Article citationsMore>>

Yi, W., Stavrinides, V., Baum, Z.M.C., Yang, Q., Barratt, D.C., Clarkson, M.J., et al. (2023) Boundary-RL: Reinforcement Learning for Weakly-Supervised Prostate Segmentation in TRUS Images. In: Cao, X., Xu, X., Rekik, I., Cui, Z. and Ouyang, X., Eds., Lecture Notes in Computer Science, Springer, 277-288.
https://doi.org/10.1007/978-3-031-45673-2_28

has been cited by the following article:

• TITLE: Multilingual Text Recognition and Assistance for Low-Resource Languages Using Computer Vision

  AUTHORS: Franck Senu Binunya, Huabing Zhou

  KEYWORDS: Long Short-Term Memory, Optical Character Recognition, Reinforced Learning

  JOURNAL NAME: Open Access Library Journal, Vol.12 No.6, June 30, 2025

  ABSTRACT: In this paper, a novel multilingual OCR (Optical Character Recognition) method for scanned papers is provided. Current open-source solutions, like Tesseract, offer extremely high accuracy when it comes to Latin letters. Nonetheless, multilingual texts using Asian characters typically have less accuracy than ones that are simply in Latin. The challenges for OCR increase when handling the logographic Chinese and Korean scripts because these languages feature complex multi-stroke characters. Text segmentation in these scripts proves challenging because their scripts lack word boundaries that Latin-based languages possess. OCR performs substantially worse at document processing when mixed English, Chinese, and Korean content exists within a single document rather than when the documents contain English-only content. The mix of complex character structures that includes no word boundaries and numerous dense character sets leads to subpar performance of current OCR systems, which process multilingual content. We provide a novel architecture that addresses these issues by using three neural blocks a segmenter and switcher as well as numerous recognizers as well as the segmenter’s reinforcement learning: Our system solves multilingual OCR challenges by implementing a segmenter to separate word images into single-character sub-images which helps minimize the difficulties of recognizing multi-stroke characters present in Chinese and Korean languages. Each sub-image goes to the switcher, which distributes it among specialized recognizers to increase accuracy due to task assignments based on character type. A new approach deals with non-Latin script word boundaries by eliminating their identification challenges. The training process of recognizers through supervised learning enhances both character recognition performance and overall output for multilingual documents. Nevertheless, there are two significant problems with the segmenter’s supervised learning: Its training necessitates a significant amount of annotation work, and its target function is not optimal. Therefore, by using the reinforcement learning method, training for the segmenter can minimise the edit distance of the final recognition results, thereby optimising overall performance. According to experimental results, the suggested approach, which does not use character boundary markers, greatly enhances performance for multilingual scripts and languages with huge character sets.