Speech Transcript - LOW-COMPLEXITY PREDICTIVE LOSSY COMPRESSION OF HYPERSPECTRAL AND ULTRASPECTRAL IMAGES

0:00:13	among everyone this is a of lossy compression on a lossy compression on a set for hyperspectral and all spectral
0:00:20	images
0:00:21	that's a joint work with and they are about the not a button you the university of C N
0:00:25	so first give some motivation apart one but was a compression and the a specific problems
0:00:31	i will describe the proposed compression algorithm then a provide some experimental results on
0:00:37	i spectral and out for spectral images and
0:00:39	finally draw some calm
0:00:42	um
0:00:42	i spectral images are a collection of pictures taken at seven different wave of the same scene
0:00:48	quite a lot of data but when it comes to compress those data on were set to like we're faced
0:00:53	with the problem that
0:00:54	we don't have many computational resources to actually do
0:00:57	impressions so
0:00:58	first that the need is some low complexity um
0:01:01	that's quite different from compression on the ground our car idle is
0:01:05	the second but it thing is that unlike typical consumer digital cameras where there is a two dimensional
0:01:12	detector are rate to
0:01:13	picture
0:01:14	for hyperspectral images to we just have one um
0:01:17	a signal one dimensional array
0:01:19	that takes one a line of the image with all it's
0:01:22	a spectral channels all the different wavelength
0:01:25	the i-th the lines of the images or the image are for by the motion of the aircraft or set
0:01:29	light
0:01:30	so it they are taken at time
0:01:32	so that what use in table is that's we to the compression we don't have the whole image available but
0:01:38	just a few lines of the image with all the spectral channel so
0:01:41	we need to do compression using a a we will about one
0:01:46	so that's do we want to compress
0:01:48	good as we can so we want to have state-of-the-art compression efficiency
0:01:52	and for hyperspectral applications we need to cover a range of bit which is relatively large
0:01:57	to be from around zero point five two
0:01:59	three three or four bit
0:02:00	a pixel
0:02:01	that should be compared with a a sensor be that is typically between twelve and
0:02:05	uh sixteen bits per se
0:02:07	so we have to cover
0:02:08	a a little bit rate and high bit rate
0:02:11	fashion and finally
0:02:12	we need some error containment uh those uh
0:02:15	compressed data packets are down only uh use a communication channel
0:02:19	which is subject to occasional back loss and we don't want
0:02:22	that's a the signal back to these disrupts the reconstruction of the whole
0:02:27	so
0:02:28	uh there are several of all buttons for performing compression of two dimensional images like
0:02:34	and hyperspectral or on for spectral images
0:02:37	a most popular approach uses three dimensional tile from coding
0:02:41	for each sample jpeg two thousand we part two with that of course a multi component transformation
0:02:46	where you can use
0:02:47	you the spectral wavelet transform or an arbitrary spectral transformation
0:02:51	and then the light jpeg two thousand on each of the transform
0:02:54	a a spectral chan
0:02:56	we we can use a way list we use the couple nonlinear transforms whatever
0:02:59	this
0:03:00	uh uh was very well at low rates of the problem of this approach
0:03:03	the high complex
0:03:05	complexity comes from the top for but you know or or from the coding and rate distortion
0:03:09	use a shouldn't G
0:03:10	a thousand where you had to
0:03:11	uh
0:03:12	do we
0:03:13	i i coding all you links and then post compression rate distortion
0:03:16	ascension so that
0:03:17	mass too complex for on more
0:03:20	that it does work well for
0:03:21	archive
0:03:22	hmmm
0:03:24	the second thing is that if we want to use J two thousand well then a for compression we don't
0:03:29	have a whole image available
0:03:31	so our spatial transformations
0:03:33	we have just to take that
0:03:34	on a few lines of that image at a time
0:03:36	which is that
0:03:37	a possible with J
0:03:38	a thousand using be line based transformation
0:03:41	but then that start from them a shouldn't can be done in a global way more are
0:03:45	you has to be done and a lot of weight just a few lines at that time and there's a
0:03:49	big performance pen
0:03:51	or that
0:03:52	oh to be
