Text Detection in Street View Images by Cascaded Convolutional Neural Networks

Po Wei Chang, Guan Xin Zeng, Po Chyi Su

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Considering traffic/shop signs in street view images convey a large amount of information such as locations of pictures taken or effects of advertisement etc., a text detection mechanism for street view images is proposed in this research. To deal with relatively complicated content of street views in urban areas, the proposed scheme consists of two major parts. First, since various interference caused by pedestrians, buildings, vehicles appearing in images will significantly affect the detection performance, a Fully Convolutional Network is employed to locate street signs. Next, another neural network, i.e., Region Proposal Network, will help to extract text lines in the identified traffic/shop signs. Both horizontal and vertical text-lines will be extracted. The experimental results show that the proposed scheme is feasible, especially in processing complex streetscape.

Original languageEnglish
Title of host publication2018 IEEE 23rd International Conference on Digital Signal Processing, DSP 2018
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781538668115
DOIs
StatePublished - 31 Jan 2019
Event23rd IEEE International Conference on Digital Signal Processing, DSP 2018 - Shanghai, China
Duration: 19 Nov 201821 Nov 2018

Publication series

NameInternational Conference on Digital Signal Processing, DSP
Volume2018-November

Conference

Conference23rd IEEE International Conference on Digital Signal Processing, DSP 2018
Country/TerritoryChina
CityShanghai
Period19/11/1821/11/18

Keywords

  • fully convolutional network
  • region proposal network
  • sign detection
  • street view
  • text detection

Fingerprint

Dive into the research topics of 'Text Detection in Street View Images by Cascaded Convolutional Neural Networks'. Together they form a unique fingerprint.

Cite this