0:00:13so um
0:00:14my name is a somewhat what an bit from japan and
0:00:17so today i would direct to talk about uh about my
0:00:22a a a little work
0:00:24uh tight so that all
0:00:26fast identification of jpeg two thousand images for distance there's to provide
0:00:34uh
0:00:35this is a i'll try all of my talk
0:00:39part a i'm direct talk about C
0:00:42but grand all of this work
0:00:44a second
0:00:45yeah
0:00:48jpeg two thousand coding system
0:00:50and it's coding the power as uh some rice
0:00:53then
0:00:54then
0:00:55the calm an image identification miss out of what chip pick two so than
0:01:00images
0:01:01is introduced
0:01:03and in
0:01:04the program all was calm based a miss not uh describe
0:01:11the procedure zero
0:01:13a missile to comes next
0:01:15for the by
0:01:16experiment in
0:01:17results
0:01:20and if
0:01:22yeah i it's at time
0:01:23i we conclude my
0:01:25what
0:01:31um
0:01:32did does him
0:01:33uh is
0:01:35no no as
0:01:36one oh
0:01:37uh
0:01:39the most successful up of case one all say pick two thousand
0:01:44uh jpeg peg two thousand has a coding
0:01:47provides
0:01:48for just as cinema application as an i mean doesn't and all that to standard
0:01:55uh in this
0:01:56provide a uh coding parameters
0:02:00uh all typical two thousand made if a
0:02:03for
0:02:04it's
0:02:04provide
0:02:06um
0:02:07in this just name and applications
0:02:10uh
0:02:11such as a
0:02:12and being or
0:02:15we in
0:02:16and so on
0:02:17uh there are needs to
0:02:20i didn't to find and exact frame
0:02:23exact the same frame which has the same or is no image as a given or is not image
0:02:29uh
0:02:31this identification called image identification and uh
0:02:35this identification of jpeg two sounds and images are
0:02:39required a it's its S name because uh
0:02:42frames
0:02:43all
0:02:44this that's the contents
0:02:46uh encoded by a typical two sets and
0:02:50so um
0:02:52image i didn't
0:02:53you meet identification is
0:02:55defined
0:02:56uh an operation
0:02:58for finding an image identical
0:03:00to a a given the image from you may database
0:03:04and that this
0:03:05processing should be done in compressed domain
0:03:10uh
0:03:11how about uh commissioner miss thoughts
0:03:13cannot handle the chip to send images support just them
0:03:18um because
0:03:20in this this smell
0:03:21provides
0:03:22uh coding parameters such as a
0:03:24the average a label on
0:03:28and that quantization step size and the call broke side or or and so on
0:03:33um may differ
0:03:37um
0:03:38now i would like to talk about jpeg two thousand coding system
0:03:45this is a a diagram or jpeg two thousand coding system
0:03:49and uh
0:03:50decoding power meters
0:03:52using and jpeg two says uh used in jpeg two thousand
0:03:56uh according to each operation broke
0:04:00uh a was that rate character
0:04:03so uh
0:04:06i
0:04:06fast uh input image is uh
0:04:09and i Z by D that would be and think ties it and they're
0:04:12them
0:04:14uh each
0:04:15transform questions uh
0:04:18divided
0:04:19uh called rock
0:04:20and then
0:04:22bit about it meet the ring
0:04:24a
0:04:27uh it's down
0:04:28so after a
0:04:29bit modeling
0:04:32the number of bit of brains
0:04:34uh
0:04:36it called in this presentation
0:04:38a a and and that B P
0:04:40uh
0:04:41is produced and the recorded it as a part of the had that information of jpeg two sounds and called
0:04:47the story
0:04:50uh in this work
0:04:51uh
0:04:52this
0:04:53and that the bp
0:04:54is used to fall image identification
0:05:00uh
0:05:01i would like to talk uh
0:05:03that that detail four
0:05:05uh and that P P
0:05:08uh this figure shows uh
0:05:10transformed questions
0:05:14uh
0:05:16this is a wherever it questions and the
0:05:19to this it's a seven one
0:05:21and uh
0:05:23the rocks
0:05:23uh here
0:05:25so called rocks
0:05:27uh
0:05:30each quantise questions is
0:05:32separated into this
0:05:34each sign
0:05:36and the i was sort it should
0:05:39and the
0:05:40where a
0:05:42we have the absolute to use of
0:05:44for tries
0:05:47a speech prince
0:05:49from the most significant bit to say the least significant country
