Accurate identification of Oudemansiella raphanipes growth stages is crucial for understanding its development and optimizing cultivation. However, deep learning methods for this task remain unexplored. This paper introduces OR-FCOS, an enhanced fully convolutional one-stage (FCOS) approach designed to improve accuracy and efficiency in identifying these growth stages. We constructed the ORaph8K dataset, containing 8,000 images of Oudemansiella raphanipes at different growth stages, used for training and validation. The OR-FCOS uses the MobileNetV3-Large backbone with an efficient multi-scale attention (EMA) module, improving feature extraction efficiency without sacrificing accuracy. A neural architecture search (NAS)-enhanced FCOS decoder replaces both the traditional feature pyramid networks (FPN) and prediction head in FCOS, optimizing feature fusion and prediction. Integrating the complete intersection over union (CIoU) loss function addresses standard IoU limitations by factoring in aspect ratio and bounding box center distance. Channel pruning further reduces the decoder’s parameters, decreasing model size and computational requirements while maintaining precision. The enhanced algorithm achieved a mean average precision (mAP) of 89.4% (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41598_2025_9303_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="50" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {mAP}_{50}\)</EquationSource> </InlineEquation>) and 78.3% (<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41598_2025_9303_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="65" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {mAP}_{50:95}\)</EquationSource> </InlineEquation>), while the number of model parameters was reduced to 9.9 M, the model size was reduced to 40.1 MB, and the number of floating point operations per second (FLOPs) was reduced to 31.2 G. These results show that OR-FCOS accurately and efficiently identifies the growth stages of Oudemansiella raphanipes. By installing cameras in cultivation facilities, our algorithm enables automated and real-time monitoring, thereby supporting large-scale factory-based production of the fungus.

OR-FCOS: an enhanced fully convolutional one-stage approach for growth stage identification of Oudemansiella raphanipes

期刊论文
出版模式：

开放获取
发表日期：

2025年7月15日

查看全文

Runze Fang,
Huamao Huang,
Nuoyan Guo,
Haichuan Wei,
Shiyi Wang,
Haiying Hu,
Ming Liu

摘要

Accurate identification of Oudemansiella raphanipes growth stages is crucial for understanding its development and optimizing cultivation. However, deep learning methods for this task remain unexplored. This paper introduces OR-FCOS, an enhanced fully convolutional one-stage (FCOS) approach designed to improve accuracy and efficiency in identifying these growth stages. We constructed the ORaph8K dataset, containing 8,000 images of Oudemansiella raphanipes at different growth stages, used for training and validation. The OR-FCOS uses the MobileNetV3-Large backbone with an efficient multi-scale attention (EMA) module, improving feature extraction efficiency without sacrificing accuracy. A neural architecture search (NAS)-enhanced FCOS decoder replaces both the traditional feature pyramid networks (FPN) and prediction head in FCOS, optimizing feature fusion and prediction. Integrating the complete intersection over union (CIoU) loss function addresses standard IoU limitations by factoring in aspect ratio and bounding box center distance. Channel pruning further reduces the decoder’s parameters, decreasing model size and computational requirements while maintaining precision. The enhanced algorithm achieved a mean average precision (mAP) of 89.4% ( \(\hbox {mAP}_{50}\) ) and 78.3% ( \(\hbox {mAP}_{50:95}\) ), while the number of model parameters was reduced to 9.9 M, the model size was reduced to 40.1 MB, and the number of floating point operations per second (FLOPs) was reduced to 31.2 G. These results show that OR-FCOS accurately and efficiently identifies the growth stages of Oudemansiella raphanipes. By installing cameras in cultivation facilities, our algorithm enables automated and real-time monitoring, thereby supporting large-scale factory-based production of the fungus.