講演抄録/キーワード |
講演名 |
2021-12-16 15:10
Neural Network-based Local Feature Descriptors for Matching Excavated Mokkan Fragments of Various Sizes ○Trung Tan Ngo・Hung Tuan Nguyen・Masaki Nakagawa(TUAT) PRMU2021-33 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
This paper presents a method to predict whether two excavated Mokkan fragments of various sizes are from the same tablet or not using an end-to-end attention-based neural network, namely A-VLAD. The method does not require any preprocessing stages such as binarization and segmentation. It has three main parts: a local feature extractor using Convolutional Neural Network from an input image, an attention filter for key-points selection, and a generalized deep neural network-based VLAD model to aggregate the extracted key-points and form a representative vector. The whole network is trained end-to-end using the stochastic gradient descent algorithm to optimize both cross-entropy and triplet losses. In the experiments, we evaluate the proposed model on 13,205 fragments broken from 556 complete wooden tablets excavated from the Heijo-Kyo Palace ruins in the Japanese Nara period. The proposed A-VLAD model achieved mean average precision of 75.5% and top-1 accuracy of 87.9% better than the state-of-the-art methods on this Mokkan dataset. Thus, it is expected to be used to support archaeologists to assemble Mokkan fragments to recover an original Mokkan tablet. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Mokkan / Historical documents / Image retrieval / Convolution neural network / / / / |
文献情報 |
信学技報, vol. 121, no. 304, PRMU2021-33, pp. 51-56, 2021年12月. |
資料番号 |
PRMU2021-33 |
発行日 |
2021-12-09 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2021-33 |