Turkish font and character recognition with deep learning

The aim of this thesis is to recognize Turkish characters and fonts from numerical images. After the numerical image obtained for the recognition process to be realized, each closed region corresponding to the Turkish letters was separated from each other. The letter image, obtained by cutting off the boundaries, was sent to the pre-trained letter and font networks respectively and the results were taken. An interface is designed to visualize these processes. Text and font information as a result when the image that containing the text is loaded is shown in the interface. Character recognition is difficult because of the accents and dots in Turkish letters. Turkish characters such as i, j, ğ, ü and ö were found in the texts at the beginning. The body and part of the letter were found to be separate. An algorithm was combined to create separate decision mechanisms for each. In addition, a data set consisting of approximately 13,000 letter images containing a total of 38 different fonts and images on the entire Turkish letters in 227 * 227 * 3 size for this study was prepared. As a result of the tests, 42% letter recognition success and 62,6% font recognition success were achieved.