基于图像融合和超分辨率重建的近红外图像彩色化  

Colorization of Near-infrared Images Based on Image Fusion and Super-Resolution Reconstruction

作　　者：杨薪 孔玲君[2] 王文举[1] 

机构地区：[1]上海理工大学出版印刷与艺术学院,上海 [2]上海出版印刷高等专科学校印刷包装工程系,上海

出　　处：《软件工程与应用》2023年第6期873-882,共10页Software Engineering and Applications

摘　　要：为了提升近红外图像彩色化的图像质量和分辨率,提出通过图像融合和超分辨率重建相结合的方式实现近红外图像彩色化。首先,将近红外图像和RGB彩色图像输入到带有通道注意力机制模块的图像融合网络中得到融合后的彩色化图像,然后再通过亮度增强模块提升图像亮度,最后通过图像超分辨率重建提升图像分辨率。通过实验,将本文算法与其他算法进行对比,结果表明主观视觉上本文彩色化结果的图像质量和分辨率都更高,客观评价上,基于KAIST数据集的对比实验表明,本文算法的PSNR值为20.617,SSIM值为0.882,与其他算法相比,PSNR值和SSIM值分别提升了137.91%和87.66%;在IVRL RGB-NIR数据集上的PSNR值为18.295,SSIM值为0.796,与其他算法相比也都有提升。本文算法能够有效地提升近红外图像彩色化后的图像质量及分辨率,具有一定的实际应用意义。In order to improve the image quality and resolution of near-infrared image colorization, a combination of image fusion and super-resolution reconstruction is proposed to achieve near-infrared image colorization. Firstly, near-infrared images and RGB color images are input into an image fusion network with channel attention mechanism modules to obtain fused colored images. Then, the brightness of the image is enhanced through a brightness enhancement module, and finally, the image resolution is improved through image super-resolution reconstruction. Through experiments, the algorithm proposed in this paper was compared with other algorithms, and the results showed that the image quality and resolution of the colorization results in this paper were higher subjectively. Objectively, based on the KAIST dataset, comparative experiments showed that the PSNR value and SSIM value of the algorithm proposed in this paper were 20.617 and 0.882, respectively. Compared with other algorithms, the PSNR value and SSIM value increased by 137.91% and 87.66%, respectively;The PSNR value on the IVRL RGB-NIR dataset is 18.295, and the SSIM value is 0.796, which is also improved compared to other algorithms. The algorithm proposed in this paper can effectively improve the image quality and resolution of near-infrared images after colorization, and has certain practical application significance.

关 键 词：近红外图像 彩色化 融合 超分辨率 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]