How does reverse image lookup work?
In: Technology
There’s a paper on the subject: [Fast Multiresolution Image Querying](https://grail.cs.washington.edu/projects/query/), by Charles E. Jacobs, Adam Finkelstein & David H. Salesin in the 1995 Siggraph proceedings. I implemented this from the paper to manage and sort my own image collection, and later worked on it at Google as my 20% project. I don’t think any of my own work actually went into Google’s reverse image search, so I can’t say if this is actually how they do it.
The gist of it is this: you scale the image down to a reasonable size, and then use a Haar Wavelet transform on it. This is similar to a Fourier transform, but easier to compute. The transform returns an array of values that basically represent the components of the image in the spatial frequency domain.
You pick out the 40 most significant values, truncate them to save space, and essentially construct a “signature” of the image. The rest of the algorithm is just a fast way to match the signature to the signatures of images in your database.
Latest Answers