Speech Transcript - Machine Vision Solutions for the Forest Industry

0:00:15	and the ladies and gentleman and it's my blessed to be here in sandy uh
0:00:22	but no and in this uh conference and then i uh i have a real
0:00:29	my presentation uh
0:00:32	i had quite practical and at the channel level i'm an engineer i'm not the
0:00:38	mathematician so hopefully this is okay to you
0:00:43	so um
0:00:45	and i will talk about the mess envision solutions and especially if although forest the
0:00:51	industry
0:00:53	and uh maybe the next slide that was more about motivation this is accomplished every
0:00:59	of my university in that in that a lot better and the and this area
0:01:05	is that so called in is like this trick area so it's full of the
0:01:09	lakes and forests
0:01:12	and the forests it is a big business big in the study in that area
0:01:18	so actually it's the area where there is the most then set at a number
0:01:24	of while males and solving this in the world
0:01:28	so it's quite natural uh since the six kilometres from our university there is a
0:01:35	big by a male and there are two other paper mills quite close
0:01:39	um about um
0:01:42	L one university uh it's one of the three a technical universities in finland and
0:01:49	here are our main focus areas
0:01:53	so green energy and technology
0:01:58	sustainable value creation and international have of R S and relations and you can see
0:02:05	because we are at the border is the closest to university to rest of the
0:02:10	so that's quite a match
0:02:14	so um and also my uh research area in must envision connected the forest industry
0:02:21	it's quite close to these calls
0:02:24	as you can later like to see
0:02:33	so this is our motivation this is probably very clear to you um what is
0:02:40	the motivation of computer vision and machine vision but from the practical point of view
0:02:45	at six i especially these and these are not of fake pictures these are real
0:02:49	pictures this is a person in uh mail paper mills laboratory to in serious uh
0:02:57	quality assurance testing
0:03:01	so it's mostly damman wiley and you as you can see it really minor
0:03:06	so there are some are based but it's
0:03:10	it's uh bring it on a piece of paper board and then this person is
0:03:15	used in magnifying glass checking
0:03:19	is it is it okay is the quality okay what this the prediction of the
0:03:24	usability of this um at the
0:03:27	and that this is how it should be done
0:03:30	so bill built in uh matching wits that's the same thing and the in the
0:03:37	standard last way one of the channel and C is there is that this depends
0:03:42	very much on human body and the uh and uh be right is based on
0:03:49	individuals a preference is still there is a slight variation
0:03:54	they are and also this monday morning and uh friday afternoon effect too
0:03:59	and this is a standardised way to do it
0:04:03	since this information goes from paper mail yeah males from this uh paper manufacturers to
0:04:11	a printing house that what is the quality of the product is any
0:04:16	and actually printing houses are doing their own similar tests so as to where E
0:04:21	Y
0:04:23	that opinion these the same
0:04:26	so this uh would be a standardised solution is
0:04:32	so uh
0:04:36	another thing of course is that this is similar don't knows task and some that
0:04:40	cell or spencer's at in the industrial environment
0:04:46	and uh this motivations i think you have a scene that it's fast the more
0:04:51	accurate less expensive i wouldn't say gee but i would say less expensive because that
0:04:57	must be sent to quality and easy to use is important
0:05:02	this
0:05:03	um people who will be systems they sometimes will get that it must be easy
0:05:08	to use
0:05:10	and sometimes the realm us to do not some in for my mation or two
0:05:15	months options which you never actually
0:05:18	so is it the use
0:05:21	and it's a good thing
0:05:23	and motley question of course is the replace human recent totally in this case that
0:05:29	somebody ways
0:05:30	so no human of we need it anymore at all a mess in instead
0:05:36	ses human-based and so sometimes to keep some hint this is mainly for example in
0:05:43	medical image processing
0:05:46	and then perform that's the time possible to human vision system you can imagine some
0:05:51	online uh solutions where you simply don't have ability to be so fast
0:05:57	when a paperweight is running the that the meters per second
0:06:02	so you that you can't handle that time sequence in your mind
0:06:07	or simply we don't have always in system can see spectral information for some
0:06:14	so this is our um
0:06:17	uh
0:06:19	some kind of uh starting point
0:06:23	and uh
