레이블 분산을 이용한 글자의 회전 각도 추정

Abstract

This paper proposed a method to increase the accuracy of the predicted orientation of text. We modified cross-entropy loss by distributing the label weight. After predicting the initial angle by using the deep-learning method, the rotation angle was more accurately predicted with a post-processing algorithm. Among the various distribution, when the label was distributed by using triangular distribution, the accuracy was the highest. The proposed model increased the accuracy of text orientation prediction, and reduced the error by using triangular distribution, and consequently increased the accuracy of final character recognition.