本文二维码信息
二维码（扫一下试试看！）
Research on Paper Defect Classification Method Based on Multi-scale Feature Fusion Attention Mechanism
Received:January 19, 2021  
DOI：10.11980/j.issn.0254-508X.2021.04.004
Key Words:multi-scale feature fusion  attentional mechanisms  convolutional neural network  image classification
Fund Project:陕西省榆林市2020年科技计划项目（CXY-2020-090）。
Author NameAffiliationPostcode
ZHANG Kaisheng School of Electrical and Control Engineering， Shaanxi University of Science & Technology， Xi’an， Shaanxi Province， 710021 710021
SONG Fan School of Electronic Information and Artificial Intelligence， Shaanxi University of Science & Technology， Xi’an， Shaanxi Province， 710021 710021
Hits: 3499
Download times: 1936
Abstract:Aiming at the problem of inaccurate classification of traditional paper defect Algorithm in paper industry， this study proposed a classification method combined with multi-scale feature fusion and attention mechanism. The method uses sharpening filter and contrast enhancement to obtain the response to straight line， and then used Sobel edge detection to obtain the response to edge. Then， the above responses were put into a convolutional neural network （CNN） separately to extract shallow local information and focuse to get global information， finally an attention mechanism was used to classify paper defect by focusing on the most characteristic parts of the image. Experiments showed that this method outperformed HOG+SVM， LBP+SVM and traditional CNN methods， achieving 96.63% classification accuracy on the self-built dataset. Compared with existing CNN-based paper defect classification algorithm， the method proposed in this study required less train data and obtained better results.
View Full Text  HTML  View/Add Comment  Download reader