layer features for small target detection.
• Training on a mixed dataset with only 16.2% real images yields comparable performance to full real-data training, reducing data collection costs while maintaining high accuracy.
• The network's output on real images surpasses manually labeled ground truth masks in shape segmentation precision, indicating superior feature learning from well-labeled synthetic data.
• The NUDT-SIRST dataset, containing diverse target sizes, shapes, and clutter backgrounds, enables more comprehensive evaluation than existing benchmarks.