Research On Recognition Method Of Chinese Dishes Based On Deep Learning

In recent years,the development of artificial intelligence is getting hotter and hotter,and the research and development in the field of computer vision is getting closer to our daily life.With China entering a well-off society in an all-round way,people pay more and more attention to the problems of diet health,such as diet monitoring,nutrition analysis,food recommendation system and so on.Therefore,the dish image recognition technology based on computer vision has become one of the current research hotspots.However,at present,the research and application of dish image recognition are mostly in western food and Japanese food.Because of the variety and complexity of Chinese dishes,different dishes may be very similar,and the same dishes may be too different.Therefore,the research on the image recognition method of Chinese dishes is still a difficult problem.In order to make the research content more suitable for practical application,this paper selects Chinese restaurants which are common in daily life as application scenarios,and carries out data collection in an efficient canteen for about one year.Then,the multi-objective Chinese food recognition method is studied,and the Chinese food recognition system is designed.In this paper,the recognition of Chinese dishes images is divided into two parts: firstly,the images of dishes tray filled with multiple dishes are input into the dish position detection network to locate and extract the positions of dishes.Then,the extracted images of multiple dishes are sequentially input into the dish recognition network model,and finally,the categories of each dish are output through the recognition network.The main research contents and innovative achievements of this paper are as follows:(1)the Food-Dnet target detection network is constructed.In this detection network,Res Net-18 convolution network with shallow network depth is selected as the backbone network,and its improvement incorporates deformable convolution.At the same time,the Head part of the detection network is also introduced into the deformable residual module,which can compensate for the accuracy loss caused by reducing the depth of the network to a certain extent.In order to improve the effectiveness of extracting spatial characteristic information,the FPN structure of the Neck network is replaced by the FPN+PAN structure.The experimental results show that the Food-Dnet target detection network achieves the m AP value of 96.2% on the data set Food-C,and the detection speed is obviously superior to the mainstream target detection networks such as YOLOv3.(2)A Chinese dish recognition model RNA-TL(resnet with attention and triple-loss)based on improved residual network is proposed.Firstly,the recognition network extracts deep image semantic information by fusing multi-scale features,and then adds a layer of attention mechanism branch to pay more attention to important parts of the image.Finally,the feature vectors are input into SVM to classify the dish images,and the improved triple-loss(TL)function is used to calculate the similarity between classes.Experiments show that the recognition accuracy of RNA-TL model can reach 83.6% in Chinese food public dataset Food208 and 90.31% in dataset Food292.