0:05:54all the samples in the bit prince uh
0:05:57is out
0:05:59zero or one
0:06:03then is a special breed brain
0:06:05uh in fish all samples are zero
0:06:10it's a P S
0:06:12and this special bit-plane is if to as a
0:06:16zero a brain
0:06:18and then when all
0:06:20it prints out
0:06:22zero be brains
0:06:23uh the cold broke
0:06:26this
0:06:28uh this is i
0:06:29coding you need to a chip pick two cells and
0:06:31uh
0:06:35defined as a not included
0:06:37broke
0:06:38because it does not include any data that to be encoded
0:06:52uh this figure shows the
0:06:54example all the
0:06:58and P P
0:07:04and not that uh and that B B
0:07:07can be easily extracted by passing to hit that all jpeg two thousand called streams
0:07:13and uh
0:07:13no know coding is it
0:07:16and uh compression late
0:07:18or was it called a stream
0:07:20has no impact to
0:07:22to bad in that P P
0:07:25uh there is a
0:07:27metal
0:07:27two
0:07:29vol you mean image identification
0:07:32using
0:07:32and that the P P
0:07:34uh i would like to summarise uh miss a
0:07:37uh this ought passed to extract the
0:07:40and that to be P using a
0:07:42hit the se
0:07:44and then
0:07:45uh
0:07:47the components and all in to be P between a
0:07:51query eight
0:07:52big two thousand images and the i image if
0:07:55mean
0:07:56and the J peg two that the images in uh
0:07:59and you made database
0:08:07uh this mess of works where for the a you is that to have the same coding power as
0:08:14however about it is not applicable with
0:08:17different coding parameters
0:08:19because uh
0:08:22barry all and that the P P in images
0:08:25compressed to is different coding parameters may be different
0:08:33so two so was uh
0:08:37this program
0:08:38uh the extended
0:08:40ms of it's probable
0:08:46uh the proposed missile
0:08:48a is composed of four steps
0:08:50at fast
0:08:51uh regions to be compared to a defined
0:08:56using the
0:08:57regions
0:08:58uh the difference of the labels and the or
0:09:03code
0:09:03called broke size can be can say
0:09:08uh here is a
0:09:10transform the sent
0:09:12uh of a query image
0:09:14and a this is the database one
0:09:17and uh
0:09:18each
0:09:19a a query images
0:09:21uh be composed to by W dwt
0:09:24and level four
0:09:25and don't data bases a three
0:09:28and we choose uh
0:09:31many
0:09:32little
0:09:33between the
0:09:34images
0:09:36uh
0:09:37for for identification
0:09:39so
0:09:41this is just have a one
0:09:43so
0:09:47in could we made this uh the rocks size is set to
0:09:51a to by set to
0:09:54but uh in
0:09:56database images uh
0:09:58that the
0:09:59uh for group size
0:10:01uh it's set to sixty four by sixty four uh
0:10:05they are
0:10:06define in each other the
0:10:09uh
0:10:13the whole so is on top and the but can sides off
0:10:17compared regions
0:10:19uh
0:10:19sit to user
0:10:21recent
0:10:22marks marks spot
0:10:24or
0:10:28this site
0:10:30so uh
0:10:32this is a example compared regions
0:10:37uh
0:10:38one regions
0:10:39uh
0:10:40has a a a a full whole blocks
0:10:43for a query images but uh
0:10:45one regions has a a one of the broke for database minute
0:10:52uh if
0:10:53but blue
0:10:55number of did zero bit to in the region like a query images
0:11:01uh or is the minimum value is used
0:11:05in some so you can't
0:11:06process
0:11:09like is
0:11:10uh a few days
0:11:12not included
0:11:13but you
0:11:14uh minimum value
0:11:17is you priest the by
0:11:18not good
0:11:24and then uh
0:11:26omission mission or
0:11:27relation vectors
0:11:28uh
0:11:29a right
0:11:30so the
0:11:31and that B P all joe for a region
0:11:36this is our region
0:11:39a region
0:11:40and that it's
0:11:41okay
0:11:42neighbouring regions are can be a and then
0:11:45so a back that is
0:11:47a produce
0:11:52this value all
0:11:54one
0:11:55means a in that B P
0:11:57or was a neighbouring regions is
0:12:00small runs and
0:12:02for or one
0:12:03and the minus one
0:12:05means a a and that to be also a neighbouring regions is run and
0:12:10folk or one
0:12:12and uh
0:12:14the components comparison
0:12:15with
0:12:16anybody value and or
0:12:18between uh
0:12:19not to include it
0:12:21oh to put
