An improved model based on YOLOX for detection of tea sprouts in natural environment

The tea industry occupies a pivotal and important position in China’s import and export trade commodities. With the improvement of people's quality of life, the demand for famous tea sprout is increasing. However, manual picking is inefficient and costly. Although mechanical picking can pick te...

Full description

Saved in:
Bibliographic Details
Published in:Evolving systems 2024-10, Vol.15 (5), p.1665-1679
Main Authors: Li, Xiutong, Liu, Ruixin, Li, Yuxin, Li, Zhilin, Yan, Peng, Yu, Mei, Dong, Xuan, Yan, Jianwei, Xie, Benliang
Format: Article
Language:English
Subjects:
Accuracy
Agricultural commodities
Algorithms
Artificial Intelligence
Complex Systems
Complexity
Confidence intervals
Deep learning
Engineering
Feature maps
Neural networks
Original Paper
Retina
Target detection
Transformers
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The tea industry occupies a pivotal and important position in China’s import and export trade commodities. With the improvement of people's quality of life, the demand for famous tea sprout is increasing. However, manual picking is inefficient and costly. Although mechanical picking can pick tea sprouts efficiently, it lacks selectivity, which leads to an increase in the workload of post-processing and screening of superior tea leaves. To address this, this paper establishes a dataset for tea sprouts in natural environments and proposes an improved YOLOX tea sprouts detection model, YOLOX-ST based on the Swin Transformer. The model employs the Swin Transformer as the backbone network to enhance overall detection accuracy. Additionally, it introduces the CBAM attention mechanism to address issues of miss-detection and false detections in complex environments. Furthermore, a small target detection layer is also incorporated to resolve the problem of incomplete information about tea sprout features learned from the deep feature map. To address the sample imbalance, we introduce the EIoU loss function and apply Focal Loss to the confidence level. The experimental results demonstrate that the proposed model in this paper achieves an accuracy of 95.45%, which is 5.73% higher than the original YOLOX model. Moreover, it outperforms other YOLO series models in terms of accuracy, while achieving a faster detection speed, reaching 93.2 FPS.
ISSN:1868-6478
1868-6486
DOI:10.1007/s12530-024-09589-2