Join GitHub today
This project uses Keras and Python to convert a grayscale image to color without any additional information.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
|Failed to load latest commit information.|
Image Colorization == Notes == Note1 : Some architectures need very large memory (We used AWS EC2 P2 instance that has 61 GB memory to do the experiments) Note2 : Our results can be view from google drive: https://drive.google.com/open?id=0BzxDU3VAcOkQck9rYThkd3JmQVE Execution time : Run at least 30 minutes (1 hour for some larger architecture) for each code Results : View the result image in corresponding folder namde predict_output_(architecture) == Dependencies == OS: Ubuntu 16.04 Software: python Deep learning framework: Theano, Keras Dependencies: openCV == Dataset Download Links: == Since we rearranged the dataset, we uploaded the dataset to the following links 1.fruitdata : https://drive.google.com/open?id=0BzxDU3VAcOkQUVRhTlJ6SXlIYnc 2.combined : https://drive.google.com/open?id=0BzxDU3VAcOkQTnpmZjA2aGszWGM After download the dataset, fruitdata should be put into the fruit_colorizer folder, and combined should be put into landscape_colorizer folder == Run Code == 1.For training and doing predicting for Fruit dataset: $ cd fruit_colorizer $ colorization_auto_encoder.py # Use deep auto-encoder architecture $ colorization_auto_encoder_shallow.py # Use shallow auto-encoder architecture 2.For training and doing predicting for Open Country(Landscape) dataset: $ cd landscape_colorizer $ python colorization_use_pretrained_conv_layers.py # Use our pre-trained conv-layer and Dense layer. # This is our baseline model $ python colorization_concate_pretrained_conv_layers.py # Use concat conv layer architecture $ python colorization_auto_encoder.py # Use auto-encoder architecture $ python colorization_use_vgg.py # Use vgg as feature extractor $ python colorization_use_concate_vgg.py # Use concat vgg conv layers architecture 3. For comparison between color images and gray scale images testing on a VGG-16 fine-tuned network. The result shows loss and accuracy. $ cd comparison $ python vgg16_cifar_RGB_gray_comparison.py