0:12:22zero eight minute
0:12:24all was a big to
0:12:27oh finally the local decision
0:12:30uh
0:12:31is made by these operation
0:12:37this local decision it's four
0:12:39one
0:12:40compared regions
0:12:42and the when uh
0:12:43for or
0:12:46uh
0:12:47this is a is
0:12:48or all or all or can the regions
0:12:51uh
0:12:53but steve
0:12:55this shows are made
0:12:56so
0:12:57the global but this is a is
0:12:59may
0:13:01if all
0:13:02local decisions uh point steve and then
0:13:07colour
0:13:08global but this isn't is
0:13:10plus still
0:13:15in the experiment uh we use the
0:13:19standard the evaluation just marked area uh it is defined as a uh
0:13:25D C i'm
0:13:26uh this never and you said tape
0:13:28uh was used
0:13:30uh
0:13:30this just it
0:13:32contained a uh over
0:13:36fifteen thousand frames
0:13:38and uh have fast
0:13:40a sequence of jpeg two thousand compressed
0:13:44you may is was to in a database
0:13:48and uh a query images once found a and uh and in five identified to an the jpeg two thousand
0:13:54images in that data
0:14:00uh the identification experiment how was performed to for all possible combinations of a query and a database images
0:14:10and the number of combinations of a query
0:14:13and the database
0:14:15well as a
0:14:16fifteen
0:14:17so was and by fifty and in total
0:14:25yeah it's a property of of the
0:14:27test C against and then the encoding power we can support query images
0:14:33uh the resolution
0:14:35all
0:14:36this C case
0:14:37is a
0:14:38four sounds and by
0:14:39two thousand
0:14:41and uh
0:14:43this is a
0:14:45so as as a coding a permit to as
0:14:48so and the
0:14:50this table
0:14:51a a is the power as support database images
0:14:55in this table
0:14:56uh
0:14:57this
0:14:59did to you piece if you've
0:15:01five three and C B
0:15:05C six people C B want to eight
0:15:07are we is
0:15:08i tape
0:15:10uh means uh images
0:15:12encoded using a different
0:15:15do you double D V you as and the core the block is
0:15:18a divide to iteration and different step sizes
0:15:26uh
0:15:27this is an experimental results
0:15:29and the
0:15:30the
0:15:31is force
0:15:32was you
0:15:33right
0:15:34if B R and the two to rate T B R
0:15:36is
0:15:37uh kinda kinda correct
0:15:41by
0:15:42these equations
0:15:47this table shows so
0:15:50the idea of identification based on the proposed a of produce to and uh
0:15:55one percent
0:15:56all
0:15:57the
0:15:59a a one percent
0:16:00if B are we of the difference of jpeg two thousand coding probably
0:16:06and the result any false negative match
0:16:15and uh
0:16:16this experiment
0:16:18a X Q to
0:16:20uh
0:16:22on a workstation with
0:16:24uh wrecked is
0:16:26and that the is processing time for one combination of a curry and uh
0:16:31database a image was about uh
0:16:34the point six
0:16:35me second pass frame
0:16:40a a a a finite i really
0:16:42i would to conclude the my talk
0:16:45uh
0:16:47and identification method of what jpeg two thousand images with uh
0:16:51different jpeg two thousand coding parameters has be
0:16:54present it
0:16:55uh
0:16:56for each as
0:16:58uh
0:17:00the proposed wasn't miss out
0:17:03is uh
0:17:05compressive domain processing
0:17:07and uh has a robust S against a difference
0:17:11encoding calling S
0:17:12and uh
0:17:13it's one
0:17:15runs
0:17:16last
0:17:17thank you for your attention
0:17:24any questions
0:17:31i i i one push
0:17:33i you should in the previous slide that but the uh i
0:17:38it T P R of the true positive
0:17:40and it was at a hundred percent
0:17:42does it mean that you could actually get results that the overall better but proving the F P R
0:17:48introducing reducing P are a little bit
0:17:51if you change the threshold
0:17:53a a actually a a
0:17:56we
0:17:57we have a
0:17:58we have a
0:18:00for home to some experiments uh a with a a different
0:18:05coding rate
0:18:07so we get the
0:18:09very low rate
0:18:10uh some
0:18:12results a
0:18:14uh oh obtained
0:18:16uh
0:18:18not hundred in
0:18:19"'kay"
0:18:21but uh
0:18:23oops
0:18:23also so
0:18:25much of different
0:18:30a a and you the questions
0:18:32Q let's the to speaker