基于深度神经网络的遥感图像彩色化方法  

Remote Sensing Image Colorization Based on Deep Neural Networks with Multi-Scale Residual Receptive Filed

作　　者：冯佳男 江倩 金鑫[1,2] 李昕洁 黄珊珊 姚绍文[1,2] Feng Jianan;Jiang Qian;Jin Xin;Lee Shin-Jye;Huang Shanshan;Yao Shaowen(School of Software,Yunnan University,Kunming 650504;Engineering Research Center of Cyberspace,Yunnan University,Kunming 650504;Institute of Technology Management,Xinzhu Chiao Tung University,china Xinzhu 30010)

机构地区：[1]云南大学软件学院,昆明650504 [2]云南大学跨境网络空间安全教育部工程研究中心,昆明650504 [3]新竹交通大学科技管理研究所,新竹30010

出　　处：《计算机辅助设计与图形学学报》2021年第11期1658-1667,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61863036,62002313,62101481);中国博士后科学基金(2019M653507,2020T130564);云南省重点领域科技计划(202001BB050076);云南大学研究生科研创新基金(2020231).

摘　　要：为了实现遥感图像彩色化,解决目前彩色化模型存在颜色不准确和颜色溢出等问题,提出一种端到端的深度神经网络模型.首先,通过构建多尺度残差感受域块提取丰富的高维特征;其次,利用U-Net、复杂残差结构、注意力机制和子像素向上卷积等结构构建一个彩色信息重建网络输出彩色化结果;最后,使用NWPU-RESISC45遥感图像数据集进行训练和验证.结果表明,与其他自动彩色化方法相比,所提方法的PSNR值平均提高6~10 dB,SSIM值增加0.05~0.11,实现了遥感图像彩色化.此外,该方法在RSSCN7和AID数据集上也取得了良好的彩色化效果.To solve the problems of mistaken coloring and color bleeding in the current colorization methods,an end-to-end deep neural network is proposed to achieve remote sensing image colorization.First,the mul-ti-scale residual receptive filed net is introduced to extract the key features of source image.Second,a color information recovery network is con-structed by using U-Net,complex residual structure,attention mecha-nism,sequeeze-and-excitation and pixel-shuffle blocks to obtain color result.NWPU-RESISC45 dataset is chosen for model training and validation.Compared with other color methods,the PSNR value of the pro-posed method is increased by 6-10 dB on average and the SSIM value is increased by 0.05-0.11.In addition,the proposed method also achieves satisfactory color results on RSSCN7 and AID datasets.

关 键 词：深度神经网络 遥感图像 图像彩色化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]