Text recognition in radiographic weld images

Abstract

Automatic recognition of text characters on radiographic images based on computer vision would be a very useful step forward as it could improve and simplify the file handling of digitised radiographs. Text recognition in radiographic weld images is challenging since there is no uniform font or character size and each character may tilt in different directions and by different amounts. Deep learning approaches for text recognition have recently achieved breakthrough performance using convolutional neural networks (CNNs). CNNs can recognise normalised characters in different fonts. However, the tilt of a character still has a strong influence on the accuracy of recognition. In this paper, a new improved algorithm is proposed based on the Radon transform, which is very effective at character rectification. The improved algorithm increases the accuracy of character recognition from 86.25% to 98.48% in the current experiments. The CNN is used to recognise the rectified characters, which achieves good accuracy and improves character recognition in radiographic weld images. A CNN greatly improves the efficiency of digital scanning and filing of radiographic film. The method proposed in this paper is also compared with other methods that are commonly used in other fields and the results show that the proposed method is better than state-of-the-art methods.