Procedure.txt

This file will answer these questions:
1. What is your area of interest? What are you specifically going to investigate?
2. What dataset are you going to use to do your investigation?
    A. Provide a link
    B. Describe the format - is it a file, a website, a database, etc.
    C. How big is the dataset? How many rows and how many columns?
3. How are you going to collect your dataset? Is it a simple download or will you write a web crawler?
4. What is your procedure for conducting your investigation?
    A. What preprocessing will you do on your data?
    B. What machine learning techniques will you use (feel free to consult an instructor on any bullet, but especially here)?
    C. Is there any other software development necessary to conduct your investigation?
5. How will you evaluate your system once it has been built?

1. My area of interest is artwork. I will investigate how to make a computer learn to create a piece of art.
The type of art will be water color. I will especially focus on what patterns appear most in watercolor art and the most prevalent
colors and items in those pieces of art. Additionally I will see what patterns appear most often and the distribution of paint to
white space on a canvas. Then it will also focus on the perspective and dimensions of the artwork weather the majority is 2d or 3d
by the most prevalent expression in the finished project. I also want to see if a computer can identify and replicate commonplace
patterns used in watercolor art to the best it can.
2. The dataset I am using is the BAM dataset at this link: https://bam-dataset.org/ . It is a sqlite database
and the one I am using has 2.2 million rows and 5 columns in the place the pictures are stored. I will also use the one that has
the images sorted by medium to get only the watercolor ones.
3. The database is a download, but it comes with 7 different types of images. I am going to sort it and 
make my own set of images of the type I want, water color images. I will have it find and download all the images
under that label and put it into a folder to use.
4. I will search and extract the mid numbers that have the media_watercolor set as positive and then use those mid numbers to extract
the links attached to them in the table with the link sources. I will then download all the images from those links. From those 
downloaded images I will make sure through a brief overview that none are unusable or don't match the correct style. This led me to
having a little over 100k images to use. Because an input is static on a neural network I have to  resize all the images to around the
same size which will cause compression and quality issues, but I don't see a way to get around it and will possibly lead to more 
interesting results. The generative network will take a 100 input and reshape it into a 65536*0.5*0.5 cube and go through 10 transposed convolutional
layers until it is a width of 3 and a size of 512 x 512. I chose this size since it is scaled to the size of
my 256x256 images instead of the papers 32x32. The output is a tanh function. The discriminator will do 
the reverse. It will take the image produced and run it through 10 convolutional layers until it is 65536*0.5*0.5
and then I will normalize it between 0 and 1, 1 being real and 0 being fake.
5. The stretch goal for this project is to have it graded by an adversarial neural network, however the first time through with the
neural network I will have a currently unknown, but hopefully 5-10 people grade it on the rubric outlined below on about 500 images
total that it will produce, I will then submit these back into the neural network and run it again. If this turns out to be too slow
or ineffective for my project, I will switch to using a DCGAN in order to speed up the process and make it more effective or efficient.
The large amount of time it will take to grade the images for my project using humans and the over subjectivity with it led me to using
the DCGANN method as it is much faster and objective.
However due to the slow nature and the small test size with using a rubric will lead to subjectivity and inconsistency. Additionally
this won't help as the network needs multiple tests and too fully test something I would have to spend the days or weeks.

1. How well does the art follow color theory. Do the colors flow nicely, is contrast used correctly, do the colors work with eachother
to provide a good looking image.
2. Does the painting replicate nature or other subjects from life.
3. Does it balance between empty space and filled space.
4. is the shading natural and flows nicely and correctly.


DCGANN guide: https://arxiv.org/pdf/1511.06434.pdf