Feature diffusion reconstruction mechanism network for crop spike head detection DOI Creative Commons
Rui Ming,

Qian Gong,

Yang Chen

et al.

Frontiers in Plant Science, Journal Year: 2024, Volume and Issue: 15

Published: Oct. 1, 2024

Introduction Monitoring crop spike growth using low-altitude remote sensing images is essential for precision agriculture, as it enables accurate health assessment and yield estimation. Despite the advancements in deep learning-based visual recognition, existing detection methods struggle to balance computational efficiency with accuracy complex multi-scale environments, particularly on resource-constrained platforms. Methods To address this gap, we propose FDRMNet, a novel feature diffusion reconstruction mechanism network designed accurately detect spikes challenging scenarios. The core innovation of FDRMNet lies its focus lightweight parameter-sharing head, which can effectively improve model while enhancing model's ability perceive shape texture.FDRMNet introduces Multi-Scale Feature Focus Reconstruction module that integrates information across different scales employs various convolutional kernels capture global context effectively. Additionally, an Attention-Enhanced Fusion Module developed interaction between map positions, leveraging adaptive average pooling convolution operations enhance critical features. ensure suitability platforms limited resources, incorporate Lightweight Parameter Sharing Detection Head, reduces parameter count by sharing weights layers. Results According evaluation experiments wheat head dataset diverse rice panicle dataset, outperforms other state-of-the-art mAP @.5 94.23%, 75.13% R 2 value 0.969, 0.963 predicted values ground truth values. In addition, frames per second parameters two datasets are 227.27,288 6.8M, respectively, maintains top three position among all compared algorithms. Discussion Extensive qualitative quantitative demonstrate significantly counting tasks, achieving higher lower complexity.The results underscore superior practicality generalization capability real-world applications. This research contributes highly efficient computationally effective solution detection, offering substantial benefits agriculture practices.

Language: Английский

FLTrans-Net: Transformer-based feature learning network for wheat head detection DOI
Samia Nawaz Yousafzai, Inzamam Mashood Nasir, Sara Tehsin

et al.

Computers and Electronics in Agriculture, Journal Year: 2024, Volume and Issue: 229, P. 109706 - 109706

Published: Dec. 3, 2024

Language: Английский

Citations

2
TasselLFANetV2: Exploring Vision Models Adaptation in Cross-Domain DOI
Zhenghong Yu, Jianxiong Ye,

Shengjie Liufu

et al.

IEEE Geoscience and Remote Sensing Letters, Journal Year: 2024, Volume and Issue: 21, P. 1 - 5

Published: Jan. 1, 2024

The datasets collected by people are always just a sampling of the real world. In this letter, we explore possibility achieving high-quality domain adaptation (DA) without explicit adaptation. As baseline, implemented significantly improved second-generation version TasselLFANet, TasselLFANetV2. This model, with indicators reaching AP 50 0.981 and R 2 0.9684, demonstrates leading performance in two typical cross-domain settings data distribution scenarios, agriculture remote sensing (RS), exhibiting strong generalization, surpassing advanced methods such as YOLOv8-UAV, PlantBiCNet, SLA, etc. We further studied combination regularization techniques feature re-mapping modules can effectively alleviate invariance model. What's more, when training set validation same, model is better, but premise that there must be proper transformation strategy. work provides new perspective for understanding solving problem difference deep learning. code, accessed at https://github.com/Ye-Sk/TasselLFANetV2.

Language: Английский

Citations

1
Oriented feature pyramid network for small and dense wheat heads detection and counting DOI Creative Commons
Junwei Yu, Weiwei Chen, Nan Liu

et al.