0:03:52	a a global optimal rate distortion
0:03:55	age
0:03:56	the you the also approach is to use of prediction techniques using three dimensional spatial and spectral prediction
0:04:02	but action has been used for a time
0:04:05	for the last this compression
0:04:07	yeah was this compression used to to go for high quality applications where you want the maximum absolute error between
0:04:13	the decoder in an original image to be a by
0:04:16	a a user select value
0:04:18	so that works very well i high rates
0:04:21	i it doesn't work well as well the little bit rate
0:04:23	and then uh
0:04:25	a two dimensional prediction is usually a couple of colour this station an on how one i entropy coding
0:04:31	have a a clear that we don't go below one
0:04:33	pixel
0:04:34	the short course
0:04:35	a code word
0:04:36	but how one on can provide is just one
0:04:38	so
0:04:38	uh that's a problem
0:04:40	to go below one
0:04:42	so uh what we propose for a compression is based on a a approach where use three dimensional
0:04:49	spatial and spectral predictor or which keeps us a low complexity that need
0:04:53	a compression
0:04:55	but then we're faced with the problem of improving performance of a low bit
0:04:59	existing schemes
0:05:00	the just don't them
0:05:02	so that we don't know how do them is she's we don't really need to perform you was let's compression
0:05:07	move to
0:05:08	truly lasting compression
0:05:09	all the prediction residuals
0:05:11	and so in order to do that we improve quantisation stage
0:05:16	we don't use a simple scalar want either
0:05:18	and we a rate distortion optimization the whole ski
0:05:23	so
0:05:24	uh
0:05:25	this is how we do it
0:05:26	uh the prediction stage you use
0:05:29	therefore forms the prediction independently on sixteen by sixteen blocks of samples
0:05:34	so the speech or shows uh and in which dividing in four sixteen by sixteen blocks
0:05:39	and uh uh of for every we look at a all channels and the court look blocks in the the
0:05:45	different spectral channels so looking
0:05:47	what where the wavelet to mention
0:05:49	we probably right
0:05:50	from the curve at decoding block the bleeding spectral chan
0:05:54	so this is quite different from the kind of prediction that is usually uh and blowing in hyperspectral image compression
0:06:00	pixel by pixel
0:06:01	and not a lot by block prediction
0:06:03	but as will see is a loss as to a very efficient to start from
0:06:07	i
0:06:08	so essentially what we do is uh that are would need to in the next slide that we calculate a
0:06:13	linear predictor that use the previous well
0:06:16	but it be current law
0:06:17	then we calculate
0:06:19	the position the prediction residual and i think about this is that it provides a spatial error containment in that
0:06:25	it some compressed where one is lot
0:06:28	should
0:06:28	that we have set the next blocks in the
0:06:31	uh the the the the block in the same spatial position in the next wave lines but it not affect
0:06:36	and now their spatial loss
0:06:38	which
0:06:39	so the prediction is actually quite seen but we got X
0:06:42	the vector of samples of the current sixteen by six
0:06:45	in lexical that ring
0:06:47	are the vector of samples of the reference plot which is as i said
0:06:51	but score okay at decoder block in previous spectral channel
0:06:55	and then we calculate at least mean square predictor or with which is defined by two parameters are are fine
0:07:00	and and
0:07:01	and is the mean value of the current block and i'll wise at least means square parameter that
0:07:05	pretty P
0:07:06	current uh one
0:07:08	probably principal
0:07:11	alright
0:07:12	so uh the first uh the other thing to move formula
0:07:17	lossy compression quantisation
0:07:19	uh uh
0:07:20	technical near lossless compression of less quantisation
0:07:24	which is almost a do not at high bit but far from optimal a low bit
0:07:29	so for a bit rate compression its customary to use one with that some which creates long sequences of zeros
0:07:36	that are back to the effectively by and three coders
0:07:39	is is the optimal at lower rates
0:07:41	not i rates
0:07:42	and to find some that works well at all rates we decide to you
0:07:46	a a kind of quantizer which is called uniform threshold quantizer or or you Q
0:07:50	which is slightly more complex than E
0:07:53	in form quantizer
