Development of Vertical Text Interpreter for Natural Scene Images

Ling, Ong Yi; WeiYen, Almon Chai; McCarthy, Christopher

doi:10.1109/access.2021.3121608

Cited by 6 publications

(1 citation statement)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In [23], they introduced a method for handling all categories of vertical texts and a new dataset that consists of all variants of vertical texts. They began by extracting features, which they then fed into an RNN and a connectionist temporal classification (CTC) to predict text.…”

Section: Orientation Correction With Recognition Methodsmentioning

confidence: 99%

Arbitrary Oriented Scene Text Recognizer (AOTR)

2022

IJIES

View full text Add to dashboard Cite

Recognizing arbitrary oriented text has grabbed researchers' attention to develop several algorithms due to its high complexity in scene images and real-time applications like language translators, reading text for blind people, navigation systems, and smart parking. Converting text in natural scene images into strings has extended its application to Natural Language Processing (NLP)-based applications like named entity recognition. There are only three methods for text recognition: Optical Character Recognition (OCR), conventional methods, and Neural Network (NN) models. OCR is only known to be a successful text recognizer for scanned images, and in recent years, NN models have outperformed traditional methods for scene text recognition. It is necessary to create an optimal model to address the issue of scene text irregularity. We present the NN and customized OCR models, which we tailor for arbitrarily oriented text recognition, thereby avoiding scene text irregularity. An Orientation Correction Model (OCM) was introduced to improve the recognition model. In place of the recognition model, we used OCR. Alternatively, we created another model that reads corrected text images and extracts low-level features using the convolutional neural network layer. A recurrent neural network then uses these features to recognize text. Experiments were conducted on ICDAR2015, Total Text, Art19, and Cute80 benchmark datasets. It is observed that the proposed model obtains 79.5 % accuracy and hence increases the result by 1.9 % compared to the existing method after adding the orientation correction model. Similarly, results are promising on other datasets compared to existing algorithms.

show abstract

Section: Orientation Correction With Recognition Methodsmentioning

confidence: 99%