結合深度學習和生物階層式分類架構的種子辨識系統;Seed Recognition System Combining Deep Learning with Biological Hierarchical Classification Architecture

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89859

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89859

Title:	結合深度學習和生物階層式分類架構的種子辨識系統;Seed Recognition System Combining Deep Learning with Biological Hierarchical Classification Architecture
Authors:	陳庭萱;Chen, Ting-Hsuan
Contributors:	資訊工程學系
Keywords:	辨識系統;階層式架構;深度神經網路;遷移學習
Date:	2022-08-01
Issue Date:	2022-10-04 12:02:35 (UTC+8)
Publisher:	國立中央大學
Abstract:	本論文提出一個結合深度學習和生物階層式分類架構的種子分類器，用於大類別數的種子辨識，我們先對種子影像進行影像前處理，擷取出單顆完整種子影像後，輸入階層架構的殘差神經網路，搭配生物分類法的科別、屬別、種別，作為三階層的分類流程，首先將種子影像進入科別的神經網路模型，分類出該種子的科別後，再將影像輸入進此科別的屬別層網路，最後再進入此分類器的第三層，進行種別的分類，輸出最終決策。我們以深度學習與階層式架構對含有衍生種的784類種子與560類的種子樣本進行分類實驗，分別可達73.13%與92.33%的辨識率，與單一神經網路(Resnet50)和混合式神經網路架構(Resnet50+Siamese)的12.7%和31.33%的辨識率相比，實驗結果顯示我們的方法具有明顯優勢，且階層式架構將分類流程分開的方式，能夠經由階層判斷分類錯誤的原因，彌補深度學習的不可解釋性。;This paper proposes a seed classifier that combines deep learning and biological hierarchical classification architecture for seed identification of large number of categories. Residual neural network, combined with the family, genus, and species of the biological taxonomy, as a three-level classification process, first enter the seed image into the neural network model of the family, and after classifying the family of the seed, then The image is input into the category layer network of this class, and finally enters the third layer of the classifier to classify the category and output the final decision. We use deep learning and hierarchical architecture to classify 784 types of seeds and 560 types of seed samples containing derived species, and the recognition rates can reach 73.13% and 92.33%, respectively. Compared with the recognition rate of 12.7% and 31.33% of the network architecture (Resnet50+Siamese), the experimental results show that our method has obvious advantages, and the hierarchical architecture separates the classification process, which can determine the cause of the classification error through the hierarchy, Compensate for the uninterpretability of deep learning.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	40	View/Open

社群 sharing

Loading...