基于深度学习的红外夜视图像超分辨率重建  

Super-resolution Reconstruction of Infrared Images in Night Environments Based on Deep-learning

作　　者：王丹 陈亮[1] WANG Dan;CHEN Liang(School of Information Science and Technology,DongHua University,Shanghai 201600,China)

机构地区：[1]东华大学信息科学与技术学院

出　　处：《红外技术》2019年第10期963-969,共7页Infrared Technology

摘　　要：针对红外夜视图像对比度低、成像质量不高的问题,提出适合红外夜视图像超分辨率重建方法。在自然图像超分辨率重建模型的基础上增加基于Retinex的对比度增强预处理步骤,并对网络模型做如下改进:构建超深卷积神经网络学习低分辨率图像与高分辨率图像之间的映射关系,增大感受野,提升网络学习能力;仅学习高低分辨率图像间的差值信息加速网络收敛。针对高分辨率红外夜视图像不易获得,数据量较少的问题,利用迁移学习理论,使用少量的高分辨率红外夜视图像为目标样本,对自然图像超分辨率重建模型进行微调,得到适合红外夜视图像重建的网络权重模型。实验结果证明:使用该方法得到的红外夜视图像信息丰富,层次分明,具有良好的视觉效果。To address the low contrast and poor image quality of infrared images recorded in a night environment(NIR),we propose an algorithm suitable for NIR super-resolution.Based on a color image super-resolution reconstruction model,a preprocessing step based on Retinex is added to enhance the contrast and improve the convolution neural network as follows:First,more layers are used to increase the receptive field to improve the learning ability of the network.Second,the residual image generated by lowand high-resolution images is learned to improve the convergence speed.Finally,the network trained by high-resolution images is fine-tuned to obtain weights that fit the NIR reconstruction with a small amount of data according to transfer learning,because the lack of data makes it difficult to get more NIR images.Simulation results shows that the proposed algorithm performs well,yielding a reconstructed image with higher contrast,richer detail,and better visuals.

关 键 词：红外夜视图像 超分辨率 预处理 超深神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]