MSGEGA: Multiscale Gaussian Enhancement and Global-Aware Network for Infrared Small Target Detection

In recent years, the development of deep learning has significantly improved the performance of infrared small targets detection, particularly in addressing the issue of detecting small infrared targets with high signal-to-clutter ratio (SCR). However, detecting infrared small targets in complex bac...

Full description

Saved in:
Bibliographic Details
Main Authors: Yuyang Xi, Liuwei Zhang, Ying Jiang, Feng Qian, Fanjiao Tan, Qingyu Hou
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11045977/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent years, the development of deep learning has significantly improved the performance of infrared small targets detection, particularly in addressing the issue of detecting small infrared targets with high signal-to-clutter ratio (SCR). However, detecting infrared small targets in complex backgrounds with low SCR and poor contrast with the background still faces challenges such as low detection rate and high false alarm rate. In this article, we propose a novel network called multiscale Gaussian enhancement and global-aware (MSGEGA) network. MSGEGA combines multiscale Gaussian enhancement with global-aware to strengthen the ability to perceive global background features and capture local detailed features of the target. We design infrared small targets feature enhancement module based on multiscale Gaussian templates, which is applied to the head and skip connections to further improve the distinguishability between IRSTs with low-SCR and background. Meanwhile, global-aware leverages its global receptive field to suppress background and clutter components with high mutual similarity in images, integrates fine structures and boundary information from different layers, thereby enhancing the local feature representation of IRSTs. To validate the effectiveness of the proposed method, publicly available datasets are filtered to form a specialized low SCR dataset named IRST-LSCR. Extensive experiments show that MSGEGA achieves outstanding performance in detecting small targets with low SCR (SCR &lt; 6). Specifically, the proposed method demonstrates significant advantages on the screened dataset, achieving an AUC of 0.992. At a detection rate of 0.871, it maintains a false alarm rate of 0.9<italic>e</italic>-5, outperforming all comparison algorithms. The model has 1.17 M parameters and 2.50 GFLOPs, achieving 39.80 frames per second on a single RTX 4090.
ISSN:1939-1404
2151-1535