I want to find image duplicates and similar images in MS SQL Server 7.
I got it running using sql cursors - it's slow but it works, thanks again for the comment. See my answer for the solution I came up with.
I have a database containing image paths and the fingerprints of the associated images computed with the help of this dhash algorithm. I'm using a variation where I'm storing 128 bits per image (horizontal and vertical gradient) in a BINARY(16)
column.
What I want to be able to do is something in the lines of:
DECLARE @id INT
SET @id = ...
DECLARE @given_hash BINARY(16)
SET @given_hash = ...
SELECT TOP 10 file_path,hash,
(hamming_distance(hash, @given_hash)) AS distance
FROM my_table
WHERE distance <= 20
ORDER BY distance ASC
It is trivial to obtain exact duplicates - just use a WHERE hash = @hash_to_compare
and that's it.
However, I want to be able to use a similarity measure (Hamming Distance) to account for small manipulations/defects/compression artifacts etc. I've come up with a stored procedure implementing the distance measure:
CREATE PROCEDURE hamming_distance128
@hash BINARY(16),
@supplied BINARY(16)
AS
DECLARE @i INT, @j INT
SET @i = 1
SET @j = 1
DECLARE @count INT
SET @count = 0
DECLARE @byte TINYINT
DECLARE @length TINYINT
SET @length = 16
WHILE @i <= @length
BEGIN
SET @j = 1
SET @byte = CAST((ASCII(SUBSTRING(@hash,@i,1)) ^ ASCII(SUBSTRING(@supplied,@i,1))) AS TINYINT)
WHILE @j < 256
BEGIN
SET @count = @count + (CASE (@byte & @j) WHEN 0 THEN 0 ELSE 1 END)
SET @j = @j * 2
END
SET @i = @i + 1
END
SELECT @count
GO
Unfortunately the DBMS (SQL Server 7 - it's impossible to upgrade/change this) won't let me use it to compute distances within a query and this piece of j*nk doesn't support user defined functions. Sure enough I didn't find anything like MySQL's BIT_COUNT
that would make this a no brainer for T-SQL.
Is there hope of getting this to work in T-SQL on SQL Server 7?
Help very much appreciated!
I got it running using SQL cursors and a temporary table. Thanks again @Tomalak
I provide my implementation for future reference, hope this is useful to anyone
CREATE PROCEDURE hamming_distance128
@hash BINARY(16),
@supplied BINARY(16),
@distance TINYINT OUTPUT
AS
DECLARE @i INT, @j INT
SET @i = 1
SET @j = 1
DECLARE @count TINYINT
SET @count = 0
DECLARE @byte TINYINT
DECLARE @length TINYINT
SET @length = 16
WHILE @i <= @length
BEGIN
SET @j = 1
SET @byte = CAST((ASCII(SUBSTRING(@hash,@i,1)) ^ ASCII(SUBSTRING(@supplied,@i,1))) AS TINYINT)
WHILE @j < 256
BEGIN
SET @count = @count + (CASE (@byte & @j) WHEN 0 THEN 0 ELSE 1 END)
SET @j = @j * 2
END
SET @i = @i + 1
END
SET @distance = @count
GO
DECLARE @min_similarity FLOAT
SET @min_similarity = 0.85 -- | 85%
DECLARE @supplied_hash BINARY(16)
SET @supplied_hash = 0x392929295B4B13371B0301272D2B2509
IF OBJECT_ID('tempdb..#distances') IS NOT NULL
BEGIN DROP TABLE #distances END
CREATE TABLE #distances
(id INT NOT NULL,
similarity FLOAT NOT NULL)
DECLARE
@tmp_id INT,
@dhash BINARY(16),
@distance TINYINT,
@similarity FLOAT
DECLARE rowCursor CURSOR
LOCAL FORWARD_ONLY READ_ONLY
FOR
SELECT id,dhash_value
FROM hashes
OPEN rowCursor
FETCH NEXT FROM rowCursor
INTO @tmp_id,@dhash
WHILE @@FETCH_STATUS = 0
BEGIN
EXECUTE hamming_distance128 @dhash, @supplied_hash, @distance OUTPUT
IF @distance < 128
BEGIN
SET @similarity = CAST(128 - @distance AS FLOAT) / 128.0
IF @similarity >= @min_similarity
BEGIN
INSERT INTO #distances (id,similarity)
VALUES (@tmp_id,@similarity)
END
END
FETCH NEXT FROM rowCursor
INTO @tmp_id,@dhash
END
CLOSE rowCursor
DEALLOCATE rowCursor
SELECT hashes.id,#distances.similarity
FROM #distances
INNER JOIN hashes
ON hashes.id = #distances.id
ORDER BY #distances.similarity DESC
DROP TABLE #distances