ResultsOur system performance is evaluated based upon theaccuracy of calculation of nutrition and calorific value. Because of the use ofpixel-based color image segmentation we use both support vector machine andfuzzy C-means as the algorithm. Since a complete food images are stored as alibrary file, the use of internet is avoided, which holds as greater advantage.Depending upon the type of diabetes, the calorific and nutritional value forthe patient is pre-fed and hence the output that the food is recommended or notis based upon those values.
Since deep learning concept is used unsupervisedlearning is done.In-order to minimize the complexity, we convert the RGB valuesto gray scale, thus increasing the processing power. The use of SIFT featuring at2 different scales results in bestperformance. Since it is a mobile app it becomes handy for the user and can useit anywhere and at anytime. It results in good user interface with outputsaligned perfectly.
DiscussionThe results are clear with modestnumber of food, the proposed technique is effective. The app is not yet released in thepublic since there are two reasons. First, by improving more computation atmore scales. As for now we have limited the scales due to computationalreasons.
Second, for the app to be more realistic huge amount of differenttypes of food has to be store. Scaling up to thousands of foods requires the developmentof new approaches to food ontology by incremental approach in upgrading the ontology anytime and anywhere whena new food is added.ConclusionsWe have developed a mobile app for recognition of nutritional andcalorific content in the food consumed by a diabetes patient.The application recognizesthe food item present only on the plateand estimates their calorific and nutrition content automatically without anyuser intervention.
To identify the food items, the user simply needs to take asnapshot of the food plate. The system automaticallydetects the salient regions base on the algorithm. Hierarchical segmentationis used to segment the images into lone region. By using SVM classifiers theapplication extracts the features at different scales and locations,classifying these region. The application crops the part of food alone from thebackground and starts to estimate the calorific and nutrition content. We haveimplemented this experiment as an Android smartphone application.
our experiments, we have achieved.Over 88% ofaccuracy in determining 16 food items. For the future work, we have planned tomake this application for personalizing the food base on user habits, health issues and othermeta-data.