Font Recognition for Image Tracing Using Transfer Learning

Lau, Chih Kei (2021) Font Recognition for Image Tracing Using Transfer Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Lau Chih Kei.pdf Restricted to Registered users only Download (3MB)

Abstract

In the current packaging and manufacturing industry, the graphic designers are still applying manual font matching to choose an identical font that matches with the original text before printing the design on corrugated boxes. This approach is a time consuming and error prone process. An automated approach is decided to be applied in order to solve the limitation of manual font matching. Therefore, an automatic font recognition and classification is proposed to classify the font style by using Convolutional Neural Network (CNN) Transfer Learning which will increase the efficiency of font matching by providing an accurate result in a short time. The development of the font matching system was partitioned into several stages which were data collection, image pre-processing, system development and system testing. The dataset was collected by using 2 approaches which were gathering real-world text images from the packaging material provided by Nixel and generating synthetic text images using Python code. The collected images will be utilized in the model training during the system development stage, a VGG16 CNN pre-trained model was chosen as the base model to train a new font classification model. In order to maximize the accuracy, transfer learning methods evaluation and selection, hyperparameters tuning and k-fold cross validation were carried out to the CNN model. The built model was implemented into the font matching system which will be used to perform font recognition in future. In the testing phase, the trained model was evaluated using synthetic test dataset and it achieved 91.26% accuracy. Besides, the model was assessed with the real-world dataset as well to test its practicality in the actual environment. However, due to the images that used to train the model were perfect resolution without distortion, hence the model could not recognize well the font in the real-world image that contained noises.