Scientific Reports, Journal Year: 2024, Volume and Issue: 14(1)

Published: April 6, 2024

Abstract Wheat head detection and counting using deep learning techniques has gained considerable attention in precision agriculture applications such as wheat growth monitoring, yield estimation, resource allocation. However, the accurate of small dense heads remains challenging due to inherent variations their size, orientation, appearance, aspect ratios, density, complexity imaging conditions. To address these challenges, we propose a novel approach called Oriented Feature Pyramid Network (OFPN) that focuses on detecting rotated by utilizing oriented bounding boxes. In order facilitate development evaluation our proposed method, introduce dataset named Rotated Global Head Dataset (RGWHD). This is constructed manually annotating images from Detection (GWHD) with Furthermore, incorporate Path-aggregation Balanced into architecture effectively extract both semantic positional information input images. achieved leveraging feature fusion at multiple scales, enhancing capabilities for heads. improve localization accuracy overlapping heads, employ Soft-NMS algorithm filter Experimental results indicate superior performance OFPN model, achieving remarkable mean average 85.77% detection, surpassing six other state-of-the-art models. Moreover, observe substantial improvement counting, an 93.97%. represents increase 3.12% compared Faster R-CNN method. Both qualitative quantitative demonstrate effectiveness model accurately localizing within various scenarios.

Language: Английский

Citations

1
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
MAR-YOLOv9: A multi-dataset object detection method for agricultural fields based on YOLOv9 DOI Creative Commons
Dunlu Lu, Yangxu Wang

PLoS ONE, Journal Year: 2024, Volume and Issue: 19(10), P. e0307643 - e0307643

Published: Oct. 29, 2024

With the development of deep learning technology, object detection has been widely applied in various fields. However, cross-dataset detection, conventional models often face performance degradation issues. This is particularly true agricultural field, where there a multitude crop types and complex variable environment. Existing technologies still bottlenecks when dealing with diverse scenarios. To address these issues, this study proposes lightweight, enhanced method for domain based on YOLOv9, named Multi-Adapt Recognition-YOLOv9 (MAR-YOLOv9). The traditional 32x downsampling Backbone network optimized, 16x innovatively designed. A more streamlined lightweight Main Neck structure introduced, along innovative methods feature extraction, up-sampling, Concat connection. hybrid connection strategy allows model to flexibly utilize features from different levels. solves issues increased training time redundant weights caused by neck auxiliary branch structures enabling MAR-YOLOv9 maintain high while reducing model’s computational complexity improving speed, making it suitable real-time tasks. In comparative experiments four plant datasets, improved [email protected] accuracy 39.18% compared seven mainstream algorithms, 1.28% YOLOv9 model. At same time, size was reduced 9.3%, number layers decreased, costs storage requirements. Additionally, demonstrated significant advantages detecting images, providing an efficient, adaptable solution tasks field. curated data code can be accessed at following link: https://github.com/YangxuWangamI/MAR-YOLOv9 .

Language: Английский

Citations

1
ムギ類育種での画像センシングの活用に向けた穂の検出の試み DOI Open Access
Haruki Nakamura, Goro Ishikawa,

Jun‐ichi Yonemaru

et al.

Breeding Research, Journal Year: 2024, Volume and Issue: 26(1), P. 5 - 16

Published: May 21, 2024

ムギ類の育種で行われる形質評価・選抜は多くの時間と労力を要するため,これらの高速化・自動化は非常に大きな役割をもつ.深層学習などの登場により飛躍的に発展した画像センシング技術は,画像から様々な情報を高速かつ高精度に取得することを可能にし,育種の効率化に貢献する.そこで,本研究ではこうした画像センシング技術を利用した育種の効率化を目的とし,その一例として物体検出技術を活用したムギ類の穂の検出と穂数調査方法の開発を試みた.穂の検出には,コムギ・オオムギを合わせて2,023枚の訓練画像と674枚の検証画像を供試し,YOLOv4を利用したモデルを作成した.作成した検出モデルは未学習のデータに対するmAP(mean Average Precision)が85.13%と良好な精度を示し,異なる麦種,熟期の画像に対し頑健と考えられた.作成したモデルとトラッキング技術を活用し,動画から穂数の推定を試みた.動画を用いた穂数の集計方法では,フレームあたり平均穂数と動画中のユニーク(固有)な穂の総数の2種類について,検出閾値を変えつつ検証した.その結果,閾値を0.35に設定した際のユニークな穂の総数による穂数推定が実測値と高い相関を示し,決定係数はオオムギで0.726,コムギで0.510だった.コムギ,オオムギの生産力検定試験区を対象に,穂揃い期以降の異なる3時点でこの手法により穂数の推定を行った.推定された穂数と生産力検定試験で得られた調査結果を比較したところ,相関係数は2年間の平均でオオムギでは0.499,コムギで0.337と全体の傾向としては一致していた.本研究で開発した手法は従来の目視による測定に比べて簡便であることに加えて反復間の再現性が優れていることから,ムギ類の穂数調査における省力化,高速化および高精度化に貢献できると考えられた.

