Dense Nested Attention Network for Infrared Small Target Detection | JoVE Visualize

Area of Science:

Computer Vision
Artificial Intelligence
Infrared Imaging

Background:

Single-frame infrared small target (SIRST) detection is crucial for separating small objects from complex backgrounds.
Convolutional Neural Network (CNN)-based methods show promise in object detection but struggle with SIRST due to information loss in pooling layers.
Existing CNN architectures are not optimized for preserving subtle features of small targets in infrared imagery.

Purpose of the Study:

To propose a novel deep learning network, the Dense Nested Attention Network (DNA-Net), specifically designed for SIRST.
To address the challenge of information loss in deep layers of CNNs for infrared small target detection.
To enhance the contextual understanding and feature representation of small targets in infrared images.

Main Methods:

Developed a Dense Nested Interactive Module (DNIM) to facilitate progressive interaction between high-level and low-level features, preserving target information.
Introduced a Cascaded Channel and Spatial Attention Module (CSAM) for adaptive enhancement of multi-level features within the network.
Implemented DNA-Net, leveraging repetitive feature fusion and enhancement to incorporate contextual information of small targets.

Main Results:

The proposed DNA-Net effectively maintains infrared small target information even in deeper network layers.
Experiments on public and a newly developed dataset (NUDT-SIRST) demonstrate superior performance compared to state-of-the-art methods.
Achieved significant improvements in probability of detection (Pd), reduction in false-alarm rate (Fa), and enhanced intersection over union (IoU).

Conclusions:

DNA-Net offers an effective solution for single-frame infrared small target detection by overcoming limitations of standard CNNs.
The proposed DNIM and CSAM modules contribute to robust feature learning and target representation.
The method shows strong potential for real-world applications requiring accurate detection of small infrared targets.