Ti-tanium/Color-Scheme-Style-Classifier

crawl images of different style, and extrac color scheme from the images, then try to train a classfier that distinguish color scheme of different style

Style_classifier

This repository trained a classifier to differentiate different style of color scheme, such as cute, fresh or technology.

Image Crawler

Prerequisite

The following libraries are required

pip install colorthief
pip install chardet
pip install http

Usage

The images used to train the classifier are crawled from ShutterStock. Three types of images are collected, including 'cute', 'fresh', 'technology'.

If you want to crawl more types of images, add your url in ditc 'url', and then add a tag to list 'style':

    styles=['technology','cute','fresh']
    url={
        'cute':'https://www.shutterstock.com/search/cute+pink?mreleased=false',
        'technology':'https://www.shutterstock.com/search/fresh+green?mreleased=false',
        'fresh':'https://www.shutterstock.com/search/technology+blue+future?mreleased=false'
    }

Change the PATH and DATA_FILENAME to where you want to save your data:

# where to save your data
PATH="./data/"
DATA_FILENAME="train_tech.csv"

And then run the following command, note that every page contains about 100 images, the Tag is the temporary file name to store the images.

python ./Image_Crawler.py startPage endPage Tag

For example:

python ./Image_Crawler.py 1 100 1

The images are stored in the 'data' folder. And also, the color scheme (palette) is extracted using Color Theif. Each palette is composed of 5 rgb colors, and stored in DATA_FILENAME, which looks like this:

Data

Color Extraction with Color Thief

Below is a example of color scheme extraction.

Image:

Extracted color scheme:

Classfier

The classifer is trained using lightGBM. The data set has 15 features(5*3 rgb). The Accuracy of the model is arround 0.9263. Any suggestion to improve the model would be appreciated😊. The parameters used to train lightGBM model are as follows:

# specify your configurations as a dict
params = {
   'boosting_type': 'gbdt',
   'objective': 'multiclass',
   'metric': {'multi_error'},
   'learning_rate': 0.01,
   'feature_fraction': 0.7,
   'bagging_fraction': 0.7,
   'bagging_freq': 4,
   'num_class':3,
   'num_leaves':400,
   'max_depth':10,
   'max_bin':30
}

The detailed code is in the 'classifyPalette.ipynb' notebook. If you are not interested in the detailed training process, you can also use 'stylePredict.py' to do color scheme pridiction.

Usage

Prerequisite

make sure you have the following module installed:

pip install lightgbm
pip install colorthief
pip install pandas

Download model.txt and stylePredict.py and you can use the style_predict function to predict color scheme style. The parameter is a list of 5 rgb tuples(eg：[(r,g,b),(r,g,b),(r,g,b),(r,g,b),(r,g,b)]). And it returns three styles ['cute','fresh','technology'].

Hear is two examples:

Image:

Palette: [(240, 224, 225),(144, 198, 216),(241, 40, 114),(221, 185, 54),(47, 177, 205)]

Predictions:

Image:

Palette: [(46, 175, 231), (76, 140, 144), (9, 102, 179), (6, 71, 130), (4, 23, 47)]

Predictions:

Because the color schemes extracted from images of different style may only differentiate slightly from each other, it's difficult for the classifier to distinguish these palettes. Given a better dataset, the results may improve a lot.