Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 105 === In the era of information explosion, multimedia has become an indispensable part of modern life. People use videos and images as digital diary and create enormous image text data consequently. Texts in image usually contain informative data, and therefore scene text recognition system would be a promising application.
This thesis presents a fast scene text localization and recognition algorithm. We have develop a system that takes images as input and recognizes texts in the input images as output. The system consists of three parts: (1) Character candidate extraction, (2) Character classification and grouping, (3) Optical character recognition. In the first stage, extremal region(ER) is used as a candidate extractor. In order to reach high recall rate, we extract ER in multiple channels such as YCrCb and their inverted channels. A non-maximum suppression skill is introduced to eliminate overlapped candidates. In the second stage, we used mean local binary pattern as feature and train our classifier by AdaBoost. Text candidates are classified as one of strong text, weak text and non-text by a 2 stages classifier. The 2 stages classifier is intended to remain high recall and precision simultaneously. We then track the weak texts with strong texts as long as they have similar properties. Our next step was to group the candidates and transform them from character level to word level. Finally, our optical character recognition is done by using chain-code direction as feature and support vector machine as classifier.
The experimental results show that our system is able to detect text in real-time and recognize text in nearly real-time. In addition, the system can detect text in different text fonts and text size, also tolerate moderate rotation, blurring and inconsistent lighting. Thus, the robustness of the system is validated.
|