基于U型多尺度Transformer网络的红外小目标检测算法  

Infrared small target detection algorithm with U-shaped multiscale transformer network

作　　者：段沛沛[1] 张严 雒明世[1] 闫效莺[1] DUAN Peipei;ZHANG Yan;LUO Mingshi;YAN Xiaoying(School of Computer Science,Xi'an Shiyou University,Xi'an 710065,China;School of Electric Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi'an 710021,China)

机构地区：[1]西安石油大学计算机学院,陕西西安710065 [2]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《西北工业大学学报》2025年第1期154-162,共9页Journal of Northwestern Polytechnical University

基　　金：陕西省自然科学基础研究计划(S2023-JC-YB-2680)资助。

摘　　要：针对红外小目标特征难以提取、易被噪声干扰及复杂背景淹没等问题,提出了一种基于U型多尺度Transformer网络的检测算法。该算法在U型多尺度网络架构下,借助卷积操作提取、强化小目标局部显著性特征,同时又基于Transformer机制对图像全局特征进行建模,以获取红外图像背景信息;通过对所生成目标置信图与特征图的自注意力运算,完成了对图像浅层和深层特征的融合,实现了对像素级红外小目标的分割及检测。实验证明,在红外序列图像弱小飞机目标检测跟踪数据集中,即使针对背景复杂且含噪的图像进行检测,所提算法性能仍然优于对比算法,呈现了良好的鲁棒性及稳定、准确的检测效果。在算法阈值选用使FM平均值最大的情况下,其检测率为0.9972,虚警率为2.82×10^(-7),精确率为0.9127,而召回率则为0.921。To solve the problem of small targets feature extraction and the susceptibility of targets to being overwhelmed by noise and complex backgrounds,a detection method with U-shaped multiscale transformer network is proposed.Based on the U-shaped multiscale network architecture,the proposed method uses convolution operations to extract and enhance local salient features of small targets.Concurrently,it uses the Transformer mechanism to model global image features,facilitating the extraction and suppression of the image background.Subsequently,through self-attention operations on target confidence maps and feature maps,fusion of shallow and deep features in images is achieved.This accomplishes pixel-level segmentation of infrared small targets,fulfilling the purpose of target detection.Experiments demonstrate in infrared sequence image dim and small aircraft target detection and tracking data set,even when applied to infrared images with complex background and noisy,our method outperforms the state-of-the-art detection methods.The method shows good robustness and high detection accuracy.When the threshold is selected to maximize the average of FM,the detection rate of our method reaches 0.9972,its false alarm rate is 2.82×10^(-7),the precision rate is 0.9127,and the recall rate is 0.921.

关 键 词：红外小目标检测 图像分割 深度学习 自注意力机制 

分 类 号：TN219[电子电信—物理电子学]