Avaliação de técnicas para detecção de produtos e preços em folhetos digitais de supermercados

Abstract

Supermarket flyers are one of the most important ways to communicate offers, however its unestructured format makes the value comparison between products from different stores arduous. With smoothing the price comparison process in mind, Ludiiprice emerged, an application developed at UFS, under this monography’s advisors supervision, that eases the price monitoring between several comercial estabilishments. Its database is build through the collaborative digitalization of invoices and the text mining of prices published on the Internet. The present paper integrates LudiiPrice, with the finality of enabling flyer images as data sources for the application as well. In order to detect the prices and products on commercial flyers, we will analize deep learning object detection technics on images, area that showed an significantly increase in the last decades. For the detection model training, 900 images of flyers from 20 different supermarkets were labeled, the classes used were Price, Description, Image and Offer. Using the labeled data, we trained the YOLOv5 and SSD detection models and, through from a comparative analysis, we selected the first for the construction of the detector. Then, we built some rules to improve the performance of our model. In this way, the model proposed had 79% of average precision for description, 93% for image, 93% for offer and 86% for price.