A Review of Optical Text Recognition from Distorted Scene Image

• Published in: 2022 4th International Conference on Cybernetics and Intelligent System (ICORIS) pp. 1 - 5
• Main Authors: Sumady, Oliver Oswin, Antoni, Brian Joe, Nasuta, Randy, Nurhasanah, Irwansyah, Edy
• Format: Conference Proceeding
• Language: English
• Published: IEEE 08.10.2022
• Subjects: CA-FCN; CNN; Convolutional neural networks; distorted image; Distortion; Image recognition; Integrated optics; Optical distortion; Optical imaging; PRISMA; scene text recognition; Text recognition
• DOI: 10.1109/ICORIS56080.2022.10031325

The growing number of images with text taken from a natural position increases the amount of text distortion. Some challenges come because of distortion, curvature, or blur which occur when images are taken from a natural position. Scene text recognition has made significant progress and improved in accuracy. However, issues arise from the nature of several images. This paper aims to review algorithms used for scene text recognition that focus on the accuracy and consistency of scene text recognition on various common datasets and compare them. In addition, to find the weakness and inconsistencies of various scene text recognition algorithms between different datasets. A PRISMA method flow diagram applies to conduct the review. The results show Convolutional Neural Network (CNN) is the most adopted approach to creating scene text recognition programs. The highest accuracy is the CA-FCN algorithm used for the SVT dataset. However, the consistency of algorithm performance varies from one dataset to another. Most algorithms struggled with the IC15 irregular or SVT regular dataset and performed best using the IC03 dataset.