MTSC-Net: A Semi-Supervised Counting Network for Estimating the Number of Slash Pine New Shoots DOI Creative Commons
Zhaoxu Zhang, Yanjie Li,

Yue Cao

et al.

Plant Phenomics, Journal Year: 2024, Volume and Issue: 6

Published: Jan. 1, 2024

The new shoot density of slash pine serves as a vital indicator for assessing its growth and photosynthetic capacity, while the number shoots offers an intuitive reflection this density. With deep learning methods becoming increasingly popular, automated counting has greatly improved in recent years but is still limited by tedious expensive data collection labeling. To resolve these issues, paper proposes semi-supervised network (MTSC-Net) estimating shoots. First, based on mean-teacher framework, we introduce VGG19 to extract multiscale features. Second, connect local feature information with global channel features, attention fusion module introduced achieve effective fusion. Finally, map probability distribution are processed fine-grained manner through dilated convolution regression head classification head. In addition, masked image modeling strategy encourage contextual understanding features improve performance. experimental results show that MTSC-Net outperforms other models labeled percentages ranging from 5% 50%. When percentage 5%, mean absolute error root square 17.71 25.49, respectively. These findings demonstrate our work can be used efficient method provide support tree breeding genetic utilization.

Language: Английский

Research on Small-Target Detection of Flax Pests and Diseases in Natural Environment by Integrating Similarity-Aware Activation Module and Bidirectional Feature Pyramid Network Module Features DOI Creative Commons

M. Zhong,

Yue Li,

Yuhong Gao

et al.

Agronomy, Journal Year: 2025, Volume and Issue: 15(1), P. 187 - 187

Published: Jan. 14, 2025

In the detection of pests and diseases flax, early wilt disease is elusive, yellow leaf symptoms are easily confusing, pest hampered by issues such as diversity in species, difficulty detection, technological bottlenecks, posing significant challenges to efforts. To address these issues, this paper proposes a flax method based on an improved YOLOv8n model. enhance accuracy generalization capability model, first employs Albumentations library for data augmentation, which strengthens model’s adaptability complex environments enriching training samples. Secondly, terms model architecture, Bidirectional Feature Pyramid Network (BiFPN) module introduced replace original feature extraction network. Through bidirectional multi-scale fusion, ability distinguish with similar features large scale differences effectively improved. Meanwhile, integration SimAM attention mechanism enables learn information from three-dimensional channels, enhancing its perception features. Additionally, adopts EIOU loss function further optimize bounding box regression, reducing distortion boxes caused high sample variability. The experimental results demonstrate that achieves performance dataset, notable improvements mean average precision compared Finally, four-headed design, significantly enhances small targets size 4 × pixels or larger introducing new heads optimizing extraction. This not only improves but also maintains computational efficiency, providing effective technical support rapid precise possessing important practical application value.

Language: Английский

Citations

2
Rice Counting and Localization in Unmanned Aerial Vehicle Imagery Using Enhanced Feature Fusion DOI Creative Commons
Mingwei Yao, Wei Li, Li Chen

et al.

Agronomy, Journal Year: 2024, Volume and Issue: 14(4), P. 868 - 868

Published: April 21, 2024

In rice cultivation and breeding, obtaining accurate information on the quantity spatial distribution of plants is crucial. However, traditional field sampling methods can only provide rough estimates plant count fail to capture precise locations. To address these problems, this paper proposes P2PNet-EFF for counting localization plants. Firstly, through introduction enhanced feature fusion (EFF), model improves its ability integrate deep semantic while preserving shallow details. This allows holistically analyze morphology rather than focusing solely their central points, substantially reducing errors caused by leaf overlap. Secondly, integrating efficient multi-scale attention (EMA) into backbone, enhances extraction capabilities suppresses interference from similar backgrounds. Finally, evaluate effectiveness method, we introduce URCAL dataset localization, gathered using UAV. consists 365 high-resolution images 173,352 point annotations. Experimental results demonstrate that proposed method achieves a 34.87% reduction in MAE 28.19% RMSE compared original P2PNet increasing R2 3.03%. Furthermore, conducted extensive experiments three frequently used datasets. The excellent performance method.

Language: Английский

Citations

5
Enhancing multilevel tea leaf recognition based on improved YOLOv8n DOI Creative Commons
Xiaoli Tang, Li Tang, Junmin Li

et al.

Frontiers in Plant Science, Journal Year: 2025, Volume and Issue: 16

Published: March 28, 2025

In the tea industry, automated picking plays a vital role in improving efficiency and ensuring quality. Tea leaf recognition significantly impacts precision success of operations. recent years, deep learning has achieved notable advancements detection, yet research on multilevel composite features remains insufficient. To meet diverse demands picking, this study aims to enhance different categories. A novel method for generating overlapping-labeled category datasets is proposed. Additionally, Tea-You Only Look Once v8n (T-YOLOv8n) model introduced detection. By incorporating Convolutional Block Attention Module (CBAM) Bidirectional Feature Pyramid Network (BiFPN) multi-scale feature fusion, improved T-YOLOv8n demonstrates superior performance detecting small overlapping targets. Moreover, integrating CIOU Focal Loss functions further optimizes accuracy stability bounding box predictions. Experimental results highlight that proposed surpasses YOLOv8, YOLOv5, YOLOv9 mAP50, achieving increase from 70.5% 74.4% recall 73.3% 75.4%. computational costs are reduced by up 19.3%, confirming its robustness suitability complex garden environment. The detection while maintaining computationally efficient operations, facilitating practical deployment resource-constrained edge computing environments. advanced fusion data augmentation techniques, enhanced adaptability lighting conditions background variations, scenarios. contributes development smart agricultural technologies, including intelligent classification, real-time monitoring, providing new opportunities sustainability production.

