Self-Checkout System Prototype for Point-of-Sale using Image Recognition with Deep Neural Networks

Abstract

This article describes the development of an industrial prototype named Self-Checkout with Image Recognition (SCIR) developed for points of sale (POS) using lightweight convolutional neural networks to solve the problem of image verification. The development was made for a company whose POS product is installed in thousands of cashier machines in department stores, supermarkets, pharmacies and similar in Latin America. This development is intended to complement the existing self-service technology based on verification by weight of products with verification by image recognition. This optimization enhances the purchasing process through a better experience for the client. Simultaneously, the amount of fraud suffered by the retailer decreases when a client scans an article and then carries one of equal weight, but higher price. The system has the advantage of being easily integrated into an existing product. Its physical dimensions are like the self-checkouts that are in operation. The prototype recognizes 10 different products, and its precision value per class is greater than 96% in all cases while the recall value stays over 79% for each product.