Аналіз та навчання YOLO-моделей для детекції та нормалізації чеків з метою підвищення точності розпізнавання тексту

Abstract

This work explores the problem of OCR models with text recognition when the text is at an angled position. To solve the problem, it is proposed to normalize the image. To do this, you need to detect the corners of the document and correct the image in perspective. The goal is to analyze and train YOLO models for finding receipt boundaries using OBB or segmentation. Both methods successfully reduce the angle deviation of the text on the receipt. OBB is slightly faster than segmentation, but segmentation better highlights the boundaries of the receipt and prospectively corrects the image.