| Title |
Research on Disaster Image Generation by Disaster Type based on Inpainting |
| Authors |
최민지(Minji Choi) ; 원루빈(Ru-Bin Won) ; 최지훈(Ji Hoon Choi) ; 배병준(Byungjun Bae) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.8.3 |
| Keywords |
Deep learning; Image segmentation; Image inpainting; Disaster visualization; Image generation |
| Abstract |
Despite the availability of real-time text-based disaster alert services, disaster marginalized groups such as the visually and hearing impaired, foreigners, children, and the elderly face challenges in understanding disaster situations, leading to inadequate responses. This paper proposes a disaster image generation algorithm to enhance disaster situation awareness for these groups with limited literacy. To minimize visual representation errors of disaster situations, disasters are categorized into surface disaster, global-scale disaster, and structure-focused disaster. The algorithm generates disaster images based on image segmentation and inpainting, considering the characteristics of each disaster type. Finally, the results of disaster image generation for each disaster type are demonstrated using actual terrain images. The proposed algorithm has been validated to outperform existing image generation models by comparing FID and CLIP scores, demonstrating its superiority. This enables rapid and accurate information delivery during disasters, assisting marginalized groups in better understanding disaster situations. |