Using a Convolutional Neural Network to Recognize Text Elements in Poor Quality Scanned Images

Abstract

The paper proposes a method for recognizing the content of scanned images of poor quality using convolutional neural networks (CNNs). The method involves the implementation of three main stages. At the first stage, image preprocessing is implemented, which consists of identifying the contours of its alphabetic and numeric elements and basic punctuation marks. At the second stage, the content of the image fragments inside the identified contours is sequentially fed to the input of the CNN, which implements a multiclass classification. At the third and final stage, the post-processing of the set of SNA responses and the formation of a text document with recognition results are implemented. An experimental study of all stages was carried out in Python using the Keras deep learning libraries and OpenCV computer vision and showed fairly good results for the main types of deterioration in the quality of a scanned image: geometric distortions, blurring of borders, the appearance of extra lines and spots during scanning, etc.