Initial Study on Classification of Japanese Paper by Kozo Name using EfficientNet with Digital Camera
Initial Study on Classification of Japanese Paper by Kozo Name using EfficientNet with Digital Camera
复制标题
使用 EfficientNet 和数码相机按 Kozo 名称对日本纸进行分类的初步研究
DOI：
10.1109/gcce50665.2020.9291930
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Kamiya Naoki
中科院分区：
文献类型：
--
作者：
Kitamura Tomohiro;Iwata Akiko;Urano Yuri;Zhou Yexin;Shibazaki Koji;Kamiya Naoki
对历史材料（如纸张）进行无损分析，对于阐明历史和文化至关重要。这是因为有可能利用从其他来源获得的信息来阐明一本书的起源背景，例如造纸方法和纸的纤维成分。本文提出了一种基于EfficientNet-B0的日本纸kozo贴片分类方法。在整篇文章中，我们将使用patch图像分类的结果进行整合，以估计kozo纤维的起源。结果表明，该方法对两类药材产地的平均准确率为93.1%，四类药材产地的平均准确率为87.5%。此外，利用在原始图像上重新排列位置的重构图像进行分类后得到的patch图像，对整篇论文的分类精度进行评价。因此，尽管实验次数有限，但可以在所有的宏观图像中对kozo的四个产区进行分类，而kozo是整个宏观图像的来源。在未来，我们将研究一种多尺度技术来考虑斑块图像重建宏图像中斑块之间的纤维连续性。
Nondestructive analysis of historical materials, such as paper, is vital in the elucidation of history and culture. This is because it is possible to elucidate the background of a book's origin using information obtained from other sources, such as paper making methods and paper's fiber composition. Herein, we propose a patch classification method for Japanese paper kozo using EfficientNet-B0. We integrate the results classified using patch images to estimate the origin of the kozo fiber in the entire paper. Consequently, the average success rate of our method was 93.1 % for two classifications and 87.5 % for four classifications of the production areas of kozo. In addition, the classification accuracy for the entire paper was evaluated from the patch image obtained after classification using the reconstructed image in which the position was rearranged on the original image. As a result, the classification of four production areas of kozo, from which the entire macro image was made, could be classified in all macro images, although the number of experiments was limited. In the future, we will investigate a multi-scale technique to consider the fiber continuity between patches in macro images reconstructed from patch images.