zyishai / image-retriever Goto Github PK

Improve readability and modularity of existing code, and add helper functions to enable debugging and visualizing of computations.

Reference: Article's section 6.1.1.

Description: For each offset, sum both Correlogram contribution as well as Histogram contribution and take the maximum.

Correlogram contribution is calculated by the difference between the autocorrelogram for query image that corresponds to the color of the offset pixel, and the local autocorrelogram of the sub image in the big image at location of the offset.

Transform image object to 2d numeric representation of the image, where each cell gets a label based on its KMean's cluster.

This has two advantages:

1. Working with numeric array is much faster than working with Numpy's arrays and OpenCV's colors.
2. Two different colors will get the same numeric representation, if they are in the same cluster (simplify equality check).

Problem: when calculating a correlogram, we should calculate specific correlogram for each two colors combination and for each distance. In an image, there are possibly 256³ different color combinations. Even for small images (e.g 160x160) that's too much (25600! options, not including different distances!!).

Suggested Solution: take only few major colors of the image, preferably colors that are close to each other (local correlations between colors are more significant than global correlations). This will also allow us to use relatively small distance.

Reference: Article's section 6.

Improvement on #1

To calculate correlogram of an image, we need to run for each two colors and for each distance the specific correlogram calculation. These calculations does not depend on each other, so we can parallelize them.

Task: move calculations to a queue and allow workers to calculate them asynchronously.