CRNN model for text detection and classification from natural scenes

Abstract

In the emerging field of computer vision, text recognition in natural settings remains a significant challenge due to variables like font, text size, and background complexity. This study introduces a method focusing on the automatic detection and classification of cursive text in multiple languages: English, Hindi, Tamil, and Kannada using a deep convolutional recurrent neural network (CRNN). The architecture combines convolutional neural networks (CNN) and long short-term memory (LSTM) networks for effective spatial and temporal learning. We employed pre-trained CNN models like VGG-16 and ResNet-18 for feature extraction and evaluated their performance. The method outperformed existing techniques, achieving an accuracy of 95.0%, 96.3%, and 96.2% on ICDAR 2015, ICDAR 2017, and a custom dataset (PDT2023), respectively. The findings not only push the boundaries of text detection technology but also offer promising prospects for practical applications.