3D 렌더링 소프트웨어를 활용한 합성 데이터 생성과 이를 통한 객체 탐지 연구 : 소나무재선충병 중심으로 = Synthetic Image Generation Using 3D Rendering Software and Object Detection Research : Focused on Pine Wilt Disease

국문 초록 (Abstract)

산림의 건강을 유지하기 위해서는 소나무재선충병(PWD, Pine Wilt Disease)과 같은 식물 질병을 조기에 발견하는 것이 매우 중요하다. 최근 딥러닝 기반 알고리즘을 산림과 농업 분야에 적용하는 연구가 활발히 진행되고 있다. 그러나, 이를 활용하기 위해서는 먼저 데이터 세트 구축 과정이 필수적이다. 본 연구에서는 딥러닝 기반의 객체 탐지 알고리즘과 함께 합성 데이터 생성 기법을 사용하여 PWD를 탐지하고자 한다. 이를 위해 3D 렌더링 소프트웨어 Blender와 Unreal Engine 5를 활용하여 합성 데이터를 생성한다. 또한, 본 연구에서는 학습 모델의 성능 개선과 합성 데이터의 유효성을 판별하기 위해 세 가지 다른 학습 전략을 적용하고 비교한다: 1) 합성 데이터와 실제 데이터를 각각 독립적으로 학습시키는 방법, 2) 두 데이터를 앙상블하여 함께 학습시키는 방법, 그리고 3) 합성 데이터를 이용해 사전 훈련된 모델을 실제 데이터로 파인튜닝하는 방법. 실험 결과, 세 번째 방법이 가장 우수한 성능을 보였다. 이 연구의 결과는 산림 분야의 딥러닝 알고리즘 훈련을 위한 대규모 데이터 구축에 도움이 될 것으로 기대한다.

다국어 초록 (Multilingual Abstract)

Maintaining the health of forests is crucial, and early detection of plant diseases like Pine Wilt Disease (PWD) is vital for this purpose. Recently, research in applying deep learning-based algorithms to forestry and agriculture has been gaining momentum. However, constructing a dataset is an essential prerequisite for utilizing these technologies. This study aims to detect PWD using deep learning-based object detection algorithms and synthetic data generation techniques. We use 3D rendering software, Blender, and Unreal Engine 5, to create synthetic data. Additionally, this research applies and compares three different training strategies to improve the learning model's performance and assess the validity of the synthetic data: 1) Training on synthetic and real data independently, 2) Ensemble learning using both data types, and 3) Pre-training the model with synthetic data and fine-tuning it with real data. Experimental results showed that the third method exhibited the best performance. The outcomes of this research are expected to assist in building large-scale datasets for training deep learning algorithms in the field of forestry.

참고문헌 (Reference)

1 C. Wang, "YOLOv7 : Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors" 7464-7475, 2022

2 W. Qiu, "UnrealCV: Connecting Computer Vision to Unreal Engine" 909-916, 2016

3 H. Kim, "Treatment for pine wilt disease, is it safe"

4 V. Grondin, "Training Deep Learning Algorithms on Synthetic Forest Images for Tree Detection" 2022

5 F. Özge Ünel, "The Power of Tiling for Small Object Detection" 582-591, 2019

6 Choi Songi, "The Dynamic Nature of Fluid Cultural Landscapes for Preservation and Management" 1 (1): 51-58, 2021

7 F. C. Akyon, "Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection" 966-970, 2022

8 G. Paulin, "Review and analysis of synthetic dataset generation methods and techniques for application in computer vision" 56 : 9221-9265, 2023

9 C. Newlands, "Procedural Generation and Rendering of Realistic, Navigable Forest Environments: An Open-Source Tool"

10 Korea Forest Service, "Plan for forecasting and controlling forest diseases and pests in 2022"

11 Korea Forest Service, "Pine Wilt Disease"

12 S. E. Ebadi, "PSP-HDRI+ : A Synthetic Dataset Generator for Pre-Training of Human-Centric Computer Vision Models" 2022

13 Z. Zou, "Object Detection in 20 Years : A Survey" 111 (111): 257-276, 2023

14 C. M. de Melo, "Next-generation deep learning based on simulators and synthetic data" 26 : 174-187, 2021

15 F. E. Nowruzi, "How much real data do we actually need: Analyzing object detection performance using synthetic and real data"

16 Korea Forest Service, "Forest Disease"

17 Hemanth G. ; Rao N. Thirupathi, "Enhancing Early Breast Cancer Detection Using Transfer Learning, Deep Feature Extraction and SVM Classification" 1 (1): 77-84, 2021

18 김동수 ; 이상명 ; 정영진 ; 최광식 ; 문일성 ; 박정규, "Emergence Ecology of Japanese Pine Sawyer, Monochamus alternatus(Coleoptera : Cerambycidae), a Vector of Pinewood Nematode, Bursaphelenchus xylophilus" 42 (42): 307-313, 2003

19 R. Yu, "Early detection of pine wilt disease using deep learning algorithms and UAV-based multispectral imagery" 497 : 2021

20 R. Zhang, "Detecting Pine Trees Damaged by Wilt Disease Using Deep Learning Techniques Applied to Multi-Spectral Images" 10 : 39108-39118, 2022

21 D. Jaswanth ; B. Dinesh Reddy, "Big Data Service Design : Collection to Analysis in the Cloud Computing Ecosystem" 1 (1): 1-12, 2021