Scene Text Detection and Recognition Using Maximally Stable Extremal Region

Abstract

In recent years, scene text detection and recognition have become important research areas in computer vision and machine learning. Traditional text detection and recognition methods may struggle with detecting and recognizing text in images with low resolution, complex backgrounds, and varying font sizes. The proposed methodology addresses these challenges by combining multiple algorithms and using deep learning techniques. In this paper, we propose a method for scene text detection based on Maximally Stable Extremal Regions (MSER) combined with Stroke Width Transform (SWT) and recognition using Convolutional Recurrent Neural Networks (CRNN). Our method consists of two stages: text detection and text recognition. To detect text, we use MSER and SWT to extract candidate text regions from the input and then, we eradicate non-text regions using image to image translation. Finally, to recognize text, CRNN is used to recognize the text present in the detected regions. Our CRNN architecture consists of convolutional and recurrent layers, which enable us to capture both spatial and temporal features of the text. The methodology is evaluated on various benchmark datasets and has obtained good results with accuracy of 96% when compared to existing methods.

Keywords

• MSER
• SWT
• Text Detection
• Text Recognition
• Deep Learning
• CRNN