0:07:54	but dead zone
0:07:55	but is the or not at all
0:07:58	yeah O you do you
0:08:00	is actually
0:08:01	very simple
0:08:02	it's
0:08:03	i
0:08:04	i i there in which every interval what decision you of rows are all the same size
0:08:09	so calculating the court were is done much the same way
0:08:12	classical
0:08:13	colour you from quantizer
0:08:15	the difference lies in the fact that that construction there is not taken as the midpoint point of the quantization
0:08:20	interval but rather the sense
0:08:22	as H
0:08:24	so uh since we are a blank these to the prediction was walls we make use some so that the
0:08:29	position was you those
0:08:30	for a two dimensional
0:08:32	a a two sided exponentially decreasing distribution lower plot distribution
0:08:37	and so we calculate the
0:08:39	uh
0:08:39	actually actual construct and that used as a it's using these
0:08:42	is to be
0:08:44	and what happens is that if should look at this speech or
0:08:46	so we you can see here are are are the different quantization interval
0:08:50	the getting but of the point of the interval that you know a construction put by a a or for
0:08:56	one times are
0:08:58	a seems we use this uh we make this some some of distribution their out that the way that is
0:09:03	of the prediction that or or more be then the high values
0:09:06	so uh what we do is we add the collection term to the the construction but
0:09:11	account for that
0:09:12	so that by is that construction to works a little bit
0:09:15	so much so as the uh quantisation indexes low so close to
0:09:20	so that
0:09:21	uh
0:09:22	the or from your last
0:09:25	and the second and was the most important one is rate distortion optimization
0:09:28	and this is where really helps to use
0:09:31	a square blocks for the prediction
0:09:33	so the eight yeah here is
0:09:35	uh essentially similar to the skip mode be the compression
0:09:39	sometimes we find certain sixteen by sixteen or that can be pretty to really very well from the you reference
0:09:45	blah
0:09:46	and in that case we don't in the prediction a so we are other keep the encoding of the prediction
0:09:51	or
0:09:51	so that we save a lot it's in the process and just signal to the decoder
0:09:55	that the decoded signal on this case we just the pretty
0:09:58	that the decoder can
0:09:59	a couple
0:10:01	in particular
0:10:02	we actually
0:10:03	prediction
0:10:04	we calculate the variance of the prediction and D
0:10:07	and he can they're this variance for the special
0:10:11	and if D X is the threshold that
0:10:13	it means that the predictor is not a good enough for speech in the current block so we don't the
0:10:18	classical and encoding of the prediction or
0:10:20	i rice
0:10:21	if
0:10:22	the D is below a actual
0:10:24	we simply to that and prediction parameters for a lot of the file but no prediction that so
0:10:29	so that
0:10:30	the the would will be a a to uh mean the petition parameter
0:10:34	from the file
0:10:34	calculate a pretty or in use the predictor as coding
0:10:41	entropy coding all B
0:10:43	but addition was used was is done using a a power of two call
0:10:47	this is a a very typical become uh thing in uh compression for set like imaging
0:10:53	because goal of two codes are much see there than any other
0:10:56	a a cool especially arithmetic medical
0:10:59	so
0:10:59	uh there not was powerful but that that's of a good compromise
0:11:03	performance and complexity
0:11:05	and calculate
0:11:06	the best coding parameter
0:11:08	every sample
0:11:09	based on the of magnitude of on my prediction residual of a window of thirty two
0:11:14	so it's not done a lot
0:11:15	well
0:11:16	a sample by sent
0:11:19	right
0:11:20	so here are some results for the proposed algorithm
0:11:24	a a tried that on a different images all show results for
0:11:28	images from two different data
0:11:30	well as i've is i have this is an images
0:11:32	using spectrum there
0:11:34	long
0:11:34	which is flown on the aircraft
0:11:37	and these images are
0:11:39	had it they have two hundred fifty four spectral channels and the spatial signs
0:11:43	six and an at times five hundred and twelve images
0:11:47	um
0:11:47	and the is a the right images as are where by this
0:11:51	they have no calibration whatsoever no of corporations and not with really image
0:11:56	it is taken by
0:11:57	i've of is
