Open AccessEstimation of protein from the images of health drink powders
Author(s) -
P. Josephin Shermila,
A. Milton
Publication year - 2019
Publication title -
journal of food science and technology/journal of food science and technology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.656
H-Index - 68
eISSN - 0975-8402
pISSN - 0022-1155
DOI - 10.1007/s13197-019-04224-4
Subject(s) - prewitt operator , support vector machine , artificial intelligence , convolutional neural network , sobel operator , artificial neural network , computer science , random forest , pattern recognition (psychology) , deep learning , local binary patterns , image processing , regression , machine learning , histogram , statistics , image (mathematics) , mathematics , edge detection
Using new technologies to know the nutrition contents of food is the new emerging area of research. Predicting protein content from the image of food is one such area that will be most useful to the human beings because monitoring the nutrition intake has many health benefits. Patients with rare diseases like maple syrup urine disease need to be in good diet practices in order to survive. Protein intake has to be monitored for those individuals. In this paper, protein measurement of health drink powder is performed using image processing techniques. In this work food images are captured and a new database with 990 images of 9 health drink powders is created. Protein content is predicted using deep learning convolutional neural network and also using image features with linear regression. Image features like first order statistics, histogram-oriented gradient, gray level co-occurrence matrix, local binary pattern and gradient magnitude and gradient direction features obtained by applying Prewitt, Sobel and Kirsch are used to estimate the protein content. Training and testing are done using linear regression model which uses support vector machine to obtain the optimal hyper plane. A tenfold cross validation is used to improve the statistical significance of the results. Experimental results show that the protein contents are predicted with an average error of ± 2.71. Deep learning improves the prediction with an average error of ± 1.96.