Advanced 3D Depth Imaging Techniques for Morphometric Analysis of Detected On-Tree Apples Based on AI Technology

This study developed non-destructive technology for predicting apple size to determine optimal harvest timing of field-grown apples. RGBD images were collected in field environments with fluctuating light conditions, and deep learning techniques were integrated to analyze morphometric parameters. After training various models, the EfficientDet D4 and Mask R-CNN ResNet101 models demonstrated the highest detection accuracy. Morphometric metrics were measured by linking boundary box information with 3D depth information to determine horizontal and vertical diameters. Without occlusion, mean absolute percentage error (MAPE) using boundary box-based methods was 6.201% and 5.164% for horizontal and vertical diameters, respectively, while mask-based methods achieved improved accuracy with MAPE of 5.667% and 4.921%. Volume and weight predictions showed MAPE of 7.183% and 6.571%, respectively. For partially occluded apples, amodal segmentation was applied to analyze morphometric parameters according to occlusion rates. While conventional models showed increasing MAPE with higher occlusion rates, the amodal segmentation-based model maintained consistent accuracy regardless of occlusion rate, demonstrating potential for automated harvest systems where fruits are frequently partially obscured by leaves and branches.