碩士 / 國立交通大學 / 電控工程研究所 / 107 / The development of computer vision technology today has been changed due to the rise of deep learning. What's more, the application of this technology had impressed lots of different aspects including object and text recognitions. Both object and text recognitions have also been applicable to both academic and industrial aspects. The text spotting applied on robot technology is an even more important development nowadays particularly in transmitting information through words in the human environment. Thus the text spotting can accelerate the integration of robotics into human civilization and increase human welfare if the robot has strong text spotting ability to help the human when required.
Regards the development of the text spotting, the current mainstream is using vocabulary as a unit, and the identification method is based on the Convolution Neural Network, which uses large amount of data to train the text spotting model.However, no matter how big the model is, it's quite difficult to include all the necessary words inside for the robot. Therefore, how to effectively add new words into the large recognition model and turn it into a new task model will be a meaningful subject worth study.On the other hand, the old process of Scene Text Spotting includes Text Detection and Text Recognition. The Text Detection is still a needed field requiring for more research efforts, even though there has been a new deep learning method to enhance the text recognition.The quality of Text Detection will directly affect the outcome of the Text Recognition. Traditionally, information is mostly obtained by image feature region, but it is susceptible to background complexity, light, and viewing angle.
In response to the two challenges above, this paper will present three methods of training: Joint Training, Fine-tuning and Feature Extraction. By adding new navigation street sign to the existing large vocabulary model data, we can analyze the recognition accuracy and the required learning time from the three methods for both the existing text and the new task text. After that, the model will be pruned and compressed in order to fit it into a resource constrained self-driving robot. And, This thesis will also modify the model from the paper of Mr. Chuang, who is a senior of my laboratory. The model with only the ability recognizing the new navigation street sign will be modified into Fully Convolutional Networks, it will combine the two processes of Text Detection and Text Recognition into one single process to achieve end-to-end text spotting. Finally, both new text task model and the end-to-end text spotting model will be deployed into the self-driving robot, and verified it with autonomous navigation tasks.
| Identifer | oai:union.ndltd.org:TW/107NCTU5449006 |
| Date | January 2018 |
| Creators | Hung, Chen-Hao, 洪禎浩 |
| Contributors | Wang, Hsueh-Cheng, 王學誠 |
| Source Sets | National Digital Library of Theses and Dissertations in Taiwan |
| Language | zh-TW |
| Detected Language | English |
| Type | 學位論文 ; thesis |
| Format | 53 |
Page generated in 0.0017 seconds