Addi-11 / Neural_Style_Transfer

Producing artistic images

Neural Style Transfer

A tensorflow implementation of the project has been described in paper : A neural Algorithm of Artistic Style
The implementation process is as follows:

• The content and style image are run through the vgg model (extractor) to get the target content and style.
• An image is generated closer to the content image (or with noise added) - Generated Image
• Loss of the generated image is computed
• Tensorflow graph is intilaised and the generated image

Requirements

Data Files

Download the pretrained model from the following link : imagenet-vgg-verydeep-19.mat. Save it in a directory pre_trained_model The above method is used in the old-approach file. (which has some bugs at present).

The data can also be loaded and used directly from keras applications.

Dependencies

You can install the dependencies using pip install -r requirements.txt. Here's the list of required packages for manual installation

• Tensorflow
• Numpy
• pandas
• Scipy

Sample Results :

Content and Style Images

Generated Images

References :

The implementation has had from the following projects

• A tutorial version of the implementation https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/15_Style_Transfer.ipynb
• Implementation related to style transfer while preserving color and vedio style transfer https://github.com/cysmith/neural-style-tf
• Fun implementation https://github.com/hwalsuklee/tensorflow-style-transfer