Image-to-Image Translation for Near-Infrared Image Colorization

被引：0

作者：

Kim, Hyeongyu ^{[1
]}

Kim, Jonghyun ^{[1
]}

Kim, Joongkyu ^{[1
]}

机构：

[1] Sungkyunkwan Univ, Dept Elect & Comp Engn, Suwon, South Korea

来源：

2022 INTERNATIONAL CONFERENCE ON ELECTRONICS, INFORMATION, AND COMMUNICATION (ICEIC) | 2022年

基金：

新加坡国家研究基金会;

关键词：

Image-to-image translation; Near-infrared image colorization; feature fusion module;

D O I：

10.1109/ICEIC54506.2022.9748773

中图分类号：

TM [电工技术]; TN [电子技术、通信技术];

学科分类号：

0808 ; 0809 ;

摘要：

In the existing near-infrared image colorization methods, the main obstacles are distortion of color consistency and texture. We propose a feature fusion module using multiscale channel attention module (MS-CAM) that extracts edges by Canny filter and fuses two colorized images with feature maps. The proposed method improves the PSNR of NIR2VC compared to the baseline on VSIAD datasets and output has consistency to the exterior wall of the building. Moreover, the proposed method well captures textures in objects. However, when predicting gray color like roads and cars, artifacts are occurred and the output tends to be slightly brighter. This method can also be applied to the field of monitoring such as CCTV and military facilities, as well as thermal imaging cameras with properties similar to near-infrared rays.

引用

页数：4