0:06:25	this is just same shortly about the low-level auditory machine vision and but the recognition
0:06:31	of auditory we have we have around the when the people
0:06:35	doing this results and we have a it you kate the twenty seven uh doctors
0:06:41	in this area so in machine vision and many times the uh these theses they
0:06:48	are connected to real problems it's comes from industry go from bases
0:06:56	and we have around one million euros but here is how much it and six
0:07:00	hundred thousand euros just
0:07:02	is um
0:07:05	um external funding and of course L one bit main goal is to produce a
0:07:12	speed of common one is where these ideas could be implemented as soon as possible
0:07:17	in practice
0:07:24	um generally however is not only supporting building a built in a visual inspection systems
0:07:31	like here we have a many other areas like a fire model correlation here is
0:07:38	a bacterial a matter
0:07:40	so that sensors have a it's a consist of that there is not on any
0:07:46	electronics and then intelligent robotics and object recognition here face recognition and one area which
0:07:54	we especialised this medical images analyses and especially analyze civility not images so to see
0:08:01	some uh this easy some problems for example cost by diabetes
0:08:07	so that is also area
0:08:13	um if we think about the visual inspection actually it's uh um
0:08:20	it's a it's a actually what white lots color correction so there are many other
0:08:27	universities involved and also many real eight comp running so a lot of a lot
0:08:33	of all operation is needed and if we need to think about how one of
0:08:39	our main objectives is the develop novel results efficient and environmentally sound image in image
0:08:47	processing and image and the lies in met thoughts for machine vision applications especially for
0:08:53	overall quality assessment and control so the idea is to all these single measurements together
0:09:01	to be able to control in general
0:09:06	the whole industrial process
0:09:10	and we aim to offline and online solutions obviously that is a lot part of
0:09:17	our example in the preview on the previous slide that it is offline
0:09:23	uh solution because it happens in the laboratory but then we need also online solutions
0:09:29	directly uh in the industry
0:09:34	and mainly we are focused on forest and printing industry
0:09:39	so at the lower product already and the factory
0:09:42	about thirty level
0:09:48	and um if we think uh how
0:09:53	what is actually happening there you see that this contains look over different the industrial
0:10:01	is text there so it's actually quite complex uh combination before you end up to
0:10:07	get to have a lot of paper there are so many things happened before that
0:10:13	so in bright this we have a some would might be real then we have
0:10:19	well and then paper and finally some print so something you can print or
0:10:27	uh and uh if you think wow why to do it okay the results efficient
0:10:34	uh resorts in efficient uh environmentally sound production with no one data using less rollment
0:10:41	the real water and energy and actually this is a cool thing because this to
0:10:47	be environment three uh one shows uh use the same as the road using it
0:10:54	efficiently so there is no contradiction between these things because if you use less raw
0:11:01	material water and energy you save money and you also save environment
0:11:07	and the no other known quality means that the uh you use should put quality
0:11:13	as low as possible and it still feasible quality so it's still suitable quality you
0:11:20	can always book the parameters so high that the quality is high but you waste
0:11:28	a raw material what the energy and so on and you waste money
0:11:32	so as low as possible but still reasonable quality that's the idea so that these
0:11:38	than on quality and we'd without having some measurements and they interpretations we'd be gone
0:11:44	we don't know what these what these that quality so that these what we try
0:11:50	to write the so
0:11:53	actually our uh meals in finland they uh environmentally in very good condition because the
0:12:01	lakes around a white lee
0:12:05	and that we actually condo for two only you of lakes because then the for
0:12:10	its we like to do so we must really think of environmentally seems to
0:12:19	so um
0:12:21	idea simple in the way imagine they the automatic analysis and the in the forest
0:12:29	industry
0:12:32	uh we talk about this will also line quality has assessment business but this done
0:12:37	right traditionally manually
0:12:40	and the campy time calm you consuming and labourers to
0:12:46	then we have online quality assessment and then tool for real time control actually these
0:12:52	two areas and the research done their support this idea
