本文二维码信息
二维码（扫一下试试看！）
Research on Classification Method for Small Target and Low Contrast Paper Defects Based on Improved YOLOv7
Received:September 07, 2024  
DOI：10.11980/j.issn.0254-508X.2025.03.018
Key Words:classification of paper defects  small target  YOLOv7  SPPFCPS  SimAM
Fund Project:国家自然科学基金（62073206）；西安市科技计划项目（2020KJRC0146）。
Author NameAffiliationPostcode
TANG Wei School of Electrical and Control Engineering， Shaanxi University of Science and Technology， Xi’an， Shaanxi Province， 710021 710021
ZHOU Guoqing* School of Electrical and Control Engineering， Shaanxi University of Science and Technology， Xi’an， Shaanxi Province， 710021 710021
WANG Mengxiao Shaanxi Ximicro Measurement and Control Engineering Co.， Ltd.， Xianyang， Shaanxi Province， 712081 712081
FANG Jianan Xi’an Jinlu Traffic Engineering Science and Technology Development Limited Liability Company， Xi’an， Shaanxi Province， 710077 710077
ZHANG Long Xi’an Xiaomi Communication Technology Co.， Ltd.， Xi’an， Shaanxi Province， 710061 710061
ZHENG Xiaohu School of Electrical and Control Engineering， Shaanxi University of Science and Technology， Xi’an， Shaanxi Province， 710021 710021
LIU Yingwei School of Electrical and Control Engineering， Shaanxi University of Science and Technology， Xi’an， Shaanxi Province， 710021 710021
Hits: 122
Download times: 49
Abstract:As the speed and width of the paper machine increasing， the frequency of paper defects rises accordingly. To effectively eliminate these defects， accurate classification is required for tracing their sources. However， due to the small size and low contrast of some paper defects， classification results are suboptimal. This paper proposed a classification method based on an improved YOLOv7 model. The core idea of the methodology was that the neck network improved the spatial pyramid pooling with fast cross stage partial connections （SPPFCSPC）， it improved classification speed without changing the receptive field. Space-to-depth non-strided convolution was used to replace the original “convolution + pooling layer” to enhance the feature extraction capability for paper defects. Additionally， by utilizing the similarity-based attention module （SimAM） directed more resources to paper defects details， thereby boosting recognition efficiency for low contrast and small target defects. The results showed that this algorithm attained a mean precision of 0.97 with real-time detection speed of 26.5 frames/s. Compared to YOLOv7， this method significantly improved both the mean precision and detection speed for small target and low contrast paper defects.
View Full Text  HTML  View/Add Comment  Download reader