Deep Learning-Based Model for Text Font Detection and Font Detection in Contents

Abstract

Despite the diverse forms and designs in which fonts are widely utilized in our daily lives, including videos, print materials, products, websites, and mobile sites, numerous copyright issues continue to arise. Despite efforts to address these concerns and take measures for improvement, a variety of issues persist. This research serves as a preliminary study to enhance this situation. The goal was to implement a detection model for fonts used in texts within challenging media such as images and videos, aiming to mitigate font copyright issues that arise across various mediums. As a preliminary step, a model for recognizing fonts used within videos, image, pdf was developed. The model consists of two primary components: a text detection and background removal model within images, and a font recognition model used within the text. The text detection model was refined through various approaches such as image processing and deep learning to identify the optimal model. For the font recognition model, comparisons were made between CNN and ResNet models to select the most suitable one. As a result, an integrated 2-stage model was constructed. Validation was performed using arbitrary video data, revealing a top-1 rate is 76% and a top-5 rate is 94%.