Authors- Anhad Mohananey, Yashovardhan Chaturvedi
Convolutional Neural Networks to predict aesthetic goodness of an image.
To train the CNN:
import build_model
predictor = build_model.trainly(path_of_images_folder_for_training, no_of_epochs)
The predictor object stores the trained model, and can be used for predicting whether a particular image is aesthetic or not, like so:
import cv2
res_image=cv2.imread(path_of_test_image).reshape(-1,3,250,250)
print predictor(res_image)
The above will print a class label, (Aesthetic/Not Aesthetic)
Note: The 'factored' folder consists of 250X250X3 images scraped from the AVA/imagechallenge dataset(http://www.lucamarchesotti.com), which can be used for training purposes. The images that were originally downloaded were cropped to bring to 250X250X3. I would recommend downloading more images, particularly if you have a GPU, and are not constrained by training time.
Dependencies: Theano, Numpy, urllib, cv2(opencv in python).
scrap.py has the code for the scrapper that I used to get data, and consequently crop and process it to 250X250X3 px. You can use it as a reference or write your own. In case this in violation of AVA's guidelines, please let me know and I will remove it.