0:12:59	to have a some kind of real two-dimensional uh quality control based and to advise
0:13:05	dimensional units
0:13:08	and uh what we have a
0:13:12	study we have had we see this as a as a collection of uh research
0:13:18	approach it so we invested
0:13:21	so we have one uh rotate which is dealing with power
0:13:26	so how to put this uh raw material in the bases for paypal make
0:13:33	and what could be done there
0:13:36	then the one thing is that what can we do um
0:13:42	connected to a paper messy so when real paper making is happening so and there
0:13:47	is a long paper my signal for one hundred be there with the long volume
0:13:52	along more what we can do there
0:13:55	and of course these are connected so if you know about the quality of raw
0:13:59	material here you can you can put this information here and you can take advantage
0:14:04	of that
0:14:05	and then this is your final product okay um here you can't do anything anymore
0:14:13	it's already they are so paper so after that you can only predict how to
0:14:17	print on and what is the pretty that uh applicability and run ability when doing
0:14:23	some pretty
0:14:25	and then
0:14:27	then uh we test the quality of the paper and actually finally interesting thing is
0:14:33	that how we see that you meets or something pretty become paper that can be
0:14:40	mortal human beings set preferences is connected to some physical much innovation measurements
0:14:48	so that's is that and the system will be nice that you put the print
0:14:53	into that find a much emission system any else that how human beings are uh
0:15:00	but if you're in this but it's
0:15:03	"'cause" uh the quality of an image but in it of course it's very important
0:15:08	you don't want to see some bad quality images but what is the bad quality
0:15:12	and what is that what did when i ask if i thought make the test
0:15:17	we you i assume there are some but em sis but there are some form
0:15:22	of that are still so this is a big question and this actually relates to
0:15:25	any image information problem whether it's print date or whether it's speech it's exactly the
0:15:32	same problem
0:15:34	problem
0:15:36	it's very uh and fess in a be much fascinating problem actually and what typical
0:15:45	i will solve some results about this before and the steps here
0:15:54	and what one wants to remember this that uh need for these kind of research
0:16:01	must come from the industry so every single project we have a start date has
0:16:06	started in that way that we have once industry what do you need and then
0:16:12	we have stopped at the check what we can
0:16:15	so it's actually it's not uh selling your ideas it's uh
0:16:21	asking needs and uh formulating the questions
0:16:33	okay us then i would like to talk about some um
0:16:38	things uh which we must consider and one is to recognize imposes phenomena
0:16:45	so actually what we should and this that is the question and the who knows
0:16:53	the ground rules and why and how
0:16:56	and
0:16:58	yeah and it is it really possible the model the expert knowledge
0:17:04	and uh this is a good question because we must we must talk to be
0:17:10	a those experts there is that the what they would like to know and what
0:17:17	we could see and uh
0:17:20	sometimes it's islands in if the data that is very mobile so even the expert
0:17:25	have an ever seen they the from some certain industrial processes
0:17:30	so this really needs need something in and actually i have an all this when
0:17:35	we are having them some uh comparisons that the and i is that where is
0:17:40	the ground rules if i'd use your data and uh and some might be just
0:17:46	looks good but that's not you know
0:17:49	you should really be fine ground and maybe to other researchers publish your mate complex
0:17:55	operation ground
0:17:57	of course in this kind of uh industrial pros project said this comes a little
0:18:02	bit be late because firstly analyze the possibilities to make a real part that are
0:18:08	based on that it is it's but it will come anyway
0:18:13	and is this possible that's and that's and other is you sometimes it might be
0:18:19	very seen that the sum intensity values that everything but
0:18:27	you don't know before you think it's
0:18:30	mister in optimality
0:18:33	so what utterances in image and so uh there might be some illumination problems or
0:18:41	it might be a typical place in there or you need some uh not reflection
0:18:48	information but that some transparent information
0:18:52	and then we call to this uh do we need multimodal information that many times
0:18:57	we need so we have a transmission in its we have reflection immediately may even
0:19:04	have some extreme images
