Dense Nested Attention Network for Infrared Small Target Detection (2106.00487v3)

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).

• The paper introduces DNA-Net, a Dense Nested Attention Network, designed to overcome the challenges of detecting small infrared targets by preserving target information in deep network layers.
• DNA-Net utilizes a Dense Nested Interactive Module (DNIM) for multi-layer feature fusion and a Cascaded Channel and Spatial Attention Module (CSAM) for adaptive feature enhancement.
• Experimental results show DNA-Net achieves superior detection probability, lower false alarm rate, and improved IoU compared to state-of-the-art methods on a new dataset, NUDT-SIRST.

An Expert Overview of Dense Nested Attention Network for Infrared Small Target Detection

The paper by Boyang Li et al. introduces a Dense Nested Attention Network (DNA-Net) aimed at addressing the challenges inherent in Single-frame Infrared Small Target (SIRST) detection. Infrared small target detection is pivotal in applications such as maritime surveillance and precise guidance systems, where targets are typically small, dim, shapeless, and can vary significantly in size and shape.

Problem and Solution

The authors identify a critical limitation in existing convolutional neural network (CNN)-based methods: the inability to effectively capture and maintain the representation of small infrared targets in deeper layers due to the pooling operations. Existing networks designed for generic object detection tend to underperform when applied to SIRST because of the reduced target size, which leads to loss of target information in deeper layers.

To overcome this, DNA-Net is proposed. This architecture incorporates a Dense Nested Interactive Module (DNIM) and a Cascaded Channel and Spatial Attention Module (CSAM). The DNIM is designed to facilitate progressive interaction between high-level and low-level features through dense nested connectivity, ensuring that small targets remain represented even in deeper layers. The CSAM further enhances multi-level feature representation by applying channel and spatial attention mechanisms, allowing for adaptive enhancement of the feature maps.

Methodology

1. Dense Nested Interactive Module (DNIM): By stacking multiple U-shaped subnetworks, the DNIM achieves repetitive feature fusion at intermediate nodes, integrating features across different layers to preserve small target information.
2. Cascaded Channel and Spatial Attention Module (CSAM): This module adaptively enhances feature representation, enabling efficient multi-layer feature fusion while maintaining the semantic information of small targets.
3. Dataset Development: The authors introduce a novel dataset, NUDT-SIRST, intended to reflect the diversity of real-world scenarios in terms of target size, type, and background clutter. This augmentation serves to evaluate the robustness of the proposed network.

Results and Implications

The experimental results presented in the paper demonstrate that DNA-Net outperforms state-of-the-art methods by achieving superior detection probability (${P}_{d}$), a lower false alarm rate (${F}_{a}$), and improved intersection over union (IoU). The strong numerical results underscore the network's ability to maintain small target information throughout deeper network layers, leading to more accurate detection in diverse and challenging environments.

The introduction of NUDT-SIRST advances the field by providing a comprehensive dataset that includes diverse clutter backgrounds and various target scenarios. It reflects an important contribution to benchmarking the effectiveness of algorithms tackling SIRST detection.

Future Directions

The advancements discussed in this paper could pave the way for further research into adaptive feature aggregation techniques, attention mechanisms, and the design of network architectures capable of addressing specific challenges associated with small-scale object detection. Additionally, incorporating these advanced methods into real-time systems could yield practical benefits for surveillance and defense applications.

In conclusion, the paper presents a methodologically sound, well-substantiated improvement over legacy SIRST detection methods. DNA-Net demonstrates particular robustness in preserving and exploiting small target information across the processing hierarchy, marking a significant contribution to the field of infrared target detection. Future research may build upon these findings to refine and expand the applicability of such networks in more complex operational environments.