Korean Text Recognition Using Probability Score and Candidate Words

Abstract

Many new proposals for Scene Text Recognition models have been introduced in recent years. Recognizing text in natural scenes has received great attention in the recognition field and industry in recent years because of its importance and challenges. The text recognition system is a technology used in the field of artificial intelligence that requires automation, such as unmanned robots and autonomous vehicles, and refers to accurately recognizing characters even in the presence of various obstacles in the surrounding environment. Scene Text Recognition remains challenging because of many factors such as complex backgrounds, various fonts, and variable imaging conditions. Language transfer plays an important role in how humans detect and perceive text in the world. Current scene text recognition use a dictionary to improve recognition performance, but there are many limitations to the current approach of converting the output to dictionary words based purely on edit distance. In this paper, we present a novel approach to generate a list of candidate words of possible outcomes and find the words most compatible with the visual appearance of the text. The proposed method leads to a robust scene text recognition model that better handles the ambiguous cases occurring in the world and improves the overall performance of the state-of-the-art scene text discovery frameworks. We present the model that can recognize a character including Korean, English, numbers, and even special characters. It combines the network deep learning techniques such as Convolution Neural Network and Recurrent Neural Network was used together. By creating a list of candidate words, it is possible to improve the problem of character misrecognition. Well as character recognition was performed by selecting only the class with the highest probability value in the previous study, the candidate word list is created in this study by considering the probability value of the next rank. For Korean recognition, various categories of Korean dataset provided by the AIhub platform is used and the data includes an image of various categories, such as Korean signboards, traffic signs, and objects captured outdoors. We validate the performance of our model in English and Korean. As a result, it achieves state-of-the-art performance on the multi-language text recognition model and comparable results on Korean text datasets. |글자 인식 시스템은 무인 로봇, 자율 주행 자동차 등 자동화를 필요로 하는 인공지능 분야에서 사용되는 기술로, 주변 환경에 여러 장애물이 있음에도 글자를 정확하게 인식하는 것을 말한다. 영어만 인식했던 기존의 연구와 달리, 본 논문은 영어, 한국어, 특수문자와 숫자를 포함한 다양한 문자가 혼재되어 있는 경우에도 강한 인식률을 보여준다. 가장 높은 확률 값을 갖는 클래스 하나 만을 선택하는 것이 아닌 차 순위의 확률도 함께 고려하여 후보 단어 리스트를 생성하고, 이로 인해 기존에 오인식되는 단어를 교정할 수 있는 방법을 제안한다.