0:19:08	we call the different scales there and how to put this together so we have
0:19:13	multimodal information and how to how to put it together and whether it supports to
0:19:18	each other
0:19:22	then understand in measurements
0:19:25	so a bible college on unclear categorization some more clustering need it so that we
0:19:33	know enough about they may not then we have the first define what are the
0:19:37	classes there
0:19:39	so clear when it's be about the like categorization then it's just classification but many
0:19:43	times as a question of this especially when it's number of data that what are
0:19:48	the other clusters
0:19:51	so this might be a challenge to all of course accuracy and computation time if
0:19:56	you think of this uh paper with running that the meters per second you have
0:20:02	two options you where infrequently uh to sampling but the image resolution is uh low
0:20:11	or you have a high resolution but then you don't do it so frequently
0:20:17	and our idea is to go to this overall quality management and approaches process control
0:20:24	so step by step we more accurately know which kind of uh paragraph we will
0:20:30	get
0:20:32	and this is this is about how been and actually this goes into those of
0:20:37	previous steps as
0:20:39	as i mentioned
0:20:41	and there are some certain uh process phenomenon so probably means that the uh we
0:20:49	make a raw material for people making
0:20:52	so we need to check some certain things like uh i other some extra particles
0:20:57	they are which are not wanting to be there and the basis actually about suspense
0:21:05	and he meets so it's a liquid these is the right this is the right
0:21:09	about C have been some i'm one D the objects
0:21:14	and uh number of a number of this and what they do or distribution of
0:21:20	this on one be the object of like say that well
0:21:25	so mess are in these things
0:21:27	different kind of uh i things even for a sense and spectral image and then
0:21:33	this analysing part
0:21:37	so here are some examples here and one thing you know very important thing of
0:21:42	course is to understand the uh fibres that fibres uh distribution and position
0:21:50	so i'm not will industrial data for the uh
0:21:56	by new image and solutions and uh i'll idea is to pull inside the pipelines
0:22:03	which is uh have not been done so our data is totally novel so wipers
0:22:09	are you know interest the particles buckles and sales these bubbles the lapel approaches there
0:22:15	is some you know their problems there
0:22:18	and then how the form and the model the ground rules and this need something
0:22:22	can because if it's a mobile make the so there is no a human expert
0:22:25	that's believable expect is about
0:22:28	but the there is six but is about the uh i mean
0:22:34	so that can be used
0:22:40	uh and uh then uh one thing there is a segmentation the fibres another particles
0:22:46	in suspense and emits is still this is really weak
0:22:49	lee we thing and we have a done that the uh results and why this
0:22:57	piper's must be uh and the last is that you should know the uh three
0:23:03	species there it may sound strange that don't they know what is there but it's
0:23:08	not necessarily always the same tree and same kind of three material so uh the
0:23:14	know about these three C Cs is important and again this means that the this
0:23:19	viper is broken in the back leeway that later when you breed
0:23:24	how can they need that kind of uh fibres it's bad property
0:23:29	so the to detect number of things
0:23:32	is important
0:23:37	and we have some applications about this
0:23:41	if you call to my list of uh applications you see that should have that
0:23:46	six
0:23:48	then another use you uh not only five course but problem side important because particles
0:23:55	tell about some approach the situation and how they daily the daily and in that
0:24:00	way that um when there is a certain amount of purple bubbles the process is
0:24:06	uh okay so that such a certain size and certain volume distribution and how they
0:24:13	do it now when they are aligned in a factory they do it in that
0:24:17	way that they'll and the process for let's say two hours to be sure that
0:24:21	everything is okay but one hour would be enough
0:24:26	but you don't know that if you don't have an image
0:24:30	nonetheless
0:24:31	so you put even got terms is processed by means of halls and that would
0:24:36	be you say in
0:24:38	but you need in leeds analyze the solution for that so we have real the
0:24:43	field of that kind of a solution and we will shortly present this is very
0:24:49	impressive results and icp our conference in japan in november this is a couple operation
0:24:55	results with the tech technical university
