Let’s look at drinks. For each drink, we’ll keep a running total of ‘points.’ If it’s hot, it gets one point. If it’s cold, it loses a point.
– Coffee is hot, +1
– Tea is hot, +1
– Milk is cold, -1
So we can see coffee and tea are more similar to one another than to milk, using this system, right?
We could make a similar system vastly more complicated, using hundreds of properties so that only a computer could check them all – color, shade, shape, texture, etc, anything you can name that makes an image unique and recognizable – to relate images to one another. You upload your image, the computer checks the image, gets the score, and then cross references that with the scores of other images – and it shows you those images under the assumption that they are similar.
Basically, the image is fed into a program that looks at the image and creates a mathematical model (sometimes called a fingerprint or hash) of the picture (how this is done is way, way beyond ELI5). The model is then compared to other models that exist from previous images that have gone through the same process. If the models are similar enough (say, 90% the same) then the reverse image searcher will display the compared images.
Each image has a level of the three primary colour e.g 94 red, 66 green, 34 blue which make up each pixel and show one colour. An example is a picture of a yellow dress so 50 red and 50 green with 0 blue. There are thousands of pixels in just a small section of every image (depending on the size).
Anyway every picture is then translated into mathematical code depending on the pixels and locations etc which a search engine compares to others.
Latest Answers