Deep learning-based recognition of Chinese dishes in a waiterless restaurant

The outbreak of COVID-19 makes it danger going out dining. Building waiterless restaurants becomes meaningful and urgent. Because automate and accurate recognition of dishes is indispensable for ordering and charging self-served food, this study investigates the feasibility of a prototype system that uses deep networks for recognizing Chinese dishes in a self-serving restaurant. Specifically, ≈ 17,000 images and 45,000 instances of 28 categories of dishes are collected using a fixed micro-camera equipment, and 5 deep learning networks are explored for dish recognition. Experimental results on the prototype system reveal that 3 transferred deep learning networks achieve high performance, and the accuracy, mean average precision and recall are larger than 97.0%. Real-time, accurate dish recognition in an automate fashion is highly related to a great customer experience, and the real-life evaluation in this study suggests that transferred deep networks are promising to fulfill this task.