0:24:58	we propose that matters
0:25:00	and actually this is a this is a quite challenging problem although that looks so
0:25:05	originally quite easy and if you have a some uh proposals or something to the
0:25:13	to this problem i'm very happy here you or your I S
0:25:20	so it is not easy
0:25:23	since it because these problems are not that we're a circular problems so they may
0:25:30	be deformed so actually we are dealing with the P Cs or for some kind
0:25:36	of circular say and we need quite the white mats thinking how to make it
0:25:42	that efficiently and then we must detect the problems in order to be able to
0:25:48	estimate what so from problems we estimate the volume has some value there and sof
0:25:54	detection of problems as important than in this case of course detection of the become
0:25:59	problems is more important than the than the small ones
0:26:03	but also this distribution map cell from distribution experts can tell something
0:26:11	and uh we have actually ended up matt's of thing you know feel that in
0:26:16	how we get some kind of suitable it speaks
0:26:21	and enough of those and how we start to combine of these pixels because then
0:26:27	never ideal unfortunately it's a liquid and it's also problem of three-dimensional imaging
0:26:36	there so but uh i must say that then we have got quite a very
0:26:41	good results and surprisingly this uh this uh false negatives some false positives they start
0:26:50	to compensate each other and uh not most of that we can detect right so
0:26:57	we can really get true positives but the so surprisingly um well then of course
0:27:03	the question these that what is the what is the ground and that because of
0:27:08	bases you we have given the experts to select a level of confidence
0:27:16	the whether there is a war or and solar or totally unsure
0:27:21	because it's not so clear what his answer to so
0:27:27	so these contains all the nearly all the possible problems i enumerate the in the
0:27:32	couple of slides ago
0:27:35	okay but that the that the say that the pc is challenging from the uh
0:27:39	point of view we mix process
0:27:43	some people sometimes say that all you test an application itself see but we are
0:27:49	dealing with the fundamental issues of image processing
0:27:56	okay then and that other thing that the uh okay fibres
0:28:02	um couples
0:28:04	so five resigned the liquid
0:28:07	and this is bob suspension bubbles in about suspension but this is the right biopsy
0:28:14	so when the power was ready each tried to be more to uh to a
0:28:19	paper mail and this is really a tried C and E there should be only
0:28:24	uh suitable rollment the we only but all these so with this dog you the
0:28:29	there's something and one it's um the particles because of what's something's there might be
0:28:35	even this and they are
0:28:38	so any anything
0:28:40	uh
0:28:42	the right of course the different kind of categorization even plastic
0:28:47	or some unwanted what would my
0:28:51	so it would be good to categorise these
0:28:54	different kind of uh the particles "'cause" that's there was the quality of the biopsy
0:29:00	if there is no the particles it's a it's a great but the problem she
0:29:05	for paper make
0:29:10	and so we need to a here the segmentation and then classification of these this
0:29:15	that the particle physics
0:29:18	once again we are dealing with the quantum fundamental issues of image processing and analysis
0:29:27	and here is a slight problem because we must do this testing that the laboratory
0:29:32	label
0:29:34	see scenes uh
0:29:36	uh if we do it online
0:29:40	then uh we wouldn't have time to do analyses so what we have done we
0:29:45	have a uh at the laboratory we have a creative israel about six and in
0:29:50	that way that the experts have really put the ground rules there
0:29:55	and then we can test whether we can recognise this separate that the particles
0:30:02	at the moment there is no analyses
0:30:05	you just put up sixteen and you hope that you will get would like
0:30:10	so that's motivation to build the system to do that and S
0:30:14	okay then let's go to the know now we worry at in the powerpoint so
0:30:19	a would deinterleaving liquid and do some processing there it can be column mechanical then
0:30:24	you get that right biopsies then you put these power seats uh to a paper
0:30:29	machine and then you start to make paper
0:30:33	and the um main question is that it is that it presents and this represents
0:30:39	a paper a bit so that what is running from the beginning to this and
0:30:43	when the product is ready uh the question is that can we see here or
0:30:47	something
0:30:49	something which we should uh is that the react to something and actually um now
