Optical Character Recognition using CRNN

Abstract

Optical Character Recognition (OCR) is a computer vision technique which recognizes text present in any form of images, such as scanned documents and photos. In recent years, OCR has improved significantly in the precise recognition of text from images. Though there are many existing applications, we plan on exploring the domain of deep learning and build an optical character recognition system using deep learning architectures. In the later stage, this OCR system is developed to form a web application which provides the functionalities. The approach applied to achieve this is to implement a hybrid model containing three components namely, the Convolutional Neural Network component, the Recurrent Neural Network component and the Transcription component which decodes the output from RNN into the corresponding label sequence. The process of solving problems involving text recognition required CNN to extract feature maps from images. These sequence of feature vectors undergo sequence modeling through the RNN component predicting label distributions which are later translated using the Connectionist Temporal Classification technique in the transcription layer. The model implemented acts as the backend of the web application developed using the Flask web framework. The complete application is later containerized into an image using Docker. This helps in easy deployment on the application along with its environment across any system.