DACNet: A Deep Automated Checkout Network with Selective Deblurring

Abstract

Automated checkout systems have become increasingly popular as the state-of-the-art deep learning models are efficient and accurate enough for this to become a reality. However, challenges still exist due to the differences between synthetic training data and real-life products, the blurred product images captured during the checkout process, and discontinuous detections due to product similarities or tracking misses. This paper presents a robust deep learning YOLO-based pipeline, DACNet, that counters the above challenges. During training, data augmentation involving overlaying training images onto expected backgrounds creates a more diverse and accurate training dataset. When inferencing, selective deblurring is also incorporated to enhance the clarity of the items to be recognized while maintaining efficiency. And to improve accuracy further, we introduced a retrospective checking algorithm that analyzes previous detections and corrects any inaccuracies due to flickering detections or incorrect tracking results. Together, this pipeline ensures a network that produces reliable training results and high prediction accuracies even in complex retail environments with multiple items present. The proposed method has been submitted to 2023 AI City Challenge by NVIDIA and achieved a top-3 finish on the test set A with an F1-score of 0.8254. Our code is open sourced here: https://github.com/cycv5/AICityChallenge.