0:11:58	used for
0:11:59	so that classification of
0:12:00	locations and
0:12:01	oh
0:12:03	the second image isn't a in image from the years and
0:12:07	some the which is operated by the not
0:12:10	which is used for a static of studies
0:12:12	these images have a a a a a much less piece spatially
0:12:16	just one hundred and thirty five times ninety pixels
0:12:19	but they have a a spectral channel signal can hundred one spec
0:12:25	is a quality metric we look at the P peak signal noise ratio and and we compare the performance of
0:12:30	the proposed algorithm with two other algorithms
0:12:33	well i he's jpeg two thousand part with the spectral
0:12:37	discrete wavelet transformation
0:12:39	in this case we do not perform the three dimensional a rate distortion optimization we're not doing any
0:12:45	line based comes from so that is also be shown for J
0:12:48	a thousand or and i'm realistic and
0:12:50	one would be actually run the set
0:12:53	a sort of upper bound of the
0:12:54	or of J
0:12:56	and the second algorithm is near lossless compression
0:12:59	use exactly the same predictor
0:13:01	and it to be colder
0:13:03	and not using the U G Q quantizer or nor the latest store
0:13:07	just a or
0:13:09	a by E D P C and we discard uniform quantization and entropy coding of the prediction was
0:13:16	are the results
0:13:17	the curve here
0:13:19	july two thousand to the wavelet transformation
0:13:22	and a continuous list compression algorithm
0:13:26	it is no you assess compression is better and transform coding at high bit-rates and you can see that here
0:13:32	will performance difference with respect to jpeg two thousand speech large over two bit per sample
0:13:38	but that were try this this is not a as good uh so essentially for two reasons
0:13:42	one is related to the fact that
0:13:44	the like to the quantization step size
0:13:47	and the were is the quality of the reference signal for the prediction so these points
0:13:51	this brings the
0:13:52	performance style
0:13:53	and low bit
0:13:54	but then and this i have is not able to achieve a rates be more one bit
0:13:58	a pixel because we're using a
0:14:01	got a call was mean got were like this one
0:14:04	there's just no way to go below that
0:14:06	the proposed algorithm seems to bring the best of both worlds here
0:14:10	better then a job but two thousand and and are a bit rate is larger than then point three or
0:14:15	do you point thirty five percent so
0:14:18	you and that's bit rates the rate distortion optimization works pretty nice here
0:14:22	and it's for for its performance tends to the performance of the yeah lost this compression at high rates and
0:14:28	that's
0:14:29	reasonable about because at high bit-rates
0:14:31	that it is a shall never select the skip model for any block the image and a uniform threshold quantizer
0:14:37	tense this colour one
0:14:39	so
0:14:39	the two algorithms essentially become pretty much the same
0:14:44	we have similar
0:14:45	you a results for the it's image
0:14:48	sort of a yes
0:14:49	do a two thousand as a little bit better or sometimes are performs by a small market
0:14:54	proposed algorithm some but is not quite as good
0:14:57	essentially
0:14:58	you know a comparable performance
0:15:01	and and that's pretty much the same you some near as compression is not a as low bit rates is
0:15:06	become pretty high bit
0:15:08	so this is still a a a a lot of jpeg two thousand for this image but jpeg two thousand
0:15:13	you and recall
0:15:14	is using a um the from a three dimensional very from
0:15:17	as a so
0:15:18	if we use the line based on from no one
0:15:22	so
0:15:23	um
0:15:25	alright right so uh the this is an example of visual quality and this is just essentially sensually goes to
0:15:30	show that we all the were using a block based pretty or we we don't have any hard
0:15:34	here
0:15:35	so this is a
0:15:36	uh a patch from one the end of every as your original signal
0:15:41	and this is not a construct signal by the proposed algorithm at zero one forty bit per piece so it's
0:15:47	is
0:15:47	you know one of the
0:15:48	oh well as
0:15:49	bit rates at the output but in can achieve
0:15:51	and as can be seen that i mean the artifacts
0:15:53	but no not not to science
0:15:55	what's that
0:15:56	the is is that a lot not the facts
0:15:58	uh come from the quantisation the transform from the a some from the coupling of one position
