Unsupervised segmentation using CNNs applied to food analysis

Abstract

[en] In the recent times, there have been numerous papers on deep segmentation algorithms for vision tasks. The main challenge of these tasks is to obtain sufficient supervised pixel-level labels for the ground truth. The main goal of this project is to explore if Convolutional Neural Networks can be used for unsupervised segmentation. We follow a novel unsupervised deep architecture, capable of facing this challenge, called the W-net and we test it on food images. The main idea of this model is to concatenate two fully convolutional networks together into an autoencoder. The encoding layer produces a k-way pixelwise prediction, and both the reconstruction error of the autoencoder as well as the error from the decoder are jointly minimized during training. We search for the best architecture for this network and we compare the results for this unsupervised network with supervised results from a well-known network.