ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Preliminary Evaluation of 3D Machine Vision for Online Inspection and Grading of Green Asparagus

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600766.(doi:10.13031/aim.202600766)
Authors:   Zhikun Kong, Yuzhen Lu
Keywords:   Green asparagus, Post-harvest online grading, RGB-D machine vision, Multi-modal feature fusion.

Abstract. Green asparagus grading in post-harvest processing remains a labor-intensive task prone to subjective inconsistency and low throughput. This paper presents an end-to-end automated grading system for green asparagus using 3D machine vision with deep learning. A custom dataset of 320 asparagus spears was constructed, featuring pixel-level segmentation masks and commercial-grade labels (U.S. No. 1, No. 2, and Cull). For instance detection, a comparative study of YOLOv8, YOLO11, and YOLO26 across multiple scales was conducted, with YOLOv8m selected for its perfect recall (100%), high localization accuracy (mAP@50-95 of 0.843), and fast inference [178.6 frames per second (fps)]. For grade classification, a novel dual-branch RGB-D fusion network was developed, featuring a spatial cross-attention module that enables RGB features (as queries) to dynamically attend to depth geometric features (as keys/values), complemented by a multi-task auxiliary supervision strategy to prevent modality representation collapse. The network adopts a non-standard 736x256 slender input resolution to preserve the complete morphological features of asparagus spears. A flexible judgment logic and RA metric were introduced to align with practical production tolerances. Systematic ablation studies identified Swin Transformer Tiny + ResNet18 as the optimal backbone combination, achieving 93.94% relaxed accuracy (RA) and 83.33% hard accuracy (HA) on cropped asparagus instances at an inference speed of 159.4 FPS. Using full-frame inputs, the end-to-end pipeline achieved 100% detection recall and 83.33% RA on 66 independent test spears. A dual-GPU pipeline implementation achieved an end-to-end processing rate of 4.47 frames per second, with 98.4% parallel efficiency. This study demonstrates the feasibility of RGB-D fusion with spatial cross-attention for automated asparagus grading and its potential for real-time, high-throughput inspection.

(Download PDF)    (Export to EndNotes)