Fine-grained And Lightweight AI Recognition Technology For Chinese Fast-food

With the large-scale application of computer vision technology,image recognition technology has begun to spread all over people’s daily life,and has brought great convenience.At the same time,some canteens and fast-food restaurants also gradually follow the pace of the times and use image recognition technology for dish recognition to realize automatic recognition of dishes for billing and cashiering.However,traditional machine learning and deep learning solutions limit dead restaurant menus in order to ensure the accuracy of dish recognition,which is contrary to the diversity of Chinese cuisine and the seasonal characteristics of dish ingredients,and seriously reduces the dining experience of customers,with obvious disadvantages.The current image recognition-based dish recognition cashier system for fast food and large canteens also has the following problems.(1)Chinese dishes are cooked in complex ways,with many different ingredients,and the same dish is made by different chefs with different percentages of ingredients and large differences;different dishes may also appear very similar in appearance,so how to identify the subtle differences between the targets of the dishes,so as to achieve the goal of fine differentiation,becomes a difficult point that has to be overcome.(2)The traditional dish recognition positioning model uses rectangular frame for positioning,but in the actual scene,oval and long-shaped plates are common,and the rectangular frame positioning is likely to frame the local content of the adjacent plates,which seriously affects the recognition accuracy.(3)When a new dish appears,the dish model needs to be retrained and replaced in order to have the recognition ability of the new dish.However,the model update of the existing cashier system based on image recognition can only be operated by professional system maintenance personnel,which is not convenient and cannot effectively recognize the new dish in time.Therefore,how to realize rapid training and timely replacement of dish recognition models becomes another difficult point that plagues this kind of application.To address the above issues,the main work of this paper is as follows.(1)The dish recognition is split into two independent steps,dish detection and dish classification,to simplify the model replacement process.In the dish detection stage,a lightweight dish rotation target detection network LR-Center Net is proposed,which can maintain better recognition accuracy and obtain more accurate dish location information while reducing the model volume to 1/28.3 of YOLO-v3;in the dish classification stage,a weakly supervised bilinear fine-grained dish classification network AB-CNN is proposed based on the attention mechanism in the dish classification stage,a weakly supervised bilinear finegrained dish classification network AB-CNN is proposed,which achieves 85.9% accuracy in the homemade Chinese fast food dataset Chinese Fast Food-65,and realizes the fine distinction of dishes.(2)Combining LR-Center Net network and AB-CNN network to build a fine-grained lightweight fast food dish AI recognition model,and developing a dish recognition selfcheckout system based on the model to achieve fast training and ready replacement of the dish recognition model,which has been practically applied in several restaurants and achieved satisfactory results.