0:16:03	first transformation
0:16:04	where i is in this case where using a block based pretty but the quantisation used and independently of the
0:16:09	signal send
0:16:10	pretty
0:16:11	and what not
0:16:12	so this is what would have a job but for example
0:16:15	which creates you know a a lot
0:16:17	here just which essentially keeps the text or
0:16:22	alright right
0:16:23	so
0:16:24	uh
0:16:25	you can can an uh the proposed is essentially
0:16:29	uh a a a and you by for compression of
0:16:31	a a hyperspectral image where we achieve lower complexity by use
0:16:34	a prediction based approach
0:16:37	which uh uh forms
0:16:38	uh
0:16:39	is known as or better than the state-of-the-art of the art three dimensional for coding with really feature
0:16:44	distortion for optimization
0:16:45	so that seems to be a nice way for for on what compression of set images
0:16:50	complex in memory requirements are significantly lower than jpeg thousand
0:16:55	a it's difficult to compare the complexity of different algorithms by top to sign this working on J two thousand
0:17:02	and seems like the proposed approach to be like and man two
0:17:05	fewer operations than J
0:17:07	for a to the same
0:17:09	on this
0:17:11	uh but used in room for improvement
0:17:13	we're not using any or i've made calling but that and certainly have the coding efficiency apartment coding
0:17:20	what most are as by some margin
0:17:23	we might use
0:17:24	know and of the ring
0:17:25	that is using for a reference uh spectrum channel for the prediction not just a three spectral channel
0:17:32	a the spectral channel but use more correlated with the colour channel
0:17:36	so this is especially on is that provides the nice performance
0:17:41	uh this algorithm is people proposed to the european space efficiency is in is of a mission for the spectral
0:17:46	image image or these it on the is i X amount for
0:17:50	is going to fly to mars
0:17:51	you
0:17:53	that
0:18:01	do have any questions
0:18:08	i can you
0:18:11	can make any comment in regards to um
0:18:14	have the compression technique might affect processing that would occur after
0:18:18	uh the images uh
0:18:20	transmitted for example
0:18:22	yeah end member extraction or some sort of classification task
0:18:24	yeah
0:18:25	uh that's something that wouldn't propose lossy compression to a remote sensing device the re scared about the potential negative
0:18:32	effects of lossy compression so
0:18:34	uh we we were that experiments in the past with that's
0:18:37	and my feeling is that if the mean square error so you have several quality metrics cd can
0:18:42	to measure that not just mean square error the maximum air
0:18:45	spectral and will and and a lot of the matrix but
0:18:48	my experience with that
0:18:50	a if the mean square error is low in a a small have then everything we were very nice
0:18:54	and it's uh in
0:18:56	for for this kind of missions you definitely want to uh to keep the is got a sufficiently small
0:19:01	uh for not a hyperspectral image but for a spectral uh says i'll um
0:19:06	existing since C systems actually use a compression
0:19:09	spurt five does use lossy compression
0:19:12	at a bit of
0:19:13	well i think a three per pixel from paid
0:19:16	and the can a set of in just lossy compression so
0:19:20	uh uh the government agencies which are using problem funding they don't really not care about a lot of compression
0:19:26	but people
0:19:27	that the private companies they they care of what's so my feeling is that
0:19:30	compression is not a big deal
0:19:32	uh are exceptions obviously if the mean square error is is small enough run problem uh comes for example from
0:19:38	applications like a novelty detection
0:19:40	where a a large at so the one on one single pixel can actually by the result of a normally
0:19:45	detection so one has to be
0:19:46	to have a in some ways but for classification
0:19:49	my feeling is that a more less goes we the mean square error if the means got is low
0:19:57	we have time for a a question
0:19:59	quick
0:20:03	"'cause" take a speaker

LOW-COMPLEXITY PREDICTIVE LOSSY COMPRESSION OF HYPERSPECTRAL AND ULTRASPECTRAL IMAGES

Image Coding

Presented by: Enrico Magli, Author(s): Andrea Abrardo, Mauro Barni, University of Siena, Italy; Enrico Magli, Politecnico di Torino, Italy