Open AccessReading Aviation Wire Text in Natural Images under Assembly Workshop via Deeplearning
Author(s) -
Shufei Li,
Lianyu Zheng,
Yiwei Wang,
RenJie Zhang
Publication year - 2019
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/563/4/042075
Subject(s) - aviation , computer science , reading (process) , bounding overwatch , artificial intelligence , noise (video) , artificial neural network , computer vision , image (mathematics) , speech recognition , engineering , aerospace engineering , political science , law
Reading aviation wire text is an important step for aircraft wire assembly and also a challenging problem. Aviation wire text reading means to detect and recognize the serial number printed on the surface of aviation wires or aircraft cables. The main challenges of this task lie on small sizes, low contrast and background noise in images. In this paper, we propose a new method for aviation wire text detection and recognition via deep neural networks. The detection network utilizes text border features to roughly locate text regions in images. Then the coordinates of bounding boxes of texts are regressed by pixel-level predicted offsets in adjacent regions. The recognition network is capable to recognize the text regions with diverse lengths after padding strategies. In addition, a dataset of aviation wires captured by machine vision camera in the assembly workshop is developed. A flexible method to generate synthetic images of text whose characters are arranged in styles of both head-to-tail and left-to-right is presented. On aviation wire datasets, the proposed method for text detection and recognition achieves comparable performance and significantly outperforms previous approaches.