0:30:55	there are some line scan or switch to body weekly scanning but that it's of
0:31:00	course different than the two-dimensional to time instant no uh real time image and uh
0:31:06	because of scanning
0:31:08	we actually can on see this kind of till the problems which you may see
0:31:16	basically simple maybe you somehow can see that and uh of course of human being
0:31:21	can see that's that the meters per second so close to plastic on C so
0:31:26	you're we must rely on these mis ordering systems and line scan risk on to
0:31:31	it because the coast line by line so idea is to take a real the
0:31:36	dual diamond two dimensional units uh challenge there is that we need eight income it
0:31:42	is
0:31:43	and put all these images together
0:31:46	and so it's quite the real-time problem and also build in an image problem too
0:31:53	because of uh some uh geometric uh imaging problems and things like that but that
0:32:01	we can do so we can find
0:32:05	fine these kind of because these problems these sideways problems they are because of some
0:32:13	uh not so accurate tuning at the beginning of the process and you can tune
0:32:18	the parameters there but you must know in which way so this gives feedback
0:32:24	they're to do this do you
0:32:29	so how to produce and use real-time two dimensional web white image state the problem
0:32:35	of or on online uh control
0:32:41	um then we can go uh further in our analysis uh we can we can
0:32:48	focus on at some uh more accurate level uh and uh get some information and
0:32:56	why we uh this is some kind of closer close up a information from that
0:33:01	with uh paper web that area
0:33:06	and there are some uh phenomena we should see that how actually the white borders
0:33:12	are grouped there that's important but if that my signal then we can then we
0:33:19	will get nice paper and print on it then we apply but there might be
0:33:22	some grouping problems and this that was also about
0:33:27	do
0:33:29	and uh here we can do this a multimodal problem because we have our lower
0:33:34	low resolution high resolution images we have a reflectance it means it means we have
0:33:40	transmit dance emits we have seven of images and how to help able to put
0:33:44	these things together and we can have some other kind of measurements to so we
0:33:48	have a bit different kind of a units in many thoughts and how we can
0:33:55	put them together so here for example the segmentation and analyze of three five or
0:34:01	flocks using transmis tense images
0:34:05	and when the orientations are nice then it's okay maybe when you have a at
0:34:10	the right that the other paper you know this that it that the there is
0:34:14	very nicely there some certain direction about another direction that some and the that's because
0:34:19	of
0:34:21	on purpose it's made in that way
0:34:27	and uh
0:34:29	so detection of surveys orientation is so one thing to uh check and then of
0:34:36	course the challenges here that how we get these correspondences when we have a different
0:34:41	kind of image state that from the same
0:34:46	so we have a correspondence problem they are
0:34:49	so how to do image registration
0:34:58	and why we are interested in surface orientation is that it's affects directly but
0:35:07	and then after we have got this paper then we can test what we call
0:35:13	and we have build the system to do this that and the vices
0:35:20	and so is to do this
0:35:23	the similar kind of solution to everybody
0:35:28	not depending on the human brain
0:35:31	and let's see what we can see from that so we can we can have
0:35:35	a different find of uh measuring devices here and the what we i'm gonna speech
0:35:41	in this uh for one is an to be in and evilness of print in
0:35:45	misery you know name it so it's modeling so we should pretty that the uh
0:35:51	if we bring the blue or lost by their how well how you nibble form
0:35:56	this
0:35:58	and can have some kind of my sermon some estimate and it's called model in
0:36:03	the in the printing industry
0:36:06	"'cause" if that was crying is not uniform it disturbs you i and you think
0:36:11	it's about three i think this is maybe
0:36:14	but pretty but you can see if you if it wasn't
0:36:21	so that's a one thing somewhat yeah modeling detect the modeling
0:36:25	another thing is that the
0:36:28	how will this step putting in happens on the survey so it's called haley a
0:36:34	test so detection of missing dots in that this but it is bought to say
0:36:39	about this that this is mostly about some uh frequency feel that in of different
0:36:44	kind of a yeah yes so that's a technique and then the question is then
0:36:48	how we weight of those different frequencies
0:36:52	it's as the linear combination or are there some coefficients with the with the combined
