ImageFiltering

September 1, 2020 · View on GitHub

Introduction

This repository contains code to train, use and evaluate binary classification models to tag images as naturally occurring/photographic or not. With this code you can train both coarse and fine-grained classification models as described in our paper.

This research is derived from and extends Houda Alberts' master thesis.

License

Apache License 2.0.

Environment

Create an environment in Python 3.6 and install the requirements as below.

python3 -m venv venv_name
pip install -r requirements.txt

Data

There are two versions of the ImagiFilter dataset: v1.0 and v1.1. Version v1.0 is the original version and the one you should use in order to reproduce the results from our paper. In the second version v1.1 we have fixed a few images that have ambiguous/incorrect labels in v1.0, which is around 6 images. V1.0 contains images and their coarse- and fine-grained annotations, as well as train and validation splits, whereas v1.1 contains the updated file names according to the new assigned coarse and fine-grained classes, which can be used later on by changing the dataset accordingly.

Download v1.0 here.
Download for version v1.1 coming soon.

Usage

We assume you want to reproduce the results in our paper.

Data Extraction

First, download ImagiFilter v1.0 if you have not yet done so and untar it. You will create a data directory with subdirectories positive_images and negative_images containing images labelled positive and negative, respectively. You will also find files train.json and test.json.

tar zxvf imagi-filtering-data.tgz

Training (coarse prediction)

To train a LeNet CNN architecture from scratch, it is necessary to set the resize argument to True and run:

python train.py --model lenet --resize True

To fine-tune a VGG19 CNN architecture pretrained on ImageNet, run:

python train.py --model vgg

To fine-tune a ResNet-152 CNN architecture pretrained on ImageNet, run:

python train.py --model resnet

To ease the experiments, automation scripts in bash are also provided.

Training (fine-grained prediction)

Training for fine-grained image classification is very similar to above. To train a LeNet CNN architecture from scratch, it is necessary to set the resize argument to True and run:

python train_finegrained.py --model lenet --resize True