Font Recognition Using CNN Approach

Chan, Kar Yong (2021) Font Recognition Using CNN Approach. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Chan Kar Yong.pdf Restricted to Registered users only Download (3MB)

Abstract

The main objective of this thesis is to carry out font recognition in an automated way by using Convolutional Neural Network (CNN) approach. The company assiociated is Nixel Tech. It has a group of talents that enable companies to make better decisions by providing solutions in various industries. The company has given a problem statement related to the manufacturing and packaging industry. There are many designers take up much time for searching the right font styles. Besides, some experienced skillsets are also needed to ensure good searching accuracy. Therefore, a task is given in order to develop an effective font recognition system which is more convenient and solve the mentioned problem in effective, efficient, and accurate manners. This project focuses on using Convolutional Neural Network (CNN) algorithms, that is differentiated into 6 main processes, which are Goal/Objectives, Data Acquisition, Data Augmentation, Deep Learning with CNN, Data Visualization and Analysis Report. Firstly, literature review and meeting was carried out to gather some domain knowledge and obtain datasets from Nixel Tech. Then, data augmentation will be performed to artificially expand the size of datasets. After the datasets are ready, a pre-trained CNN network called Xception will be imported and its fully-connected layers will be customized to carry out transfer learning on font recognition task. After that, the trained model will be used to recognize and classify each font image with correct labelling. The results will be presented in a visualized way. In order to achieve accuracy as high as possible, our team is tried the best to collect more font image samples to increase the variety of datasets before producing the final outcome of the system. In the latest update, our proposed system achieved a font recognition accuracy of up to 90.30%.