Food vs. Non-Food Image Classifier

Food:

Non-Food:

Notebook description:

The following Notebook was created in order to classify images into two categories: Food, and Non-Food. It relies on Keras, a high-level neural networks API.

Approach:

The following implementation compares the performance of a Convolutional Neural Network with the following parameters:

• Number of layers: 1, 2, 3
• Learning rate: 1e^-3
• Batch Size: 32, 100
• Epochs: 20
• Image Size: 32x32px, 64x64px
• Number of kernels in a layer: 10, 20 in the first layer; 50, 100 in the second layer
• Kernel size: 5x5 small image; 7x7 large image

Each layer follows CONV -> RELU -> MAXPOOL

Datasets:

The datasets is split three-way: Training: 60% Test: 20% Validation: 20%

Images used in the training/testing/validation are a mix of publicly available datasets.

Data Preprocessing:

The dataset has been preprocessed by resizing the images into 32x32 and 64x64px. The images remain as RGB, not gray-scale. The argument for keeping the representation as RGB is supported by Kiyoharu Aizawa, in the paper "Food Detection and Recognition Using Convolutional Neural Network," and can be found here.

Since the dataset used in training is sizable (~5.2k Food, ~4.5k Non-Food), there is no need for augumentation. However, on a limited dataset it would be beneficial to apply augumentation with Keras ImageDataGenerator, using the following techniques:

• rotation
• width/height shift
• sheer range
• zoom range
• horizontal flip

---

Results:

The best model with one layer has the following parameters:

Layer Size: 20 neurons Neuron Size: 5x5 Batch Size: 32

The best model with two layers has the following parameters:

Layer One Size: 20 neurons Layer Two Size: 100 neurons Neuron Size: 7x7 Batch Size: 100

Classifier usage:

$ python classifier -m model.h5 -i path_to_image.jpg