基于重参数化的红外小目标检测网络  

Infrared Small Target Detection Network Based on Reparameterization

作　　者：朱翊翔 闵志方[1] 朱雪琼 王翔[1] ZHU Yi-xiang;MIN Zhi-fang;ZHU Xue-qiong;WANG Xiang(Huazhong Institue of Electro-Optics—Wuhan National Laboratory for Optoelectronics,Wuhan 430223,China;Wuhan Institute of Design and Science,School of Information Engineering,Wuhan 430205,China)

机构地区：[1]华中光电技术研究所—武汉光电国家研究中心,湖北武汉430223 [2]武汉设计工程学院信息工程学院,湖北武汉430225

出　　处：《光学与光电技术》2024年第1期10-17,共8页Optics & Optoelectronic Technology

摘　　要：红外小目标检测通常受制于较远的成像距离,使得提取目标特征成为了一种困难,如何增强目标的特征表达是近些年的主要方向之一。而过于复杂的特征表达会损失推理速度,这对于有实时性要求的红外小目标检测任务是不利的。通过使用重参数化技术结合领域中常用的残差网络作为特征提取网络,再使用额外注意力与通道注意力作为特征增强模块与特征融合模块,在数据集上取得了较好的结果。提出的模型在SIRST与IRSTD-1K数据集上分别取得了0.734与0.638的mIoU值,同时参数量和计算复杂度只有0.306 M与1.114 G FLOPs。该模型能够在推理阶段保持较少参数的同时拥有和其他领先的方法相近甚至领先的性能,在串行运行的环境上有着明显的优势。Infrared small target detection is usually limited by a long imaging distance,which makes it difficult to extract target features.How to enhance target feature expression is one of the main research directions in recent years.However,too complex feature representation will lose the speed of inference.In this paper,we use reparameterization technology and residual network as feature enhancement module and feature fusion module,and achieve good results on the datasets.On SIRST and IRSTD-1K datasets,the proposed method achieves 0.734 and 0.638 mIoU,while having only 0.306M and 1.114G FLOPs in parameter number and computational complexity.Our model can maintain fewer parameters in the inference stage while having performance similar to or even leading other leading methods,which has obvious advantages in a serial environment.

关 键 词：红外小目标检测 深度学习 卷积神经网络 模型压缩 注意力机制 

分 类 号：TN219[电子电信—物理电子学]