Research On Fine-Grained Food Recognition With Weak Supervision

According to the proverb,food is the primary necessity for humans,and dining is an indispensable activity in our day-to-day existence.With the pace of modern life accelerating,an increasing number of young people are inclined to dine out.The data shows that Chinese residents’ expenditure on dining out accounts for 49.59% of the total expenditure on food.While the demand for dining out is soaring,the restaurant still uses traditional manual settlement to charge,which is inefficient and slow to settle,which is easy to cause users to queue for a long time and have a poor dining experience.Therefore,it has become the consensus of the catering industry to use machine vision technology to realize automatic recognition and settlement of dishes to improve the efficiency of dining.However,the existing dish recognition technology cannot accurately identify the dishes.The first reason is that China is a big country in catering,with a wide variety of dishes,and sample data collection is difficult.Second,regional differences lead to different practices of the same dish in different regions,manufacturing process,raw material selection and other dimensions are different.(For example,Stir-fried tomato and scrambled eggs in some areas of Sichuan will be served with chilli).Third,some dishes are very similar in appearance and difficult to distinguish.The above reasons have brought great challenges to the dish recognition technology.Taking the above-mentioned issues into consideration,this paper conducts a study on weak supervision and fine-grained dish identification,and the main innovation points are as follows:(1)Aiming at the problem that the recognition accuracy of current dish recognition models for very similar dishes is not high,an improved weight amplification mechanism based on multi-head attention mechanism is proposed.Enlarge the regional weight with distinguishing ability,and let the model pay more attention to the key areas with distinguishing ability in the dishes,so as to increase the recognition ability of similar dishes.According to the ablation experiment,compared with the Vi T network using the original multi-head attention mechanism,the Vi T network using the weighted amplification multi-head attention has a higher recognition accuracy.And the probability mask module is proposed,this module adopts the idea of "randomly occluding areas with high attention weight".Enable the model to not only focus on highly recognizable key domains but also analyze global information of dishes,thus improving the recognition rate of the model.The experiments conducted on the Food1 K dataset have provided strong evidence to support this claim.(2)To further enhance the model’s capability in differentiating between similar dishes,the "feature comparison loss function FCLoss" is used,and balance the proportion of positive and negative samples by constructing positive sample pairs and modifying sampling methods.The idea of deep supervision is adopted to let FCLoss supervise the middle layer to improve the directness and transparency of the middle layer learning process and reduce the gradient explosion or gradient disappearance.And carry out the hard sample mining based on Loss,dynamically select the hard samples according to Loss during the training process,and improve the frequency of the hard samples.As a result,the model’s generalization ability and robustness have been enhanced.The experiments conducted on the Food1 K dataset have provided strong evidence to support this claim.(3)The health analysis system is built based on the dish identification technology according to the actual needs.The system dishes can identify the dishes taken by the user,analyze the nutritional composition according to the identification results,and give diet suggestions based on the user’s personal information,health status and other data.