Open AccessDetection of Ethnic Minority’s Symbols Based on Deep Learning
Author(s) -
Feng Ye,
Z. Shi,
Qian Kong,
Mingrui Li,
Ming–Hsuan Yang,
Mengxue Zhang,
Shuzhen Zeng
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1646/1/012033
Subject(s) - ethnic group , artificial intelligence , computer science , bounding overwatch , feature (linguistics) , symbol (formal) , pattern recognition (psychology) , pooling , sociology , linguistics , philosophy , anthropology , programming language
Symbols are the embodiments of the ideology of ethnic minorities, the detection of ethnic minority’s symbols is an important part of the protection and inheritance of ethnic cultural heritage. Nevertheless, the ethnic symbols are miscellaneous and referred to some special combination style, it is time-consuming and inaccurate to detect them only by traditional detection methods based on machine learning. Therefore, in this paper, we propose a deep learning method based on Faster R-CNN to detect various kinds of Zhuang minority’s symbols from the Z-S dataset. Firstly, the original images of Zhuang minority’s symbols are prepocessed. Secondly, the processed images are sent to the ResNet-50 with FPN to extract feature maps. Moreover, RPN processes those feature maps to generate bounding boxes. Finally, ROI pooling layer in R-CNN converts those bounding boxes into fixed-length feature vectors, which fed into two sibling fully-connected layers for further detection tasks. Through the ablation experiments based on a certain amount of ethnic symbol images, the results indicate that the proposed method has higher detection quality (mAP), and can effectively reduce the training cost.