2017.10.17

In course CS348, Computer Vision, our coursework is to write a paper and submit it to a conference.

My Partners are Siqi Liu and Mengyao Cao. After discussion, we decided our draft topic as:

A Painting AI based on Cascaded Refinment Network

The PPT illustrates our scheme:

Download File

The final topic will be confirm in tomorrow’s class.

Wish our project a brilliant success!

2017.10.25

This time, I read a paper named A Neural Algorithm of Artistic Style.

This paper is about generating a picture with given real picture and a Painting. The generated picture will have both the feature of the painting and the initial picture.

For example, this is the painting used:

Then, the Algorithm will produce a picture like this:

dome-afremov

It is so amazing!!

The paper

The method of the paper is:

Use the VGG-19 network to process the initial picture, noted by $\mathbf{p}$, and the painting, noted by $\mathbf{a}$, then at each layer, the nerwork will have some feature maps corresponding to $\mathbf{p}$ and $\mathbf{a}$.

Input a noise picture $\mathbf{x}$ to the network, also, $\mathbf{x}$ will also have some featuremaps at every layers.

Define a loss function between $\mathbf{x}$ and $\mathbf{p}$, called content loss. And a loss function between $\mathbf{x}$ and $\mathbf{a}$, called style loss.Then, the author define a compound loss function:

$$L_{total}(\mathbf{p}, \mathbf{a}, \mathbf{x})=\alpha L_{content}(\mathbf{x}, \mathbf{p}) + \beta L_{style}(\mathbf{x}, \mathbf{a})$$

Using the optimization method to maximize the $L_{total}$, and with this process, fix the noise picture $\mathbf{x}$. Then, the noise picture will become the finally result.

We can actually change the ratio between $\alpha$ and $\beta$ to change the ratio of content and style.