0:36:57	or something like that here this is very simple test the new print of on
0:37:02	a paper sleep uh don't red dots and then you simply calculate which starts are
0:37:08	missing because it means that the so face is not so cool and this simple
0:37:12	man in this woman in the picture was really count in and where is that
0:37:17	when the at the missing dot from the beginning and he may shirt she mesh
0:37:22	shirt that in millimetres and role there that ninety seven millimetres from the peak that's
0:37:27	the quality while the longer the better
0:37:32	so and we also made it is uh there's the one we used it is
0:37:36	there's another and we automated the this column
0:37:41	so ago that was automatically this of course the question there's a one challenge that
0:37:47	what is the mission of how much must be missing or and then person be
0:37:53	left there "'cause" of course this is natural situation this is not by nist
0:38:01	then uh um some small be there is how they sexy carissa must of these
0:38:08	is about that how these uh
0:38:12	this actual but it looks
0:38:16	what we can see from that from a single small not
0:38:22	it's obvious where about the predictability this is a part of an ability in printing
0:38:27	the it's called the peeking test that you press harder and harder that the that
0:38:33	uh test sleep until it breaks
0:38:39	or some i'm a other that meets this will start to happen to
0:38:43	so for usually uh and this is the border it's they're starting to see some
0:38:49	phenomenon the subways and when it's a paper it's imprecise the der of
0:38:56	so it totally breaks and the question is that how long you grow until it
0:39:01	starts to add to your break and once again people are doing by and calculating
0:39:08	how many millimetres from the beginning to that place and of course it's a at
0:39:12	the request and we have this happens this a one D uh phenomenon of the
0:39:18	early
0:39:20	and we have a thesis about
0:39:24	this topic to so this terms of butter unappealing the longer the better
0:39:28	from the printed point of view
0:39:34	and uh here is an example about the speaking tests he actually here you see
0:39:39	what happens the paper it simply start to tear off and this is not what
0:39:46	if the value and what and or it usually neighbour will S is that the
0:39:51	overall but there will become sums of place the effects there and this is able
0:39:57	to a little bit harder to define because there is no except order here is
0:40:02	quite clear that simply breaks here and that's the distance
0:40:08	and actually this company's we sell in this device it's called laplacian technologies our doctors
0:40:14	have a used up list a all pole one and they yeah selling this product
0:40:20	uh for this text
0:40:26	and this is about this missing don't the problem that
0:40:31	this print starts here and you put in the pixels and then you start to
0:40:36	check that how many pixels up the uh missing and that when the L one
0:40:41	is the
0:40:43	is the definition i is that where these come from and i think in a
0:40:47	laboratory level in those paper means they have just decide that it's twenty
0:40:55	so it's some kind of practical
0:40:57	practical value but you should always have some but i think the uh this twenties
0:41:02	in that way that you would never go to the end
0:41:06	it's comes from that it's better not to say that like one of our doctoral
0:41:11	candidate set in the doctoral dissertation examination that why is this threshold speech P and
0:41:17	he replied that that's a good number
0:41:19	but i don't recommend this was there will prefer that with questions after that
0:41:26	okay then uh and next issue when it's really pretty print it on um how
0:41:32	we can uh estimate the quality so which one which looks better and why because
0:41:39	you may have a different print the brain there's different papers different images how do
0:41:46	you can say that which one is better
0:41:49	so what we need to we uh selected some S in recent measurements
0:41:56	and we have the department of psychology from helsinki university to do in experiments with
0:42:03	humans asking their opinions and then you know how to do it we don't know
0:42:07	as engenders but they know how to do it so they collect be it's a
0:42:11	human being subjective but uh
0:42:15	what if the attributes and then to be put them together to get some kind
0:42:20	of well overall image quality index and with this index we can we can have
0:42:25	that
0:42:27	double for example then images and we completely in order and hope that it goes
0:42:31	in the same way as human being would have prefect
0:42:36	this is quite challenging results research to
