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

FD-DETR: Fruit Stem and Calyx Detection Utilizing Cost-Effective 3D Information from Monocular Depth Estimation

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

Citation:  2026 ASABE Annual International Meeting  2600298.(doi:10.13031/aim.202600298)
Authors:   Kai Guo, Ruokai Guo, Xiuqin Rao
Keywords:   Automated sorting, CLAHE, Fruit stem and calyx detection, Monocular depth estimation, RT-DETR, Space-to-depth convolution.

Abstract. In the modern fruit postharvest processing and automated packaging, 2D texture confusion between surface defects and natural structures (stems and calyxes) often leads to severe misclassification. Additionally, micro-targets like thin stems are frequently missed during multi-layer feature downsampling. To address these challenges, this study proposes a precise localization network, FD-DETR, combining monocular depth estimation with a lightweight detection architecture. A lightweight Depth-UNet is constructed using knowledge distillation, integrating Edge-Aware Feature Extraction (EAFE), High-Resolution Depth Encoding (HRDE), and Gradient-Guided Smoothing (GGS) modules to reconstruct high-fidelity 3D topological features. The Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm is applied to further amplify subtle depth variations. Subsequently, an improved RT-DETR model is employed as the detector, incorporating Space-to-Depth Convolution (SPD-Conv) to mitigate the loss of fine-grained spatial information during downsampling. Experimental results demonstrate that the proposed model effectively decouples surface defects from stems and calyxes using 3D spatial priors. On enhanced depth maps, RT-DETR-SPD-Conv achieved an mAP0.5 of 0.957 and an mAP0.5~0.95 of 0.759, outperforming YOLO11 and other lightweight baselines. Generalization tests on pomelos indicated that relying solely on single-modality depth maps has limitations for fruits with inconspicuous depth variations, highlighting the need for future multimodal fusion. This monocular depth-guided strategy provides crucial visual technology support for high-precision automated fruit orientation and intelligent sorting.

(Download PDF)    (Export to EndNotes)