Pasta Image Recognition System Based On Deep Learning

According to the 2018 global obesity statistics,more than 2 billion adults are plagued by obesity,which is mainly caused by the excessive calorie intake from unhealthy eating habits.Therefore,accurately recording calorie intake in people,s daily diet is critical to weight control.The latest research shows hat the increasing amount of a person's carbohydrate,which is the main ingredient of the pasta,will fasten the rising speed of his or her blood sugar level.For the plentiful diabetic patients in China,they need to effectively estimate the calorie of the pasta,so as to select the dose of insulin more accurately and control the blood sugar in the body effectively.Currently,the food image captured by the mobile terminal is mainly used to improve the accuracy of the dietary evaluation,but there is no related pasta dataset in the market.Pasta as a basic food in daily life,plays an important role in people's Iife.Besides,the amount of calories in different types of pasta varies greatly.In order to solve the current problem of the lacking Chinese pasta datasets in the market,this thesis will introduce a Chinese pasta image dataset called“ChineseNoodleFood,,which can be used for calorie estimation of Chinese pasta to help people arrange their meals reasonablyThe main work of this paper includes three parts:dataset establishment,pasta type identification and pasta calorie estimation.(1)The establishment of the dataset is achieved in two steps.The first step is to get the main 60 categories of pasta through the market research.Then the image data is collected through the web crawler and cleaning data and setting labels.The third step is to increase the number of images through image enhancement methods and improve the diversity of data sets through the BEGAN.Finally,more than 189,600 images are obtained.(2)Pasta species identification:based on a number of excellent deep learning models for initial training,the optimal inception v3 and Inception_Resnet_v2 models are selected for the parameter adjustment,and the image enhancement method is used to improve the recognition accuracy.The top-1 recognition accuracy rate reaches 92.5%.The top-5 recognition accuracy rate is over 99%.(3)Estimation of calorie calories:in order to obtain the calorie information more accurately and intuitively,this paper uses the binocular camera to make a rough measurement of the food volume,and derives the calorie information of the food according to the density and calorie conversion formula.This method ean more accurately estimate the true heat of the pasta.This thesis also proposes two points for the future research,which are to optimize the network model for higher top-1 recognition efficiency and to try to implement recognition algorithms on mobile devices and calculate more accurate food volumes to derive calorie information.