STUDY OF MACHINE LEARNING TOOLS AND ALGORITHMS FOR RECOGNITION AND DIGITALISATION OF SALES RECEIPTS

Abstract

This article discusses the issue of processing images of sales receipts for subsequent text information extraction using OCR methods. This application is helpful for maintaining a family budget or for conducting accounting in small companies. The main problem with recognising receipts is the low quality of ink and printing paper, which is why it wrinkles and tears easily, and printed letters quickly fade. The study is based on a series of algorithms based on stepwise methods and integrated image transformation methods that can significantly improve the resulting character recognition. The step-by-step methods localise the text, carry out classification, segmentation, and text recognition, and remove the background part at each algorithm stage. Since they do not depend on the size of the dictionary, they can be used to recognise text from images regardless of its size. To solve the problem, we proposed a unique algorithm for image normalisation, which includes finding a receipt in the image, processing the resulting image area, removing shooting defects and media defects, and using a neural network to process and restore characters. We used the EAST (Efficient and Accurate Scene Text Detector) algorithm implemented using a convolutional neural network (CNN) for the text-finding process. Based on a comparison of the performance of the models in terms of their size and H-mean value, we selected the ddrnet23-slim neural network for the test images. The developed application can significantly increase the accuracy of text information recognition and, simultaneously, is small in size. The developed system recognises characters with reasonably high accuracy and shows the accuracy of the recognition result at a level of 97% and higher. The proposed system can be used: to detect and recognise characters by automatically scanning and updating invoice fields in the database; to extract text from an image and automatically convert it to digital format and update it in the database; as a tool for detecting, recognising, and understanding texts. Keywords: dataset, neural network, digital technologies, binarisation, sales receipt, classification, OCR.