SCENE TEXT DETECTION AND RECOGNITION USING OCR AND DEEP LEARNING

Abstract

Text serves as the most potent source for high-level semantic information extraction. Comprehending natural scene text is a prominent subject in the field of computer vision. Natural scenes contain diverse specific information presented as text, applicable in various real-world applications. This study will examine all aspects of scene text comprehension while incorporating novel machine learning algorithms in a semi-pipelined or fully pipelined approach. The principal objective of this study is to develop and implement integrated algorithms that autonomously execute image processing and computer vision techniques to comprehend, rectify, and address all text-related challenges within a unified framework, culminating in a comprehensive end product. In the initial phase of this research, the YOLOv5 object detection model is utilized on the ASAYAR dataset to localize road scene text images using bounding boxes. This model has demonstrated superior efficacy, achieving an accuracy level of up to 99% on textual images. During the second phase, preprocessing techniques are implemented to enhance the quality of the image dataset through K-Means color segmentation. The improved images are subsequently processed using Maximally Stable Extremal Region (MSER), a feature region detector for text-based images. Upon identifying text regions, Optical Character Recognition (OCR) is utilized for the ultimate text recognition.
Keywords:, ,, ,