基于图像翻译的红外弱小目标数据增强算法  

Infrared small dim target data augmentation algorithm based on image translation

作　　者：廖彦彬 季钰翔 傅志凌 杨海 王喆[1] LIAO Yan-bin;JI Yu-xiang;FU Zhi-ling;YANG Hai;WANG Zhe(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China;Shanghai Science and Technology Museum,Shanghai 200127,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237 [2]上海科技馆,上海200127

出　　处：《激光与红外》2025年第2期304-312,共9页Laser & Infrared

基　　金：上海科技馆科研专项项目(No.SSTM/SOPZY-03-R2KZ-20230100001);中国科技国防计划项目(No.2021-JCJQ-JJ-0041)资助。

摘　　要：为了缓解在红外弱小目标检测问题中图像数据稀缺的问题,提出了一种基于图像翻译的红外弱小目标图像数据增强算法。该方法是一个两阶段的图像生成算法,首先引入额外的可见光图像,通过U-GAT-IT模型学习可见光和红外图像之间的映射,将可见光图像转化为红外背景图像。为了解决图像翻译过程中的过拟合问题,提出了通道正则化方法,使红外和可见光图像的通道信息量保持一致。接着,设计了一个基于视觉Transformer结构的自编码器,学习红外小目标的分布特征,以遮挡重构的方式在得到的红外背景图像上合成弱小目标。本方法在SIATD数据集上进行训练和测试,实验结果表明提出的数据增强方法在三个模型上使检测指标得到了一定提升,其中在YOLOv3模型上AP指标提高了1.37%,证明了提出的数据增强算法的有效性,能够提高目标检测模型在红外弱小目标检测任务中的表现。In order to alleviate the scarcity of image datain infrared small dimtarget detection,an image augmentation algorithm based on image-to-image translation is proposed.This method is a two-stage image generation algorithm.First,additional visible image is introduced,and the mapping between visible and infrared image is learned by U-GAT-IT model,converting the visible image into infrared background image.In order to solve the problem of overfitting in image translation,a channel regularization method is proposed to make the channel information of infrared and visible images consistent Then,an auto-encoder based on vision Transformer structure is designed to learn the distribution characteristics of infrared small targets and synthesize small targets on the obtained infrared background images in the way of masking and reconstructing.The method is trained and tested on SIATD data sets.The experimental results show that the proposed data augmentation algorithm can improve the detection indexes on three models to a certain extent,among which the AP index of the YOLOv3 model increases by 1.37%,which proves the effectiveness of the proposed data augmentation algorithm and can improve the performance of the target detection model in the infrared small dim target detection task.

关 键 词：图像翻译 数据增强 目标检测 红外弱小目标检测 生成对抗网络 自编码器 转换器网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN219[自动化与计算机技术—计算机科学与技术]