JOURNAL of the JAPAN RESEARCH ASSOCIATION for TEXTILE END-USES
Online ISSN : 1884-6599
Print ISSN : 0037-2072
ISSN-L : 0037-2072
Search
Advanced search
OR
Browse
Search
Volume 63, Issue 4
Displaying 1-12 of 12 articles from this issue
Show all abstracts
    • |<
    • <
    • 1
    • >
    • >|
  • Classification of Fibers by the Appearance Shape Using Deep Learning
    kazuya OKAMOTO, Kanya KURAMOTO, Motoshi HONDA, Satoru HIROSAWA, Tetsuy ...
    2022 Volume 63 Issue 4 Pages 242-249
    Published: April 25, 2022
    Released on J-STAGE: April 25, 2022
    DOIhttps://doi.org/10.11419/senshoshi.63.4_242
    JOURNAL FREE ACCESS
    Show abstractHide abstract

    Optical microscopy is one of the methods for fiber identification. It identifies the fiber types based on their shapes which are observed from the optical microscope images. In order to use AI (artificial intelligence) for fiber identification, deep learning was applied to a fiber classification based on the shape of optical microscope images. The microscopic images of the various shapes of fibers were captured at an objective lens magnification of 20x and 40x. These images were then pre-processed into a suitable format. The images were transferred to a pre-trained network Resnet50. From the result of the 40x images, the validation accuracy reached about 97% on a randomly sorted dataset. It validated the performance of deep learning. The high performance was also obtained from the dataset consisting of both 20x and 40x images. Therefore, the images with the different magnifications can be included in the same dataset. It is suggested that the application of deep learning can be applied for textile fiber identification.

    View full abstract
    Download PDF (1182K)
  • The Influence of Incident Light Angles on Pilling Grading from Optical Image by Visual Assessment and Convolutional Neural Network
    Motoshi HONDA, Misaki MUROSE, Satoru HIROSAWA, Kanya KURAMOTO, Saori K ...
    2022 Volume 63 Issue 4 Pages 250-257
    Published: April 25, 2022
    Released on J-STAGE: April 25, 2022
    DOIhttps://doi.org/10.11419/senshoshi.63.4_250
    JOURNAL FREE ACCESS
    Show abstractHide abstract

    The influence of incident light angles and shadows on pilling grading was investigated. A pilling grade of fabric samples was evaluated from their images by visual assessments and by using a neural network. A series of images of a fabric sample was captured by changing the light-source angle with respect to a sample. Pilling grades of the images, even from the same fabric sample, was varied greatly according to the light-source angles. The greater the difference in angle, the greater the difference in grade. In addition, some samples were difficult to grade due to the light-source angle, therefore an angle of 15 or 30 degrees was considered to be appropriate for grading. The neural network was able to learn only with the low lightsource angle images that had clear shadows, and was able to predict the same level as experts in about 70％of the samples. Thus, the shadows were found to be an extremely important factor in recognizing pills for both humans and neural networks.

    View full abstract
    Download PDF (1676K)
    • |<
    • <
    • 1
    • >
    • >|
feedback
 Top