Scholarlink

<p>In this study, a dual-branch network combined with a multi-head self-attention mechanism (MSA) is introduced to identify the geographic origin of the Fuji apple for the first time. The Fuji apple samples are collected from four different geographic origins in China. The NIR spectra data were obtained, and then the image data were obtained by the recurrence plot (RP) method. Firstly, the spectral feature extraction branch based on bidirectional gated recurrent units (Bi-GRU) is constructed. The image feature extraction branch based on the improved two-dimensional convolutional neural network (2D-CNN) is realized. Finally, the features of the spectra and image are concatenated and inputted into a MSA to compute a variety of weighted associations, automatically focusing on effective features and filtering interference. The results verify that the dual-branch network can achieve feature complementarity and enrich feature characteristics. MSA can deal with information redundancy and improve the model’s performance. The proposed method performs better than a single spectral and image model. The accuracy of training data is 0.9916, and the accuracy, precision, recall and F1 of testing data are 0.9861, 0.9833, 0.9861, and 0.9846 respectively. The proposed method can effectively improve the classification accuracy of Fuji apple geographic origin and provide a new idea for the applications of quality assurance, supply chain transparency and authentication and anti-counterfeiting in the agricultural field.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Identification of geographic origin of Fuji apple based on dual-branch deep learning model and multi-head self-attention mechanism

  • 期刊论文

  • 发表日期：

    2025年5月8日

查看全文
  • Fuqun Zou,
  • Chunhua Peng

摘要

In this study, a dual-branch network combined with a multi-head self-attention mechanism (MSA) is introduced to identify the geographic origin of the Fuji apple for the first time. The Fuji apple samples are collected from four different geographic origins in China. The NIR spectra data were obtained, and then the image data were obtained by the recurrence plot (RP) method. Firstly, the spectral feature extraction branch based on bidirectional gated recurrent units (Bi-GRU) is constructed. The image feature extraction branch based on the improved two-dimensional convolutional neural network (2D-CNN) is realized. Finally, the features of the spectra and image are concatenated and inputted into a MSA to compute a variety of weighted associations, automatically focusing on effective features and filtering interference. The results verify that the dual-branch network can achieve feature complementarity and enrich feature characteristics. MSA can deal with information redundancy and improve the model’s performance. The proposed method performs better than a single spectral and image model. The accuracy of training data is 0.9916, and the accuracy, precision, recall and F1 of testing data are 0.9861, 0.9833, 0.9861, and 0.9846 respectively. The proposed method can effectively improve the classification accuracy of Fuji apple geographic origin and provide a new idea for the applications of quality assurance, supply chain transparency and authentication and anti-counterfeiting in the agricultural field.