과실 인식을 위한 3D 디지털 팜 기반 가상 이미지 자동 생성 및 정밀 라벨링 시스템

초록

Developing artificial intelligence (AI) models for object detection, ripeness classification, and posture estimation in agriculture requires large-scale, precisely annotated datasets. However, collecting such data in real farm environments is challenging due to seasonal limitations and high costs. Existing synthetic data generation methods often suffer from insufficient realism, imprecise annotations, and limited automation. To address these issues, this study proposes a virtual image generation and automatic annotation system that utilizes three-dimensional (3D) scanned fruit models optimized in Blender and placed within a Unity-based digital farm environment. Object locations, bounding boxes, landmarks, ripeness levels, and posture information were automatically annotated using a combination of the flood fill algorithm and vector inner product calculations. Experimental results demonstrated that the proposed system reduced the data generation and annotation time by up to 97% compared to manual methods (from approximately 66 seconds to 0.23 seconds per image). In addition, the You Only Look Once (YOLO)-based object detection and pose estimation models trained on the generated synthetic data achieved high performance, with 98.0% accuracy for object detection and 97.2% for pose estimation on the synthetic dataset. Furthermore, they maintained excellent performance on real-world data, achieving 95.7% and 96.0% accuracy, respectively, thereby demonstrating strong domain generalization capability. The proposed system effectively mitigates the issue of data scarcity and supports the broader application of AI technologies in agriculture and related domains.

키워드