딥러닝을 이용한 항공영상 기반 무허가 건축물 탐지 알고리즘 비교 연구

초록

Purpose: This study aims to identify the most effective deep learning algorithm for detecting unauthorized buildings in aerial images by comparing the performance of various image recognition models. While previous research primarily utilized YOLO-based object detection systems, those studies did not analyze algorithmic differences in depth and thus lacked insight into selecting the optimal model for practical administrative use. Methods: Using aerial images labeled with unauthorized buildings and their corresponding historical images, this study compares CNN-based AlexNet, ResNet (ResNet18, ResNet50, ResNet152), and Transformer-based Vision Transformer (ViT) models. Given the limited number of annotated images, Image Augmentation techniques were applied to expand the training dataset. Pre-trained models were fine-tuned for the task, and training was optimized using Early Stopping and a ReduceLROnPlateau scheduler to prevent overfitting. Results: Among the evaluated models, ViT achieved the highest prediction accuracy of 98.66%, completing fine-tuning in only 9 epochs. All models demonstrated high reliability, with ROC curves yielding AUC scores of 1.0. ViT’s superior performance is attributed to its global visual attention mechanism, which is particularly effective in detecting structural changes across entire aerial scenes. Conclusion: This study provides a comparative analysis of deep learning models for detecting unauthorized construction, offering valuable insights into selecting practical and accurate algorithms for urban planning and building regulation. ViT, in particular, shows strong potential for real-world deployment due to its high precision and efficient training characteristics, contributing to automated building monitoring and improved public safety.

키워드