Flower-DETR: relation-aware DETR with position priors for king and lateral flower detection in apple blossom thinning

Apple blossom thinning is a critical crop load management practice for improving fruit quality, and automated thinning shows strong potential to replace manual labor. However, a key challenge for the vision module of robotic systems is to effectively distinguish between king and lateral flowers during apple blossom detection. To address this challenge, an efficient relation-aware DETR model, termed Flower-DETR, is proposed for detecting king and lateral flowers in apple inflorescences. To capture spatial correlations among individual flower instances, a relation-assisted decoder with position relation embeddings is developed. Position relations are incorporated as attention bias into the DETR decoder to establish instance-level relation priors among flowers. Through progressive attention refinement, the model learns discriminative geometric features of flower spatial positions, which helps it distinguish between king and lateral flowers during detection. The proposed Flower-DETR achieves AP 50 scores of 0.970, 0.873 and 0.928 for inflorescences, king flowers, and lateral flowers, respectively. The overall mAP 50 and mAP 50:95 across the three categories are 0.924 and 0.770, respectively. Comparative experiments show that Flower-DETR achieves the highest mAP 50 and mAP 50:95 among all compared methods while maintaining competitive computational efficiency, with 9.664 M parameters, 26.0 GFLOPs, a model size of 156.4 MB, and an inference speed of 30 FPS. The model provides a favorable balance between detection precision and computational efficiency for multi-class flower detection in apple inflorescences. This work offers a useful reference for efficient detection of king and lateral flowers in automated blossom thinning applications.

Authors

Institutions

Publication Details

Journal
Frontiers in Plant Science
Published
2026-11-16
DOI
https://doi.org/10.3389/fpls.2026.1965633
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Flower-DETR: relation-aware DETR with position priors for king and lateral flower detection in apple blossom thinning

CHAI Xiujuan, Zhuang Li, Xue Xia, Zhaoxin Li
3 links
Frontiers in Plant Science
Smart Agriculture and AI
article

Flower-DETR: relation-aware DETR with position priors for king and lateral flower detection in apple blossom thinning

CHAI Xiujuan, Zhuang Li, Xue Xia, Zhaoxin Li
article en

Abstract

Apple blossom thinning is a critical crop load management practice for improving fruit quality, and automated thinning shows strong potential to replace manual labor. However, a key challenge for the vision module of robotic systems is to effectively distinguish between king and lateral flowers during apple blossom detection. To address this challenge, an efficient relation-aware DETR model, termed Flower-DETR, is proposed for detecting king and lateral flowers in apple inflorescences. To capture spatial correlations among individual flower instances, a relation-assisted decoder with position relation embeddings is developed. Position relations are incorporated as attention bias into the DETR decoder to establish instance-level relation priors among flowers. Through progressive attention refinement, the model learns discriminative geometric features of flower spatial positions, which helps it distinguish between king and lateral flowers during detection. The proposed Flower-DETR achieves AP 50 scores of 0.970, 0.873 and 0.928 for inflorescences, king flowers, and lateral flowers, respectively. The overall mAP 50 and mAP 50:95 across the three categories are 0.924 and 0.770, respectively. Comparative experiments show that Flower-DETR achieves the highest mAP 50 and mAP 50:95 among all compared methods while maintaining competitive computational efficiency, with 9.664 M parameters, 26.0 GFLOPs, a model size of 156.4 MB, and an inference speed of 30 FPS. The model provides a favorable balance between detection precision and computational efficiency for multi-class flower detection in apple inflorescences. This work offers a useful reference for efficient detection of king and lateral flowers in automated blossom thinning applications.

Frontiers in Plant ScienceVol. 17
Agricultural Information Institute (CN), Institute of Pomology (CN), Chinese Academy of Agricultural Sciences (CN), Ministry of Agriculture and Rural Affairs (CN)
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon