최근 공간정보 수집 환경 및 활용 도메인 다양화에 따라 공간정보의 중요성은 커지고 있다. 그중 항공사진은 공간정보 중에서 다양한 수요처에 사용되는데, 가시화 기반자료로 시뮬레이션과 3차원 공간정보 오픈 플랫폼 서비스에 사용될 뿐만 아니라 연구기반자료로 객체 검출 및 식생, 환경 조사에 사용된다. 항공사진을 민간, 공공에 무료로 제공해주고 있는 공간정보 오픈 플랫폼인 브이월드는 항공사진 수집과정 및 서비스 전처리 단계에 데이터 누락오류가 발생하는 문제가 있다. 이러한 누락오류를 탐지, 분석 및 재수집하기엔 많은 양의 데이터, 시간, 지역과 비용 문제 때문에 제한요소가 많다. 특히 항공사진의 누락오류는 다양한 이유로 발생하며, 가시적인 측면의 요구 및 수요처가 가시화 기반자료로 많이 사용되기에 다른 오류에 비해 누락오류 탐지 및 복원 연구가 필요하다. 최근 항공사진, 촬영 시 발생할 수 있는 오류를 보정 및 복구 연구가 이루어지고 있다. 하지만, 이러한 연구는 촬영된 지역이 다양하지 않고, 기존에 촬영되어 보관 중이거나, 서비스 중인 항공사진에 대해선 적용하기 어려운 점이 있다. 이러한 한계점을 극복하고자 본 연구에선 브이월드 항공사진 데이터를 기준으로 누락오류 탐지기법과 복구기법을 제시한다. 누락오류 탐지는 합성 곱 신경 네트워크 모델을 이용하여 누락오류와 정상을 구분한다. 누락오류 복구는 누락오류가 있는 항공사진과 관련된 참고 이미지를 이용하여 복원하는 U-Net/Ref를 사용한다. 브이월드 데이터셋 에서 실험결과 누락오류 탐지는 ResNet18 네트워크가 90.41% 분류 정확도를 보였다. 다양한 누락오류 유형에 대한 복원은 본 연구에서 제안된 U-Net/Ref와 U-Net, VAE 복원 결과를 비교하였다. 복원 결과가 원본 대비 휘도, 구조, 밝기 수치 등 얼마나 다른지 확인하기 위해 가시적 지표로서 SSIM을 측정하였고, 얼마나 왜곡되어 복원되었는지 확인하기 위해 PSNR을 계산하였다. 결과적으로 U-Net/Ref가 기존 U-Net, VAE보다 사각형 누락 에러 유형에 대해 원본과 가깝게 복구하였고, PSNR, SSIM은 기존 연구보다 11%, 6%의 성능 개선을 보여주었다.
Recently, the importance of spatial data is increasing due to the diversification of the spatial data collection environment and utilization domain. Among those various spatial data, aerial photograph is used for various demands, such as simulation, 3D spatial data open platform services, object detection, and environmental surveys as research infrastructure. VWorld, a spatial data open platform that provides aerial photographs to the private and public free of charge, has a problem of data omission during collecting and preprocessing aerial photograph. However, in huge aerial photograph datasets, there are many limitations to inspect and correct these missing error, in respect of cost and time. In particular, the missing error of aerial photographs is inevitably missed for a variety of reasons and is often used as a basis for visible aspects of demand and visualization, requiring detection and restoration studies compared to other errors. Recently, many studies have been conducted to correct and restore errors that may occur in aerial photography. However, these studies do not consider a variety in the scene, and they are difficult to apply to aerial photographs that have been taken and stored or are being serviced. To overcome these limitations, this study presents a missing error detector and recovery technique. Missing error detection uses a CNN model to distinguish images with missing error from normal images. Missing error recovery uses U-Net/Ref with a reference images to restore missing error areas in the image. Tests on VWorld datasets showed that the ResNet18 network showed 90.41% classification accuracy. Restores for various types of omissions were compared with the U-Net/Ref proposed in this study, with U-Net and VAE restore results. SSIM was measured as a visible indicator to determine how different the restore results were, such as luminance, structure, and brightness figures from the original, and PSNR was calculated to determine how distorted the restoration was. As a result, U-Net/Ref recovered closer to the original for the missing square error type than the existing U-Net and VAE, while PSNR and SSIM showed 11% and 6% performance improvement over the existing study.
Ⅰ.서 론 ································································ 11. 연구 배경 ······················································ 12.연구 목표 ······················································ 33.논문 구성 ······················································ 5Ⅱ. 선행 연구 ··························································· 61. 공간정보 ························································ 62. 딥러닝 - GeoAI ··············································· 163. 이미지 평가 ··················································· 234. 기존 선행 연구 한계점 ···································· 26Ⅲ. 항공사진 누락오류 탐지 및 복구 ··························· 301. 항공사진 누락오류 학습데이터 생성 ················· 302. 항공사진 누락오류 탐지 및 복구 설계 ··············· 333. 항공사진 누락오류 탐지 및 복구 결과 ··············· 40ⅤI. 결론 ·································································· 46참고문헌 ···························································· 48