Citations

0
Rapid identification of medicinal plants via visual feature-based deep learning DOI Creative Commons

Chaoqun Tan,

Long Tian, Chunjie Wu

et al.

Plant Methods, Journal Year: 2024, Volume and Issue: 20(1)

Published: May 31, 2024

Abstract Background Traditional Chinese Medicinal Plants (CMPs) hold a significant and core status for the healthcare system cultural heritage in China. It has been practiced refined with history of exceeding thousands years health-protective affection clinical treatment plays an indispensable role traditional health landscape modern medical care. is important to accurately identify CMPs avoiding affected safety medication efficacy by different processed conditions cultivation environment confusion. Results In this study, we utilize self-developed device obtain high-resolution data. Furthermore, constructed visual multi-varieties image dataset. Firstly, random local data enhancement preprocessing method proposed enrich feature representation imbalanced cropping shadowing. Then, novel hybrid supervised pre-training network expand integration global features within Masked Autoencoders (MAE) incorporating parallel classification branch. can effectively enhance capture capabilities integrating details. Besides, newly designed losses are strengthen training efficiency improve learning capacity, based on reconstruction loss loss. Conclusions Extensive experiments performed our dataset as well public Experimental results demonstrate that achieves best performance among state-of-the-art methods, highlighting advantages efficient implementation plant technology having good prospects real-world applications.

Language: Английский

Citations

0
YOLO-LF: a lightweight multi-scale feature fusion algorithm for wheat spike detection DOI
Shuren Zhou,

Shengzhen Long

Journal of Real-Time Image Processing, Journal Year: 2024, Volume and Issue: 21(4)

Published: Aug. 1, 2024

Language: Английский

Citations

0
Feature diffusion reconstruction mechanism network for crop spike head detection DOI Creative Commons
Rui Ming,

Qian Gong,

Yang Chen

et al.

Frontiers in Plant Science, Journal Year: 2024, Volume and Issue: 15

Published: Oct. 1, 2024

Introduction Monitoring crop spike growth using low-altitude remote sensing images is essential for precision agriculture, as it enables accurate health assessment and yield estimation. Despite the advancements in deep learning-based visual recognition, existing detection methods struggle to balance computational efficiency with accuracy complex multi-scale environments, particularly on resource-constrained platforms. Methods To address this gap, we propose FDRMNet, a novel feature diffusion reconstruction mechanism network designed accurately detect spikes challenging scenarios. The core innovation of FDRMNet lies its focus lightweight parameter-sharing head, which can effectively improve model while enhancing model's ability perceive shape texture.FDRMNet introduces Multi-Scale Feature Focus Reconstruction module that integrates information across different scales employs various convolutional kernels capture global context effectively. Additionally, an Attention-Enhanced Fusion Module developed interaction between map positions, leveraging adaptive average pooling convolution operations enhance critical features. ensure suitability platforms limited resources, incorporate Lightweight Parameter Sharing Detection Head, reduces parameter count by sharing weights layers. Results According evaluation experiments wheat head dataset diverse rice panicle dataset, outperforms other state-of-the-art mAP @.5 94.23%, 75.13% R 2 value 0.969, 0.963 predicted values ground truth values. In addition, frames per second parameters two datasets are 227.27,288 6.8M, respectively, maintains top three position among all compared algorithms. Discussion Extensive qualitative quantitative demonstrate significantly counting tasks, achieving higher lower complexity.The results underscore superior practicality generalization capability real-world applications. This research contributes highly efficient computationally effective solution detection, offering substantial benefits agriculture practices.

Language: Английский

Citations

0