A lightweight CRNN for end-to-end scene text recognition
Scene Text Recognition (STR) is a daunting task in computer vision, where starting from an image taken in any context out in the street or 'in the wild', any instances of text must be detected and its characters recognised. The advent of Convolutional Neural Networks has allowed impressive progress...
| Author: | |
|---|---|
| Format: | Master's thesis |
| Publication date: | 2022 |
| Country: | España |
| Institution: | Universitat Oberta de Catalunya (UOC) |
| Repository: | O2, repositorio institucional de la UOC |
| OAI Identifier: | oai:openaccess.uoc.edu:10609/145367 |
| Online Access: | http://hdl.handle.net/10609/145367 |
| Access Level: | Open access |
| Keyword: | optical character recognition convolutional neural networks computer vision reconocimiento óptico de caracteres redes neuronales convolucionales visión por ordenador reconeixement òptic de caràcters xarxes neuronals convolucionals visió per ordinador Optical character recognition -- TFM Reconeixement òptic de caràcters -- TFM Reconocimiento óptico de caracteres -- TFM |
| Summary: | Scene Text Recognition (STR) is a daunting task in computer vision, where starting from an image taken in any context out in the street or 'in the wild', any instances of text must be detected and its characters recognised. The advent of Convolutional Neural Networks has allowed impressive progress in this field, but many of the STR algorithms remain very heavy and computationally expensive. In this project we have developed lightweight algorithms to detect text in the wild, and to then recognise it. Starting from a very basic knowledge of TensorFlow, we have first studied well established implementations, and then built and trained a detection algorithm from scratch first; and an end-to-end detection and recognition network later. The detection algorithm has achieved remarkable results while being over three times faster than other state-of-the-art algorithms and keeping computation cost and requirements much lower. |
|---|