PMC

Four example images from the real-world testing dataset and their corresponding class label outputs from the NutriNet model.

A diagram of the deep learning training process, including the online training component, which keeps the model updated.

Illustration of the NutriNet architecture used on an image from the recognition dataset with a few example class labels as the output.

Example images from the two classes of the food and drink image detection dataset, obtained by merging recipe website images and a subset of the ImageNet dataset.

Example images from the final food and drink image recognition dataset, built from Google image searches. Each one of these images represents a different food or drink class.

Visual representation of the loss results from . Similarly to , the number 512 indicates a model that accepts 512 × 512 pixel images as input.

Visual representation of the classification accuracy results from . The number 512 at the end of some deep learning architecture names indicates a variant of the model that accepts 512 × 512 pixel images as input.

Example filters by Krizhevsky et al. []. Because these filters were learned using the first convolutional layer of the neural network, the represented features are simple, such as the edge orientation and frequency (learned features become progressively more complex with each additional convolutional layer). Reproduced with permission from Alex Krizhevsky, Advances in NIPS 25; published by Curran Associates, Inc., 2012.