Getting Glitchy
Computers are these machines that we have engineered to be totally
deterministic. No matter how many times a computer is provided with with some input, it should
always give the same output. This is why glitch art is so fascinating - it happens when computers
dont work the way they are supposed to. Of course computers crap out all the time, but that is
usually due to poor programming for end-user applications. Actual data corruption is much rarer, and
can both wreak havoc and be utterly entrancing.
The thumbnail for this particular post is not a true glitch. Rather it is merely a simulation - an attempt at recreating the effects of data corruption. This instance in paricular uses a technique called pixel sorting. I am not the first person to have done this - not by far. However, I pride myself on writing original code to accomplish this task based only on example images that I have seen on the internet.
My first observation on pixel sorting was that while pixels are sorted along a single axis, they are not sorted across the full length of that axis. If they were, the image would be a fairly boring gradient. Instead, the sorting seems to respect an image’s features in some way. I decided to use edge detection to maintain the “shape” of the image and delimit the areas in which pixels may be sorted.
Edge Detection
After a bit of research, the Canny Edge Detector seemed to be the way to go. It produces as its output a very clean 1-bit image. That is to say, it does a good job of ignoring noise, and gives a definitive single-pixel wide edge.
The wikipedia article does a fairly good job of describing the algorithm. The main part I had to implement on my own was the hysteresis function, which determines whether or not an edge ought to be included in the final output.
1void chain(image& img, int r, int c, bool** visited){
2 if(visited[r][c]) return; //already been here, don't bother
3 visited[r][c] = true;
4 //we can only get here from a strong pixel, so we can make this one strong.
5 img[r][c] = pixel(1.0);
6 //look at the 3x3 grid around [r][c]
7 for(int i = r - 1; i <= r + 1; i++){
8 for(int j = c - 1; j <=c + 1; j++){
9 //if the pixel is out of bounds, ignore it.
10 if(i < 0 || j < 0 || i >= img.r() || j >= img.c()) continue;
11 //if the next pixel is strong, or is a candidate, add it to the chain.
12 if(!visited[i][j] && img[i][j].y >= 0.5) chain(img, i, j, visited);
13 }
14 }
15 return;
16}
17
18void hysteresis(image& img, std::vector<coord> slist){
19 bool** visited = new bool*[img.r()];
20 for(int i = 0; i < img.r(); i++){
21 visited[i] = new bool[img.c()];
22 for(int j = 0; j < img.c(); j++){
23 visited[i][j] = false;
24 }
25 }
26
27 for(int i = 0; i < slist.size(); i++){
28 //start a chain IFF the pixel is strong.
29 if(img[slist[i].row][slist[i].col].y == 1.0){
30 chain(img, slist[i].row, slist[i].col, visited);
31 }
32 }
33 for(int i = 0; i < img.r(); i++){
34 delete[] visited[i];
35 }
36 delete[] visited;
37 //remove any unconnected weak edges
38 threshold(img, 1.0);
39}
The hysteresis() function accepts an image of the preliminary edges in which an edge is either
strong, with a value of 1, or weak with a value of 0.5. All other pixels are set to 0. The parameter
slist is simply a list of all the pixels that were determined to be a strong edge. This prevents
scanning over the entire image again.
The chain() function is called by the hysteresis() function on all strong edges. It then looks
at the 8 pixels around the one given. If it finds a weak pixel, since it is bordering a strong pixel
it will be set to a strong pixel. chain() is then called recursively on the newly set pixel. In
this way, we move along an edge, creating a “chain” of strong pixels until the edge ends. Once the
hysteresis() function returns, the image only consists of strong edges.
Now that we have well-defined edges, we can sort pixels by brigtness. In the original image, we take the top row of pixels and starting from the left, scan until we hit either a strong edge or the end of the row. This section of pixels is then sorted. We then start from where we last stopped and repeat until the whole row is finished.
This simple process can create some rather stunning images:
Other Fun
Once I had the program working, I had the backbone for a rather nice general image processing program. Therefore, I decided to take a stab at some dithering, and the code amounted to an impressive 18 lines:
1//Floyd-Steinberg
2void dither(image& img){
3 pixel oldpixel;
4 pixel newpixel;
5 double q_error;
6 for(int i = 0; i < img.r(); i++){
7 for(int j = 0; j < img.c(); j++){
8 oldpixel = img[i][j];
9 newpixel = pixel(oldpixel.y > 0.5);
10 img[i][j] = newpixel;
11 q_error = oldpixel.y - newpixel.y;
12 if (j < img.c() - 1) img[i ][j+1].y += q_error * 7.0 / 16.0;
13 if (j > 0 && i < img.r() - 1) img[i+1][j-1].y += q_error * 3.0 / 16.0;
14 if (i < img.r() - 1) img[i+1][j ].y += q_error * 5.0 / 16.0;
15 if (i<img.r()-1 && j<img.c()-1) img[i+1][j+1].y += q_error * 1.0 / 16.0;
16 }
17 }
18 img.set_format("P1");
19}
I personally find dithered images to be rather pleasing to look at. (Scaling ruins the effect. Click
in the image for a direct link and make sure it is at 100% zoom exactly)
Download & Play
the code is available on github. All that is required for
compilation is g++. By default, when you run make, it will produce an executable called glitch.
The program operates Netbpm images, which are saved as ASCII text files. If you wish to convert an
existing image to a ppm image, I recommend using imagemagick like so:
convert image.jpg -resize 50% -compress none image.ppm
Omitting the -compress none flag will produce a compressed image that glitch does not know how
to handle. Most images are pretty big these days, and resizing to 50% just speeds everything up.
glitch can either read from standard input or open a file specified by the -i flag:
some-command | glitch -s > sorted.ppm
glitch < some-file.ppm -s > sorted.ppm
glitch -i some-file -s > sorted.ppm
There are a few more modes and effects beyond sorting and dithering. These can be found using -h
for help.