Yasmine Probst, University of Wollongong, Wollongong, Australia Duc Thanh Nguyen, University of Wollongong, Wollongong, Australia Wanqing Li, University of Wollongong, Wollongong, Australia Megan Rollo*, University of Newcastle, Newcastle, Australia Track: Research Presentation Topic: Mobile & Tablet Health Applications Presentation Type: Oral presentation Submission Type: Single Presentation Building: Sheraton Maui Resort Room: B - Kapalua Date: 2014-11-14 09:45 AM – 10:30 AM Last modified: 2014-11-24

If you are the presenter of this abstract (or if you cite this abstract in a talk or on a poster), please show the QR code in your slide or poster (QR code contains this URL).

Background: There is an increase in the incidence of obesity in the world and a concomitant increase in the health budget of governments as they struggle to provide healthcare for their citizens. In Australia alone, more than 70% of males and 55% of females are presently overweight costing the country $120 billion. Apart from eating nutritious food, there is a need to eat appropriate portions in order to stay healthy and avoid obesity. An innovative approach to managing people's food consumption is to educate and empower them. The prevalence of mobile phones and more recently Smartphones, has paved the way for mass education on how to manage one's food consumption at the point of need. The portability, improved efficiencies create potential for increased accuracy and provide the added benefits of improved reporting of food intake information when monitoring the impact of nutrition education programs. Using the concepts of volume computation and visual recognition creates the possibility of using a SmartPhone to automate the food identification and portion size estimation processes, key to the food record methodology identified worldwide as the most commonly employed dietary assessment method from a research perspective.
Objective: To compare a Bag of Words (BoW) model of visual recognition with an extended BoW model for visual recognition to determine the most useful parameters for correctly classifying particular food types into the appropriate food groupings.
Methods: The use of visual recognition technology requires the development of a database to underpin its processing. Further to this, in image classification, codewords are represented by image features such as shape, colour and texture for example. This research applied a BoW model to food image classification and compared it with an extended BoW model with a focus on context-based features. The classification underwent 2 phases: training and testing. The later was also required the development of a codebook following which a discriminative training exercise was undertaken using 410 colour food photographs.
Results: The most useful parameters for a food item were deemed to be the colour and texture of the food. This allowed vegetables for example to be separated from other foods on a plate. If those other foods included meats, the textural information was the most useful parameter for muscle meats in particular. The extended BoW model (r=0.57) was found to significantly improve the performance of food classification in comparison to the standard BoW model (r-0.37). The strongest recognition capability for the extended BOW model was r=0.95 for the milk category (r=0.95) likely due the limited variability of its appearance.
Conclusion: Increased algorithm training is required to allow for improved recognition and across a broader range of food images and categories.