Language: Английский

Citations

0
Robust soybean seed yield estimation using high-throughput ground robot videos DOI Creative Commons
Jiale Feng,

Samuel W. Blair,

Timilehin T. Ayanlade

et al.

Frontiers in Plant Science, Journal Year: 2025, Volume and Issue: 16

Published: March 31, 2025

We present a novel method for soybean [Glycine max (L.) Merr.] yield estimation leveraging high-throughput seed counting via computer vision and deep learning techniques. Traditional methods collecting data are labor-intensive, costly, prone to equipment failures at critical collection times require transportation of across field sites. Computer vision, the teaching computers interpret visual data, allows us extract detailed information directly from images. By treating it as task, we report more efficient alternative, employing ground robot equipped with fisheye cameras capture comprehensive videos plots which images extracted in variety development programs. These processed through P2PNet-Yield model, framework, where combined feature extraction module (the backbone P2PNet-Soy) regression estimate yields plots. Our results built on 2 years testing plot data-8,500 2021 650 2023. With these datasets, our approach incorporates several innovations further improve accuracy generalizability architecture, such image correction augmentation random sensor effects. The model achieved genotype ranking score up 83%. It demonstrates 32% reduction time collect well costs associated traditional estimation, offering scalable solution breeding programs agricultural productivity enhancement.

Language: Английский

Citations

0
SmartPod: An Automated Framework for High-Precision Soybean Pod Counting in Field Phenotyping DOI Creative Commons
Fei Liu, Shudong Wang, Shanchen Pang

et al.

Agronomy, Journal Year: 2025, Volume and Issue: 15(4), P. 791 - 791

Published: March 24, 2025

Accurate soybean pod counting remains a significant challenge in field-based phenotyping due to complex factors such as occlusion, dense distributions, and background interference. We present SmartPod, an advanced deep learning framework that addresses these challenges through three key innovations: (1) novel vision Transformer architecture for enhanced feature representation, (2) efficient attention mechanism the improved detection of overlapping pods, (3) semi-supervised strategy maximizes performance with limited annotated data. Extensive evaluations demonstrate SmartPod achieves state-of-the-art Average Precision at IoU threshold 0.5 (AP@IoU = 0.5) 94.1%, outperforming existing methods by 1.7–4.6% across various field conditions. This improvement, combined framework’s robustness environments, positions transformative tool large-scale precision breeding applications.

Language: Английский

Citations

0
Practical framework for generative on-branch soybean pod detection in occlusion and class imbalance scenes DOI

Kanglei Wu,

Tan Wang, Yuan Rao

et al.

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 139, P. 109613 - 109613

Published: Nov. 12, 2024

Language: Английский

Citations

2
Vision foundation model for agricultural applications with efficient layer aggregation network DOI
Jianxiong Ye, Zhenghong Yu,

Jiewu Lin

et al.

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 257, P. 124972 - 124972

Published: Aug. 10, 2024

Language: Английский

Citations

1
SPCN: An Innovative Soybean Pod Counting Network Based on HDC Strategy and Attention Mechanism DOI Creative Commons
Ximing Li, Yitao Zhuang, Jingye Li

et al.

Agriculture, Journal Year: 2024, Volume and Issue: 14(8), P. 1347 - 1347

Published: Aug. 12, 2024

Soybean pod count is a crucial aspect of soybean plant phenotyping, offering valuable reference information for breeding and planting management. Traditional manual counting methods are not only costly but also prone to errors. Existing detection-based face challenges due the crowded uneven distribution pods on plants. To tackle this issue, we propose Pod Counting Network (SPCN) accurate counting. SPCN density map-based architecture based Hybrid Dilated Convolution (HDC) strategy attention mechanism feature extraction, using Unbalanced Optimal Transport (UOT) loss function supervising map generation. Additionally, introduce new diverse dataset, BeanCount-1500, comprising 24,684 images 316 varieties with various backgrounds lighting conditions. Extensive experiments BeanCount-1500 demonstrate advantages in an Mean Absolute Error(MAE) Squared Error(MSE) 4.37 6.45, respectively, significantly outperforming current competing method by substantial margin. Its excellent performance Renshou2021 dataset further confirms its outstanding generalization potential. Overall, proposed can provide technical support intelligent management soybean, promoting digital precise agriculture general.

Language: Английский

Citations

1
MTSC-Net: A Semi-Supervised Counting Network for Estimating the Number of Slash Pine New Shoots DOI Creative Commons
Zhaoxu Zhang, Yanjie Li,

Yue Cao

et al.

Plant Phenomics, Journal Year: 2024, Volume and Issue: 6

Published: Jan. 1, 2024

The new shoot density of slash pine serves as a vital indicator for assessing its growth and photosynthetic capacity, while the number shoots offers an intuitive reflection this density. With deep learning methods becoming increasingly popular, automated counting has greatly improved in recent years but is still limited by tedious expensive data collection labeling. To resolve these issues, paper proposes semi-supervised network (MTSC-Net) estimating shoots. First, based on mean-teacher framework, we introduce VGG19 to extract multiscale features. Second, connect local feature information with global channel features, attention fusion module introduced achieve effective fusion. Finally, map probability distribution are processed fine-grained manner through dilated convolution regression head classification head. In addition, masked image modeling strategy encourage contextual understanding features improve performance. experimental results show that MTSC-Net outperforms other models labeled percentages ranging from 5% 50%. When percentage 5%, mean absolute error root square 17.71 25.49, respectively. These findings demonstrate our work can be used efficient method provide support tree breeding genetic utilization.

Language: Английский

Citations

0