基于空洞循环卷积的近红外图像彩色化方法  

Near infrared images colorization method based on Dilated-Cycle convolution

作　　者：高美玲 段锦[1] 莫苏新 刘高天 赵伟强 张昊 GAO Meiling;DUAN Jin;MO Suxin;LIU Gaotian;ZHAO Weiqiang;ZHANG Hao(College of Electronic and Information Engineering,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,吉林长春130022

出　　处：《光学技术》2022年第6期742-748,共7页Optical Technique

基　　金：国家重大科研仪器研制项目(62127813)。

摘　　要：针对近红外图像彩色化过程中存在近红外与可见光图像间模态差异较大、图像域样式不佳而导致其彩色化结果出现色彩纹理不吻合的问题,提出一种基于空洞循环卷积的近红外图像彩色化方法。改进CycleGAN网络,利用级联结构和空洞卷积块的优势设计了一种名为空洞级联模块,该模块采用编解码级联结构,代替原模型残差网络中的单向连接结构,进行特征通道的级联。在编解码级联层中引入空洞卷积模块,利用空洞卷积不损失图像纹理细节的优势,进一步提取不同尺度的近红外图像特征信息,最后通过解码将近红外灰度图像上色成彩色近红外图像。算法在生成网络中采用空洞级联方式改善误着色问题;在判别网络中引用感知损失函数改善网络收敛速度慢的问题。并且在NIR_VIS数据集上开展验证和分析。实验结果表明,所提方法提升效果明显,更好地保持了原目标的结构及色彩纹理特征,有效地提升了近红外图像的可视化效果。In the process of colorization of NIR images,there are large modal differences between NIR images and visible images in poor images domain styles,which lead to color texture mismatch in the colorization results.The CycleGAN network is improved,a concatenated module called the Dilated Cascade Block is designed by taking advantage of the cascade structure and the Dilated convolution block.This module adopts the encode and decode cascade structure to replace the one-way connection structure in the original model residual network.A Dilated Convolution module is introduced into the coding-decoding cascade layer to further extract feature information of NIR images of different scales by utilizing the advantage of cavity convolution without losing texture details of the images.Finally,NIR of colorful images are obtained by decoding the NIR gray images.The algorithm uses the Dilated cascade method to solve the texture mismatch problem in the generation network.The perceptive loss function is used to improve the slow convergence of discriminant networks.Validation and analysis are carried out on NIR_VIS dataset.Experimental results show that the proposed method can improve the structure and color texture of the original object better,and effectively improve the visualization effect of NIR images.

关 键 词：近红外图像 彩色化 空洞级联 对抗生成网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]