Food region segmentation in meal images using touch points

We propose an interactive scheme for segmenting meal images for automated dietary assessment. A smartphone user photographs a meal and marks a few touch points on the resulting image. The segmentation algorithm initializes a set of food segments with the touch points, and grows them using local image features. We evaluate the algorithm with a data set consisting of 300 manually segmented meal images. The precision of segmentation is 0.87, compared with 0.70 for fully automatic segmentation. The results show that the precision of segmentation was significantly improved by incorporating minimal user intervention.