Leveraging Libraries for a Deep Learning Based Font Recognition Model

Abstract

Motivation:

Several progresses have been made in the field of text recognition/Optical Character Recognition (OCR) in recent

years, but the task of font recognition remains a significant challenge.

ProblemStatement: With a variety of fonts available in numerous shapes and styles, and several fonts with only

minor differences in characteristics, font identification becomes complex. The most difficult part is distinguishing

the small differences between these similar fonts. Conventional OCR text recognition systems exhibit low

recognition rates and several errors, struggling to differentiate between different text fonts. This has made font

recognition a persistent challenge. Modern text recognition systems should be able to recognize text and font

information across a variety of fonts with high accuracy.

Approach/Method:

The proposed work aims to precisely recognize font shapes and styles of text in images by employing a method

based on Convolutional Neural Networks (CNN). This approach is tuned for multi-classification and involves two

primary steps. The first step involves splitting the dataset in a 3:1 ratio for training and testing purposes, followed

by the training of the CNN-based model. The second step involves feeding the model an image to identify the

font. The dataset comprises five classes: Lato, Raleway, Roboto, Sansation, and Walkway, with 100 randomly

generated string images per class.

Results and Findings:

The final model achieves an accuracy of 98.93%, with improved precision, recall, and F1 score.

Conclusion:

The proposed CNN-based method demonstrates a high accuracy in font recognition, addressing the challenges of

distinguishing between similar fonts and improving the performance of text recognition systems.