FVEstimator: A novel food volume estimator Wellness model for calorie measurement and healthy living

Identifying the calorific value of food requires a correct estimate of its volume and size dimensions. The food volumetric estimation can be done rationally and efficiently by measuring the food dimensions in terms of surface parameters. Food volume estimation can be effectively implemented with a computer vision-based application. The food image size can be estimated for its volumetric and calorific calibration with food area measures. However, studies in this area are limited to finding dimensions of a food item with geometrically regular, irregular, amorphous, and solid food shapes. There is a particular challenge with amorphous food items which do not have any shape and are usually calibrated with subjective container sizes by the dietitians and hence cause relative measures. Instance segmentation techniques are implemented at the pixel level and classify a pixel into a food type leading to higher accuracy in classification and segmentation of food over the background. In this work, mask-based RCNN is employed that helps accurate segmentation of food images with regular and irregular shapes in multi-food dish scenarios. The RCNN based food segmentation is applied as a volume estimator model. It is developed by fine-tuning the pre-trained ResNet model and trained over a dataset of 8 different classes of Indian breakfast food images in all shapes. The estimator model yields a precision of 90.9% for convex-shaped food images, 90.46% for amorphous food images in regular serving containers, and 98.5% to 98.9% for regular shaped (square and circle) food items. The accuracy of the presented volume estimator thus opens opportunities for further research with diverse food types and shapes.