The Combination of Background Subtraction and Convolutional Neural Network for Product Recognition

Tin Trung Thai, Synh Viet-Uyen Ha, Thong Duy-Minh Nguyen, Huy Dinh-Anh Le, Nhat Minh Chung, Quang Qui-Vinh Nguyen, Vuong Ai Nguyen

Abstract: Multi-class retail product recognition is an important Computer Vision application for the retail industry. Track 4 of the AICITY challenge is introduced for the retail industry. This track focuses on the accuracy and efficiency of the automatic checkout process. However, due to the lack of training data for retail items in the real world, a synthetic data set is usually generated based on the 3d scanned items to produce training data for an automated checkout system. To overcome the difference informative representative appearance between training data and the real-world scenario in the test set provided by the AICITY organizer, our research focuses on analyzing and recognizing retail items by combining the traditional method and state-of-the-art Convolutional Neural Network (CNN) approach. This paper presents our proposed system for product counting and recognition for automated retail checkout. Our proposed method is ranked top 8 in the experimental evaluation in the 2022 AI City challenge Track-4 with an F1-score 0.4082.