人工智能
计算机科学
深度学习
稳健性(进化)
特征提取
红外线的
模式识别(心理学)
特征(语言学)
计算机视觉
目标检测
分割
领域(数学)
融合
图像(数学)
卷积神经网络
图像分割
图像处理
传感器融合
图像融合
遥感
自动目标识别
作者
Qiuyue Xu,Haoran Li,Yuyu Zhu,Qiang‐Ming Cai,Yue Pan,Qicao Chen
出处
期刊:2019 PhotonIcs & Electromagnetics Research Symposium - Spring (PIERS-Spring)
[Institute of Electrical and Electronics Engineers]
日期:2025-05-04
卷期号:: 1-6
标识
DOI:10.1109/piers-spring66516.2025.11276614
摘要
Infrared detection imaging technology primarily relies on the interaction between objects and infrared radiation, featuring strong anti-interference capabilities and excellent concealment. Therefore, infrared small target detection (IRSTD) is widely applied in fields such as precision guidance, security monitoring, and military reconnaissance. However, due to the variability in the shape, size, and signal-to-noise ratio of small targets, traditional IRSTD methods adopt a modeldriven approach by first making appropriate assumptions about the characteristics of infrared targets in images and then designing algorithms manually to detect these targets. Limited by these preliminary assumptions, traditional methods struggle to achieve accurate and effective detection and lack robustness and applicability. In recent years, methods based on deep learning have become increasingly popular and have achieved promising results in the field of target image detection. Deep learning methods adopt a data-driven approach to separate targets from the background by learning the features of the targets, thereby achieving the purpose of detection. This paper discusses different methods for segmenting and detecting infrared small targets using deep learning and proposes a deep network that the extraction and combination of deep and shallow features for infrared small target detection (ECDSF). The proposed ECDSF algorithm comprises a feature extraction module, an attention perception module, and an information fusion module. It initially performs deep and shallow feature extraction, leveraging an attention strategy to capture feature information at different levels, and by combining the informational characteristics of both deep and shallow features, it achieves effective detection while reducing algorithm complexity. Extensive experimental results on infrared single-frame image datasets demonstrate that, compared to existing methods, ECDSF-Net exhibits good Intersection over Union (IoU) and normalized IoU (nIoU), demonstrating superior detection performance and practicality.
科研通智能强力驱动
Strongly Powered by AbleSci AI