基于残差编解码网络的红外图像自适应校正算法  

Adaptive Correction Algorithm of Infrared Image Based on Encoding and Decoding Residual Network

作　　者：牟新刚[1] 陆俊杰 周晓[1] MOU Xingang;LU Junjie;ZHOU Xiao(College of Mechanical and Electrical Engineering,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学机电工程学院,湖北武汉430070

出　　处：《红外技术》2020年第9期833-839,共7页Infrared Technology

基　　金：国家基金项目(61701357);中央高校基本科研业务费专项资金资助(183204007)。

摘　　要：针对基于场景的非均匀性校正算法存在非均匀性残余和鬼影等问题,本文提出了一种基于残差编解码网络的红外图像自适应算法。该算法针对自适应校正问题的特点,基于UNet结构,通过多尺度采样学习残差映射生成非均匀性残差图像,加入批标准化和PReLU激活函数提高校正效果,最后使用全局跳跃连接得到最终的校正结果。通过对模拟红外图像序列和真实红外图像序列校正的实验结果表明,相对于目前已有的非均匀性校正算法,该方法在PSNR(Peak Signal to Noise Ratio)和粗糙度的客观数据上都有所提升,主观视觉效果也更加清晰,细节保留程度高。Traditional scene-based non-uniformity correction algorithms generally suffer from non-uniformity residuals and ghosts.In view of this,we propose an infrared image adaptive algorithm based on the encoding and decoding residual network.The algorithm focuses on the characteristics of the adaptive correction problem.Following the UNet structure,the residual image is generated through multiscale sampling and learning residual mapping.Batch normalization and PReLU are used to improve the correction effect.Finally,the global skip connection is used to obtain the final correction result.The experimental results of correcting the simulated non-uniform infrared image sequence and the real infrared image sequence showed that this method improved the objective data of the peak signal to noise ratio(PSNR)and roughness,compared with existing non-uniformity correction algorithms.Moreover,the subjective visual effect was clearer,and the degree of detail retention was high.

关 键 词：红外图像 非均匀性校正 多尺度采样 残差学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]