Research On Algorithms And Systems For Food Ingredient Recognition Based On Deep Learning

With the development of science and technology in recent years,artificial intelligence is ubiquitous and considered the core technology of the future.By 2025,the development scale of artificial intelligence is expected to exceed 190 billion US dollars.According to the "2022 Annual Report on China’s Catering Industry" statistical results,the total national catering revenue in 2021 will be 4,689.5 billion yuan,one-third of which will come from purchasing raw food materials.Under such a trend,smart catering combining artificial intelligence and traditional catering is becoming increasingly popular and a research hotspot.The identification of fresh ingredients is a vital link.This thesis designs a recognition algorithm based on the deep learning method and explores its improvement in the case of open collection.It realizes high-accuracy recognition of fresh ingredients,which can effectively reduce labour costs in the acceptance stage of the catering industry and improve the efficiency of the entire workflow.The main work and achievements of this thesis are as follows:(1)Construct a large-scale fresh ingredients dataset with 450 categories and over90,000 pictures,and preprocess the pictures.According to the scientific naming method formulated by the national standard,each picture is named and numbered,and data enhancement processing is carried out by comprehensively using cropping,flipping,rotating,brightness transformation,contrast transformation,erasing transformation,etc.We are missing data sets.(2)The problem of high-accuracy recognition of multi-category fresh ingredients is solved based on the deep learning method.At the same time,the direction of solving the problems of similarity between classes,the difference within classes and noise difference in the recognition problem is explored.This thesis selects the optimal network from a series of backbone networks such as Res Net,Dense Net,and Efficient Net.An end-to-end recognition method is designed by integrating and improving multiple attention mechanisms,such as channel and channel space,according to the data situation of this study.The algorithm achieved a top1 accuracy rate of 93.28% in the recognition task of 220 types of fresh ingredients and used the visualization method to highlight the effect of the attention mechanism.In addition,it also solves the problem of multi-category fresh food identification in the case of open collection.Aiming at the problem of new categories other than data sets appearing in practical application scenarios,open set recognition is applied to the field of food recognition for the first time.Combining the Open Max open set module with the recognition algorithm designed in this thesis,and improving the weight distribution of its inter-class distance representation method,an accuracy rate of 74.6% was achieved in the recognition task in the open set case.(3)Based on the recognition algorithm in this thesis,combined with the shooting equipment,a set of intelligent food recognition systems is designed and developed based on Spring Boot,Vue and Uni-APP framework,which is divided into web terminal and small program terminal,including user management and food image collection,Foodstuff image recognition and recognition result display and other functional modules,complete the automatic recognition of ingredients by calling the recognition algorithm designed in this thesis.The system has been put into practical use in multiple application points,with an average recognition accuracy rate of 88%and an average time saving of 60%.