Automated Checkout (ACO) systems have high requirements for accuracy and speed in real retail scenarios. The research of ACO based on image recognition technology faces great challenges due to the lack of high-quality datasets, high merchandise granularity, and high model training costs. This project adopts PP-ShiTu algorithm based on mainbody detection, metric learning, and vector retrieval, and incorporates knowledge distillation and data enhancement strategies so as to carry out the implementation of merchandise recognition capability and effectively improve the prediction speed and recognition accuracy. With the feature learning dataset of retail scenes, the method adopted in this project can effectively balance the retrieval accuracy and speed, and improve the security and stability of the recognition process. Meanwhile, this project combines AIoT to connect “cloud, edge and end”, forming an integrated and intelligent retail settlement platform.