基于生成对抗网络的红外图像自适应校正算法  

Adaptive correction algorithm for infrared image based on generative adversarial network

作　　者：牟新刚[1] 崔健 周晓[1] MOU Xin-gang;CUI Jian;ZHOU Xiao(School of Mechanical and Electronic 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：《激光与红外》2022年第3期427-434,共8页Laser & Infrared

基　　金：国家自然科学基金项目(No.61701357)资助。

摘　　要：红外相机经过两点校正后会发生漂移,导致图像产生非均匀性噪声。传统的基于场景的非均匀性校正算法适应性较强,但会产生鬼影问题。基于深度学习的非均匀性校正方法,面对红外图像特殊的非均匀性噪声,随着网络深度增加,很容易出现图像模糊,对比度变小,细节丢失问题。针对此问题,提出一种基于生成对抗网络的非均匀性校正算法,网络分为生成网络和判别网络两部分,生成网络部分使用多尺度残差连接,将各解码层分别与不同深度的特征图进行残差连接,不断生成图片,判别网络判别生成网络生成的图片和真实图片,二者不断博弈和迭代训练。生成网络通过多尺度残差连接,很好地保留了网络各层的信息。通过评价红外图像非均匀性校正效果后表明,该网络去除红外图像的非均匀性噪声效果良好,图片清晰,边缘锐利,很好的保留了原图片的各种细节,不会产生鬼影问题。The infrared camera drifts after two-point correction,which will cause non-uniform noise in the image.The traditional scene-based non-uniformity correction algorithm is more adaptable,but it will cause“ghost imaging”.Aiming at the special non-uniformity noise of infrared images,the non-uniformity correction method based on deep learning will cause image blur,reduced contrast,and loss of details as the network depth increases.A non-uniformity correction algorithm based on generative adversarial network is proposed.The network contains two parts:generation network and discriminant network.The generation network use multi-scale skip connections to connect each decoder layer and feature maps of different depths,and generate pictures.The discriminant network discriminates the pictures generated by the generation network and the real pictures.Two parts of the network game each other and are trained iteratively.The information of each layer of the network is well preserved through the generation network with multi-scale skip connections.After evaluating the effect of infrared image non-uniformity correction,it is found that the network can well remove the non-uniformity noise of infrared image.The image is clear and the edge is sharp.It retains all kinds of details of the original image and will not produce“ghost imaging”.

关 键 词：生成对抗网络 红外图像 非均匀性校正 深度学习 

分 类 号：TN219[电子电信—物理电子学]