Optical Character Recognition Performance Comparison of CNNs and Tesseract = Convolutional Neural Networks 와 Tesseract 방법의 문자인식 성능 비교

다국어 초록 (Multilingual Abstract)

Optical Character Recognition (OCR) automatically recognizes texts in an image. OCR is still a challenging problem in computer vision. A successful solution to OCR has important scanner applications such as Text-To-Speech (TTS) conversion and automatic document classification.
In this work, we evaluate the current state-of-the-art OCR methods. One is based on convolutional neural networks (CNNs) and the other is Tesseract that is developed by HP. For this, we have designed a CNNs architecture for OCR and built our own dataset that contains upper and lower case characters. We have experimented in the presence of Salt and Pepper noise or Gaussian noise and reported their performance comparison in terms of recognition accuracy and processing time. Experimental results indicate that CNNs based OCR outperforms Tesseract in recognition accuracy but takes much more computational resources than Tesseract. In case that processing time has priority, we recommend Tesseract due to its processing speed.
In addition, CNNs is showed 97.49 % recognition accuracy using four-fold cross-validation.

목차 (Table of Contents)

• Abstract 4
• 1. Introduction 6
• 2. OCR (Optical Character Recognition) Methods 9
• 2.1. Overview of OCR 9
• 2.2. CNNs 10
• 2.3. Tesseract 21
• 2.4. OCR Database 23
• 3. Method and Performance Comparison 26
• 3.1. CNNs 26
• 3.2. Tesseract 33
• 4. Conclusion 37
• Reference 39
• Korean Abstract 42