CRNN-based Word Recognition Model for Reading Algerian Drug Labels

The recognition of text and characters in images, commonly called OCR (Optical Character Recognition) is still today a relevant field of image analysis. In this work, we tackle the text recognition problem by developing a system for reading Algerian drug labels. The processing of the label’s text starts by detecting and extracting the text regions by using MSER and SWT algorithms. The text detection procedure locates all text zones based on text properties. It selects the areas resembling letters, numbers, and some special characters, then it forms the words. The word-forming approach uses the STROCK algorithm which assembles nearby and adjacent detected text regions to determine the word regions that feed the recognition module. Our text reading model converts pixel text into string format (word or sequence of words). It uses a deep learning model consisting of a CNN, RNN, and a CTC layer to process visual data, recognize individual words from drug labels as strings of characters, and convert them into digital text. The experiments carried out on our VIG_Net dataset where our model trained on 30000 Algerian drug labels have allowed us to reach an acceptable word recognition accuracy of 63%.