Evaluating image-to-image search
Project by Misha Dowd, supervised by Peter Saveliev
It is the exact opposite of the text-to-image search we are familiar with. Given an image, visual image search engines find images in a given collection that are similar, in some way, to the query image. So far, these engines exist mostly as experimental prototypes with the exceptions of facial recognition software and finger print analysis. Most of these demo programs work with small collections of images and, frequently, without an upload feature, which makes testing impossible. Meanwhile, when testing is possible, the results are questionable.
The approach is based on the topological methods of image analysis used in Pixcavator. The difference is that the distribution of the sizes of all possible objects in the image is found and then compared to those of other images. The details about the program are here: Pixcavator image search.
The purpose of this project is to test and analysis the Pixcavator results and then determine what in the algorithm of the program can be altered.
Software (Windows): PxSearch
- creating datasets for various, medium-size image collections:
- there are public one on the web,
- collections of synthetic images can be created;
- developing a comprehensive review of the literature on the subject;
- evaluating the quality of the matching (see precision and recall in particular);
- modifying the matching criteria (bins for the distributions, thresholds for noise, color, etc).
Consider Visual image search engines.