학술논문
딥러닝 모델 성능 향상을 위한 3차원 가상 합성 데이터 활용에 관한 연구 : 소나무재선충병 피해목 탐지 중심으로 / Research on the Utilization of 3D Virtual Synthetic Data for Enhancing Deep Learning Model Performance
Document Type
Dissertation/ Thesis
Author
Source
Subject
Language
Korean
Abstract
소나무재선충병(PWD, Pine Wilt Disease)은 산림 생태계에 큰 피해를 주는 질병으로, 치료법이 없어 조기 발견을 통한 확산 억제가 중요하다. 최근, 딥러닝 기반 알고리즘을 활용하여 PWD를 조기에 발견하려는 노력이 지속되고 있다. 그러나 데이터 수집, 주석 과정 비용은 크고 복잡하다. 특히 PWD에 대한 데이터 수집 기간의 한계로 인해 데이터셋의 다양성과 양은 제한적이다. 본 연구에서는 이 문제를 해결하기 위해 3D 렌더링 툴로 PWD를 고려한 가상 숲을 생성하였으며, 이를 통해 합성 데이터셋을 구축하였다. 더불어, 합성 데이터의 실제 데이터와의 유사성을 보장하기 위해 이미지 대 이미지(I2I, Image-to-Image) 변환 기법을 도입하였다. EfficientNetv2-S 모델을 활용하여 각 데이터셋의 결과를 분석한 결과, PWD 합성 데이터셋만을 사용한 학습 모델의 PWD 탐지 가능성을 확인하였다. 실제 데이터와 합성 데이터를 앙상블하여 학습한 모델은 F1 Score 92.88%로 성능이 향상되었으며, 이로써 본 연구에서 제안한 합성 데이터의 유효성과 실용성이 입증되었다. 이러한 결과는 다른 산림 질병 탐지 및 농업 분야에도 활용될 것으로 기대되며, 전반적인 산림 생태계 보호와 농업 관리에 기여할 것으로 예상된다.
Pine Wilt Disease (PWD) is a highly damaging disease to forest ecosystems, and with no cure, early detection is critical to contain its spread. Recently, efforts have been made to utilize deep learning-based algorithms for early detection of PWD. However, the cost of data collection and annotation is high and complex, and the variety and quantity of datasets are limited, especially due to the limited data collection period for PWD. To solve this problem, this study used a 3D rendering tool to create a virtual forest considering PWDs, which was used to build a synthetic dataset. In addition, an image-to-image (I2I) conversion technique is introduced to ensure the similarity of the synthetic data with the real data. By analyzing the results of each dataset using the EfficientNetv2-S model, we confirmed the PWD detection potential of the training model using only the PWD synthetic dataset. The model trained by combining real and synthetic data improved its performance with an F1 score of 92.88%, which proves the validity and practicality of the synthetic data proposed in this study. These results are expected to be utilized in other forest disease detection and agricultural fields, contributing to overall forest ecosystem protection and agricultural management.
Pine Wilt Disease (PWD) is a highly damaging disease to forest ecosystems, and with no cure, early detection is critical to contain its spread. Recently, efforts have been made to utilize deep learning-based algorithms for early detection of PWD. However, the cost of data collection and annotation is high and complex, and the variety and quantity of datasets are limited, especially due to the limited data collection period for PWD. To solve this problem, this study used a 3D rendering tool to create a virtual forest considering PWDs, which was used to build a synthetic dataset. In addition, an image-to-image (I2I) conversion technique is introduced to ensure the similarity of the synthetic data with the real data. By analyzing the results of each dataset using the EfficientNetv2-S model, we confirmed the PWD detection potential of the training model using only the PWD synthetic dataset. The model trained by combining real and synthetic data improved its performance with an F1 score of 92.88%, which proves the validity and practicality of the synthetic data proposed in this study. These results are expected to be utilized in other forest disease detection and agricultural fields, contributing to overall forest ecosystem protection and agricultural management.