CNN-based food image segmentation without pixel-wise annotation

Abstract

We propose a CNN-based food image segmentation which requires no pixel-wise annotation. The proposed method consists of food region proposals by selective search and bounding box clustering, back propagation based saliency map estimation with the CNN model fine-tuned with the UEC-FOOD100 dataset, GrabCut guided by the estimated saliency maps and region integration by non-maximum suppression. In the experiments, the proposed method outperformed RCNN regarding food region detection as well as the PASCAL VOC detection task.