Research On Food Image Semantic Fusion Classification Algorithm Based On Hierarchical Semantic Graph Embedding

Food image recognition is the key technology of dietary health system.Hierarchical graph embedding algorithm is one of the most important points in the field of classification algorithm research.The food image recognition algorithm combined with the hierarchical graph embedding algorithm has wide application prospects in application scenarios such as diet health and medical detection.As the country promotes dietary health to the national strategic level,the optimized food recognition algorithm embedded in the hierarchical structure diagram will show an increasingly important role in real life.Food recognition accuracy is a very important indicator,but because of the limitations of the accuracy and redundancy of the classification model,it is costly to improve the recognition accuracy.It is valuable to improve the user's acceptance by combining the classification of food thickness and classification with hierarchical information.How to make better use of hierarchical information is difficult.Therefore,the food recognition direction based on the hierarchical graph embedded structure is a very valuable research topic.This paper studies the cutting-edge hierarchical classification algorithms,and studies the top-down hierarchical classification algorithms and the difficulties encountered in image semantic fusion.Fully researched and borrowed the hierarchical semantic graph embedding algorithm and combined it with convolutional neural network.This paper designs a food recognition algorithm based on hierarchical semantic graph embedding.The main work of this thesis is as follows: In order to reduce the classification accuracy loss of hierarchical deep nodes caused by the current hierarchical classification from top to bottom,a bottom-up hierarchical classification process based on hierarchical graph embedding is proposed.On the data set proposed in this paper,the algorithm proposed in this paper improves the accuracy of fine classification Top-1 by 7.79%.This paper proposes a hierarchical search threshold to fine-tune the regulation of the coarse and fine classification ratios for different use scenarios.Weakly supervised recognition of food images requires sufficient category semantic information.Food corpora are scarce in different countries,and corpora are difficult to collect.In view of this lack of corpus,this paper adopts the method of constructing hierarchical semantic graph structure to alleviate.At present,graph embedding depends on random walk,but the embedding method based on random walk cannot sufficiently extract information from sparse graphs.Using hier-archical semantic graph embedding can effectively extract the structural information of sparse hierarchical graphs.Compared with the baseline model,the model proposed in this paper has improved by 2.70% on the Top-5 index.At the same time,combined with the hierarchical search algorithm,the weakly supervised food image classification algorithm based on the hierarchical graph embedding can be taken a step further and achieve fuzzy processing on a large scale.Under the guidance of Dr.Chen Jingjing,this paper collects and sorts out the most extensive and largest food data sets.And do corresponding research on this data set.