CNN 기반 이미지 임베딩 모델을 이용한 한글 획요소 기반 유사 글꼴 추천 기법

Abstract

Font design has emerged to become more competitive in the 'Content Smog'. In this Study, we propose a technique that identifies fonts and automatically recommends the similar, using the extracted 'Stroke element' from previous studies. First, we implement an image embedding model that extracts features of an image in vector form through CNN(Convolution Neural Network). we analyze the similarity between fonts through ‘multidimensional cosine similarity calculations’ on the basis of each extracted font’s feature vector through the model. From the final analysis, the fonts of a particular character could be identified and similar fonts are automatically recommended. Since Font Recommendation is subjective, we establish the ‘Ground-Truth’ through survey. Comparing with the ground-truth, a high result close to 1(0.938). Plus, compared with character-based recommendations to verify whether the stroke element-based technique is a significant method, the results were lower(0.893). Therefore, it was verified that the technique implemented through this paper is a system that automatically recommends fonts of similar shapes that meet the subjective criteria.