ImgSmlr – similar images search for PostgreSQL

Introduction

ImgSmlr – is a PostgreSQL extension which implements similar images searching functionality.

ImgSmlr method is based on Haar wavelet transform. The goal of ImgSmlr is not to provide most advanced state of art similar images searching methods. ImgSmlr was written as sample extension which illustrate how PostgreSQL extendability could cover such untypical tasks for RDBMS as similar images search.

Authors

• Alexander Korotkov [email protected], Postgres Professional, Moscow, Russia

Availability

ImgSmlr is released as an extension and not available in default PostgreSQL installation. It is available from github under the same license as PostgreSQL and supports PostgreSQL 9.1+.

Installation

Before build and install ImgSmlr you should ensure following:

• PostgreSQL version is 9.1 or higher.
• You have development package of PostgreSQL installed or you built PostgreSQL from source.
• You have gd2 library installed on your system.
• Your PATH variable is configured so that pg_config command available.

Typical installation procedure may look like this:

$ git clone https://github.com/postgrespro/imgsmlr.git
$ cd imgsmlr
$ make USE_PGXS=1
$ sudo make USE_PGXS=1 install
$ make USE_PGXS=1 installcheck
$ psql DB -c "CREATE EXTENSION imgsmlr;"

Usage

ImgSmlr offers two datatypes: pattern and signature.

Datatype	Storage length	Description
pattern	16388 bytes	Result of Haar wavelet transform on the image
signature	64 bytes	Short representation of pattern for fast search using GiST indexes

There is set of functions *2pattern(bytea) which converts bynary data in given format into pattern. Convertion into pattern consists of following steps.

• Decompress image.
• Make image black&white.
• Resize image to 64x64 pixels.
• Apply Haar wavelet transform to the image.

Pattern could be converted into signature and shuffled for less sensitivity to image shift.

Function	Return type	Description
jpeg2pattern(bytea)	pattern	Convert jpeg image into pattern
png2pattern(bytea)	pattern	Convert png image into pattern
gif2pattern(bytea)	pattern	Convert gif image into pattern
pattern2signature(pattern)	signature	Create signature from pattern
shuffle_pattern(pattern)	pattern	Shuffle pattern for less sensitivity to image shift

Both pattern and signature datatypes supports <-> operator for eucledian distance. Signature also supports GiST indexing with KNN on <-> operator.

Operator	Left type	Right type	Return type	Description
<->	pattern	pattern	float8	Eucledian distance between two patterns
<->	signature	signature	float8	Eucledian distance between two signatures

The idea is to find top N similar images by signature using GiST index. Then find top n (n < N) similar images by pattern from top N similar images by signature.

Example

Let us assume we have an image table with columns id and data where data column contains binary jpeg data. We can create pat table with patterns and signatures of given images using following query.

CREATE TABLE pat AS (
	SELECT
		id,
		shuffle_pattern(pattern) AS pattern, 
		pattern2signature(pattern) AS signature 
	FROM (
		SELECT 
			id, 
			jpeg2pattern(data) AS pattern 
		FROM 
			image
	) x 
);

Then let's create primary key for pat table and GiST index for signatures.

ALTER TABLE pat ADD PRIMARY KEY (id);
CREATE INDEX pat_signature_idx ON pat USING gist (signature);

Prelimimary work is done. Now we can search for top 10 similar images to given image with specified id using following query.

SELECT
	id,
	smlr
FROM
(
	SELECT
		id,
		pattern <-> (SELECT pattern FROM pat WHERE id = :id) AS smlr
	FROM pat
	WHERE id <> :id
	ORDER BY
		signature <-> (SELECT signature FROM pat WHERE id = :id)
	LIMIT 100
) x
ORDER BY x.smlr ASC 
LIMIT 10

Inner query selects top 100 images by signature using GiST index. Outer query search for top 10 images by pattern from images found by inner query. You can adjust both of number to achieve better search results on your images collection.