A python script capable of blending style attributes from one image with the structural content of another.
A style image and content image exist as complete images. A blank "output" image is also created. The thrust of the script is to minimize the content difference between the content image and the output image, whilst simulatenously minimizing the style difference between the style image and the output image. The ultimate goal is to project the different qualities of the two images onto the blank image, fusing the two within the new image.
Aggregating the sub-loss functions into a larger loss function allows the problem to be abstracted away as an optimization problem. A truncated, pre-trained VGG16 model created by Oxfords Visual Geometry Group was used as the convolution network. Whilst scipy's L-BFGS-B optimization algorithm was used to hone the total loss experienced by the combination image as it was developed.
Some of the results have been included, though the quality would benefit from a more well-endowed computer and/or some AWS credits..
 |
 |
 |
|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Ensure that all required libraries are available for use within the environment. In particular, ensure that Keras is using Tensorflow as the backend.
Place content and style images in the appropriate directories before running. The script will process and create all image combinations based upon the images supplied within these directories. Within the outputImage directory, additional directories corresponding to the content images will be created, inside these directories a predictable organizational structure will also be created.
Developed using Tensorflow 1.2.0, Keras 2.0.2, PIL 4.1.1, Numpy and Scipy All of which are freely available for use.
Python 3.6.0 - All scripting Spyder 3.1.2 - Scientific computing IDE Anaconda (Continum Analytics) - Popular data science platform
Sam Hardy - Implementation
Big thanks to Siraj Raval for managing to compress and spice dense topics like Neural Networks into reasonably short youtube clips, they're an excellent starting point. Another big thanks to Harish Narayanan for his blog entry which formed the basis for the entire script, an excellent overview of CNN as well as some very necessary pre-material (linear, non-linear classifiers, score functions)
| Title/Author | Link |
|---|---|
| "Convolutional neural networks for artistic style transfer" - Harish Narayanan | https://harishnarayanan.org/writing/artistic-style-transfer/ |
| "How Convolutional neural networks see the world" - Keras Blog | https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html |
| "MNIST for ML Beginners" - Tensorflow Blog | https://www.tensorflow.org/get_started/mnist/beginners |
| "How to Generate Art - Intro to Deep Learning #8" - Siraj Raval | https://www.youtube.com/watch?v=Oex0eWoU7AQ |