Recognizing Text with a CNN
Kulsoom Mansoor, Clark F. Olson · 2019
We seek to detect text in images using multiple techniques and recognize characters using a Convolutional Neural Network (CNN). Individual characters are combined to form words, which can then be used in a variety of applications, such as automated translation. Text recognition is difficult when different types of text formats and conditions are involved, such as fonts, orientation, color, complex backgrounds, and low-quality images. Our contribution is a novel combination of techniques to perform text detection and a CNN model to classify text characters. Experiments show that both the detection algorithm and machine learning model generally succeed with clear text. The system has more difficulty detecting text from complex and low-resolution images, as well as parsing words whose characters are connected together, since this causes segmentation issues.