ML-DETR: Multiscale-Lite Detection Transformer for Identification of Mature Cherry Tomatoes
IEEE Transactions on Instrumentation and Measurement · 2025 · 74 · 1-18
DOI: 10.1109/tim.2025.3608317 ↗At a glance
Journal article in IEEE Transactions on Instrumentation and Measurement, 2025. Written by Zihao Tang and 4 co-authors. Cited 12 times so far (Crossref); 48 references. Funded by: Provincial-Level Quality Engineering Project with New Era Education of Anhui, National Natural Science Foundation of China.
Abstract (original language)
The YOLO family and Real-Time Detection Transformer (RT-DETR) represent the leading frameworks for real-time object detection of vegetables in precision agricultural engineering. However, both models face limitations in speed, recognition accuracy, and computational cost, restricting their application in embedded systems and hindering optimal utilization of non-maximum suppression-free approaches. To address these limitations, a novel Multiscale-Lite Detection Transformer (ML-DETR) is proposed for mature cherry tomato identification. ML-DETR consists of two modules: a novel Self-Attention Multi-Scale Collection-Allocation Mechanism (SAMS-CAM) and a Stacked Cross-Stage Feature Enhancement Network (SCFENet). Specifically, SAMS-CAM is integrated into a baseline model that combines the YOLOv8 backbone with the RT-DETR structure, leveraging convolutional and self-attention operations to improve detection accuracy while minimizing model parameters. SCFENet is designed to integrate feature maps, accommodating gradient variability and the multi-feature outputs of simple convolutions and serial linear transformations. A self-developed RE Parameterization (REP) module is incorporated into the gradient flow branch to enhance detection rates without sacrificing accuracy. Experimental results on an 8580-image dataset of cherry tomatoes demonstrate that ML-DETR achieves 91.36% accuracy, a mean Average Precision at 50% Intersection over Union (mAP50) of 0.8875, 13.9 Giga Floating-Point Operations (GFLOPs), and 81.2 frames per second (FPS), outperforming the original RT-DETR model. These findings provide essential support for real-time recognition and harvesting of ripe cherry tomatoes in precision agricultural engineering.
Funding
- Provincial-Level Quality Engineering Project with New Era Education of Anhui
- National Natural Science Foundation of China
Cite
APA
Tang, Z., Fang, L., Sun, S., Gong, Y., & Li, Q. (2025). ML-DETR: Multiscale-Lite Detection Transformer for Identification of Mature Cherry Tomatoes. IEEE Transactions on Instrumentation and Measurement, 74, 1-18. https://doi.org/10.1109/tim.2025.3608317
Vancouver
Tang Z, Fang L, Sun S, Gong Y, Li Q. ML-DETR: Multiscale-Lite Detection Transformer for Identification of Mature Cherry Tomatoes. IEEE Transactions on Instrumentation and Measurement. 2025;74:1-18. doi:10.1109/tim.2025.3608317
BibTeX
@article{Tang2025,
author = {Tang, Zihao and Fang, Liangfei and Sun, Shoujun and Gong, Yue and Li, Qing},
title = {ML-DETR: Multiscale-Lite Detection Transformer for Identification of Mature Cherry Tomatoes},
journal = {IEEE Transactions on Instrumentation and Measurement},
year = {2025},
volume = {74},
pages = {1--18},
doi = {10.1109/tim.2025.3608317}
}
EUVIMED