I am working with digital images that people have manually categorized on a number of parameters.
I want to be able to find images that are similar or dissimilar based on certain categories. For instance, we might want to pick two random images from our database that have similar "significance", but are otherwise as different as possible.
As an example, this picture of Stonehenge would be categorized like this:
colors: blue, green, grey themes: natural, land, sky brightness: light location: outdoor shapes: angular texture: hard lines: vertical significance: historical
The categories are mostly fixed and small, so when manually categorized the person is choosing the main colors from a list of ten colors (red, orange, yellow, green, blue, purplse, white, black, grey, brown), and the same is true of the other categories. Most of the categories only have a small number of options, such as location is indoor, outdoor, or unknown, and brightness is light, dark, or neutral.
My naive approach is something like this:
- Query the database for pictures with historical significance
- Pick a random starting image from the historically significant images
- Of the remaining historical images, do a random sampling of N images. Different values of N will give us more variety at the expense of accuracy.
- For each randomly selected image, create an ordered binary vector for all possible category values (ex. red=0, blue=1, outdoor=1, etc), producing some binary value.
- Compare the binary values for the starting image and the random image, for example by using a bitwise XOR operation (or equivalent set operation), then count the number of 1's in the result.
I am not necessarily interested in finding the most dissimilar images. Any two images that are mostly dissimilar will be fine, and in fact we typically do want some randomness so the same images pairs are not frequently selected.
This naive approach works okay. Are there any proper algorithms or techniques I should look at that might apply to this problem?