Image preprocessing for optical character recognition using neural networks

Primary task of this master’s thesis is to create a theoretical and practical basis of preprocessing of printed text for optical character recognition using forward-feed neural networks. Demonstration application was created and its parameters were set according to results of realized experiments. Project definition and task determination 1. Write a introduction about the problematics of optical character recognition of characters and the methods of image preprocessing before optical character recognition. 2. Design a system for image preprocessing using neural networks. 3. Implement the designed system and simulator of printed text. 4. Realize experiments to determine the settings of the system and to compare the different approaches. 5. Evaluate the realized experiments and their possible practical use. 6. Write a documentation according to the supervisor’s instructions. Introduction Almost everyone who is working with computers has to input some text to the computer from the paper. There is not only one way to do that. The smartest 1 Obr. 1: This is what you get when you scan the printed text. way is to scan the document and let software for optical character recognition (shortened: OCR) transform the scanned image into editable text. The OCR software can use methods like: matrix comparation of image with letter examples from library feature extraction from image recognition of characters using neural networks hybrid and combined methods other methods Each method listed above has some advantages and disagvantages, so if you are using OCR software which uses any of those methods, you know what you can expect. Flexibility of the methods listed above varies from one to another, but even the less flexible method’s success can be improved using image preprocessing before the OCR. The most used methods of image preprocessing before OCR are: thresholding based on histogram smoothing other 2D matrix filters I won’t describe any of these methods of preprocessing here because they’re pretty much known to the public.