Prior knowledge-based DMV model for few-shot and multi-category wood recognition

Due to the time-consuming and labor-intensive characteristic of wood collection, especially the high cost associated with collecting precious wood, utilizing prior knowledge becomes more effective when facing limitations such as few-shot samples, multi-category samples, and unbalanced samples during...

Full description

Saved in:
Bibliographic Details
Published in:Wood science and technology 2024-07, Vol.58 (4), p.1517-1533
Main Authors: Niu, Jiashun, Zhuang, Pengyan, Wang, Bingzhen, You, Guanglin, Sun, Jianping, He, Tuo
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Biomedical and Life Sciences
Ceramics
Composites
Deep learning
Digital imaging
Effectiveness
Feature recognition
Glass
Image enhancement
Image processing
Knowledge
Life Sciences
Machine learning
Machines
Manufacturing
Natural Materials
Original
Processes
Texture recognition
Training
Wood Science & Technology
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Due to the time-consuming and labor-intensive characteristic of wood collection, especially the high cost associated with collecting precious wood, utilizing prior knowledge becomes more effective when facing limitations such as few-shot samples, multi-category samples, and unbalanced samples during recognition training. Prior knowledge is a technique that helps algorithms to adapt new data quickly, generalize better to new situations, and understand the results of learning models more effectively. In this study, the DMV (Dual-input MobileViT) model, which incorporates prior knowledge into the MobileViT model, is proposed to improve the recognition accuracy of few-shot samples of wood. The incorporation of texture features as prior knowledge in the deep learning model is motivated by their high discriminative capability in distinguishing various types of wood, supported by mature techniques and algorithms in digital image processing. This integration ultimately enhances the efficiency and accuracy of the recognition system. The effectiveness of incorporating texture features as structural prior knowledge into the model is demonstrated by a final training accuracy of 97.8% and a testing accuracy of 92%. To enhance robustness, the texture loss is weighted with the original loss function, creating a new loss function applied to the model. Extensive experiments have shown promising results, demonstrating the advantages of the proposed approach.
ISSN:0043-7719
1432-5225
DOI:10.1007/s00226-024-01581-y