| Title |
Learning-Based Daytime Image Simulation Improvement of
Thermal Images Using Unsupervised Segmentation |
| Authors |
원태연(Won, Taeyeon);조수민(Jo, Su Min);정지헌(Jung, Ji Heon);장명도(Jang, Meongdo);김용민(Kim, Yongmin) |
| DOI |
https://doi.org/10.12652/Ksce.2025.45.3.0403 |
| Keywords |
영상변환, 무감독 세그멘테이션, UNSB, 열적외 영상, 라이다 Image translation, Unsupervised segmentation, UNSB, Thermal image, LiDAR |
| Abstract |
Nightyime image serves as critical visual information in a wide range of applications, including disaster response and surveillance
operations. To improve the clarity and visibility of such imagery, extensive research has been conducted on imaging devices, sensor
technologies, and image processing techniques. This study proposes a method to simulate daytime images from nighttime image
acquired by various sensors by learning object-specific feature regions within the image. Considering that temperature differences
between daytime and nighttime images vary depending on the object, a CNN-based unsupervised segmentation technique, and the
UNSB (Unpaired Neural Schrodinger Bridge) model were applied. The experimental setting for the unsupervised segmentation
technique was designed to satisfy three constraints: feature similarity, spatial continuity, and a limited number of clusters. Training data
were composed by compositing thermal infrared images, LiDAR intensity images, LiDAR range images, and segmentation images,
which were then learned in conjunction with corresponding optical images using the UNSB model to simulate daytime optical image.
The experimental results showed SSIM 0.593, PSNR 14.34, and R² 0.218 based on 120 epochs of training. For each object detail, the
colors were similar to the original image, proving that the conversion result is excellent when the object features are clear. |