|国家科技期刊平台
首页|期刊导航|红外技术|基于密集残差生成对抗网络的红外图像去模糊

基于密集残差生成对抗网络的红外图像去模糊OA北大核心CSTPCD

Infrared Image Deblurring Based on Dense Residual Generation Adversarial Network

中文摘要英文摘要

红外图像拍摄过程中,由于摄像设备抖动或目标快速移动会导致图像出现运动模糊,极大影响了有效信息的提取和识别.针对上述问题,本文在 DeblurGAN 基础上提出一种基于密集残差生成对抗网络的红外图像去模糊方法.该方法首先采用多尺度卷积核,提取红外图像不同尺度和层次的特征.其次,采用密集残差块(residual-in-residual dense block,RRDB)代替原生成网络中的残差单元,改善恢复红外图像的细节信息.通过本课题组自制的红外图像数据集进行实验,结果表明所提出的方法与DeblurGAN相比PSNR提高 3.60 dB,SSIM提高 0.09,主观视觉去模糊效果较好,恢复后的红外图像边缘轮廓清晰且细节信息明显.

During infrared(IR)image capture,the shaking of camera equipment or rapid movement of the target causes motion blur in the image,significantly affecting the extraction and recognition of effective information.To address these problems,this study proposes an infrared image deblurring method based on a dense residual generation adversarial network(DeblurGAN).First,multiscale convolution kernels are employed to extract features at different scales and levels from infrared images.Second,a residual-in-residual dense block(RRDB)is used,instead of the residual unit in the original generation network,to improve the detail of the recovered IR images.Experiments were conducted on the infrared image dataset collected by our group,and the results show that compared to DeblurGAN,the proposed method improves PSNR by 3.60 dB and SSIM by 0.09.The subjective deblurring effect is better,and the recovered infrared images have clear edge contours and detail information.

李立;易诗;刘茜;程兴豪;王铖

成都理工大学 机电工程学院,四川 成都 610059

计算机与自动化

生成对抗网络密集残差块红外图像去运动模糊

generative adversarial networkresidual-in-residual dense blockinfrared imagemotion deblurring

《红外技术》 2024 (006)

663-671 / 9

四川省自然科学基金面上项目(24NSFSC1481),成都理工大学高等教育人才培养质量和教学改革项目(JG2130216).

评论