author: Tilemachos Bontzorlos Python version: 2.7.11

This program is able to automatically find the tourism theme of a country using supervised machine learning methods, like Random Forests. The dataset used were images that can be gathered online (using the Instagram API for example).

The images used for the training and testing of the system are not included in this repository. Unfortunately, the generated training and test feature matrices from these images are also not included. However, each user can generate the feature matrices through the code in this repository using his own images. In future, I might add the feature matrices but for now the files are quite big in size (2.2GB).

The features used are RGB values and Histogram of Gradient (HOG) features. RGB features contain the information about colour and HOG features encode shape information. The preprocessing step is consisted of image resize to 50×50. The size of the RGB features per image equals to 50×50×3=7500. Then HOG features with 8 orientations resulting in 288 features per image are extracted. In total each image contributes to 7788 features. For training 6000 images (1000 images for each of the 6 themes) classified by theme through their labels were used. Themes selected are:

• Architecture
• Desert
• Beach
• Wildlife
• Nature
• Winter landscape

For each country up to 100 images that will be used for classifying the country were selected.

For the machine learning part the capabilities of scikit-learn library was utilized.

Example match of theme for a country:

World map of all tested countries colored in corresponding theme:

• Python 2.7
• numpy
• matplotlib
• glob
• sci-kit learn (sklearn)
• PIL
• skimage
• iso3166
• cartopy
• colorsys

train_data : Contains all the images used for training. It contains subfolders which are the labels of the images they contain. For example, all images in folder "beach" are labeled as beach images. The tourist themes selected and as a result the (empty) subfolders contained currently are: * Architecture * Desert * Beach * Wildlife * Nature * Winter landscape

test_data : Contains all the images used for classifying the countries. Each s ubfolders contains all the images for a country. For example (empty) folder "visit_austria" contains all images for Austria. The code assumes that these subfolders are named in form "visit_X", where X is the official name of the country. WARNING: if the name of the country is not the official then the world map won't be able to color this country (since its using a certain ISO). Example: Russia is officially called "RUSSIAN FEDERATION".

train_data_matrices : Contains all the data concerning the training.

• train_X.txt : the feature matrix of the training images.

• train_y.txt : the classification of the training images.

• class_themes.txt : the classes. Classification equal to 0 corresponds to first entry of this file and so on.

  All these files are created by "create_images_feature_matrices.py". Currently, this folder only contains dummy files for train_X.txt and train_y.txt, however you can generate new feature matrices for training by using your own images and the provided code in this repository.

test_data_matrices : Contains all the feature matrices for the countries (test data). Example: "test_X_france.txt" contains the features of the images for France. All these files are created by "create_images_feature_matrices.py". Currently, it only contains a dummy file, however you can generate new feature matrices for countries by using your own images and the provided code in this repository.

countries_graphs : In this folders the code outputs the theme matching graphs for each country.

results : This folder contains some results I have taken.

country_main_themes.txt : Contains the major tourist theme that a country has been classified to, for all the tested countries. Example entry: "Finland:winterlandscape"

world_map.png : The world map that "create_world_map_graph.py" outputs.

bar_chart_graph.py : Helper program that creates a graph showing the theme match of a country.

create_images_feature_matrices.py : This is a program that creates feature matrices for training and test data images contained in the corresponding folders. The features used are RGB and HOG features. Labels of the training data are the names of the folders that contain these images. Example: images in folder "beach" will be labelled as beach photos. For the test data there are no labels. The folder containing the photos defines the country that these photos belong to. Example: images in folder "visit_austria" belong to country Austria. The code in written in the way of assuming that the folders names of countries are "visit_X" where X is the country name. Also, it outputs class_themes.txt that contains all the classes. This program can run indepedently.

create_world_map_graph.py : This is a program that creates a world map where each tested country is colored to its best matching tourist theme and the rest countries are colored grey. This program can run indepedently.

classify_countries.py : This program is used to classify a list of countries to a tourist theme using supervised learning. It creates the feature matrices for the training and test data. It tests the accuracy of the classifier. Then it classifies the countries based on their feature matrices, creates a bar chart for each country matching and finally it outputs a world map depicted all the tested countries colored in the matching tourist theme. This program utilizes all the above mentioned scripts to achieve this functionality.