0:42:42	uh in this results we have a couple of things to do is of course
0:42:46	because we need to expect around rule so we need to digitise image so we
0:42:51	must the register very uh accurately this data test images units and that's a one
0:42:57	uh thing is this uh and uh occur missile inequality and uh locate in those
0:43:04	areas which we have miss that one doctoral thesis is about that
0:43:11	so the predicting human humans experience of image quality for print the image
0:43:20	and the then we calculate the this uh like a bayesian network model that
0:43:26	weights the all these simple things uh seem not seem well in a way simple
0:43:31	measurements affect the most and um it's maybe not surprising that close is one of
0:43:37	those modeling is very important and colour but this time it is a really experimentally
0:43:44	proven because we really fast
0:43:47	is a human being set a group to tell they opinions and these opinions were
0:43:52	really model and compared to
0:43:55	real but uh physical my matching pursuit measurements
0:43:59	so this is not the case this is really the result
0:44:05	and in this thesis by almost and allow you can be more
0:44:13	um then uh i'm approaching the and only a couple of slides left here an
0:44:19	actual level with you with numbers because these are selected slides not fifty so far
0:44:27	so um and that next question is that what i've mentioned is at that the
0:44:32	how well does this depend on which kind of image you
0:44:37	and yes it does very much
0:44:41	so then we would need the system to finalise what is the content of an
0:44:47	image
0:44:49	to have a
0:44:51	E meets content based this one lady
0:44:55	so it would change based on image content because many print this or see it
0:45:01	on screen it's it has different preferences we deal than scene this and this close
0:45:08	to very current hot topic uh visual object categorization so how to and especially in
0:45:15	this unsupervised manner because we are not interested what are the objects at all we
0:45:20	are just interested what objects are important
0:45:23	and uh
0:45:26	we don't even without knowing what though they are
0:45:29	well let them to affect this decision making
0:45:33	so we are dealing with unsupervised the visual object that the crisis and then we
0:45:39	need the ground rules so we did i tracking
0:45:43	and to see what are the salient features for human brains beings and one in
0:45:51	a thing is that what are these images how what is the perfect test images
0:45:57	that that's a that's a big question because uh you may have a some kind
0:46:03	of side effects if you don't the
0:46:06	landay's in which is carefully enough and that's why we have from all the university's
0:46:10	media technology specialists who were preparing these images for us
0:46:17	and uh
0:46:20	uh
0:46:21	and then uh we ended up really the to do this uh visual object categorization
0:46:27	results and um well idea is uh use in self organizing map uh of for
0:46:35	a clustering of these objects and finally get these categories so but the visual codebook
0:46:43	use in
0:46:44	self organizing map the component it and the also some uh evaluation techniques then we
0:46:52	have a new uh we use uh published new images uh it said so randomized
0:46:59	object so we make the tricks with the for foreground some background images and also
0:47:05	saliency detection
0:47:08	and this was connected to the sidetracking the results
0:47:12	and all also
0:47:15	also better techniques for what this categorization
0:47:19	and uh we have also published the uh totally new limits that it's dealing with
0:47:27	up static images where we are interested this ci tracking when we have a collapsed
0:47:32	of images so we don't have any background to
0:47:36	there we have learned to do something
0:47:39	and then we take that how i look at the conversation goes with the results
0:47:45	from human being which was of course a lot of variation
0:47:49	because they were absent images
0:47:52	this is will be published in icp
0:47:56	we sat besides this doctoral dissertation this
0:48:02	there are there um many other details yeah in with these applications for the uh
0:48:13	forest industry but since i have limited time these i want to the selected so
0:48:20	so thank you very much
0:48:40	yeah
0:48:51	that i
0:48:57	i
0:49:03	it is difficult to the question and answer is that to my both are important
0:49:09	so actually what we finally would like to know is the volume of a of
0:49:15	those problems there
0:49:17	so then uh we need to detect the problems and but also distribution tells about
0:49:25	the process so about been experts can deal when the distribution changes
0:49:32	so both are unimportant and of course we make a quite a big assumption here
0:49:37	that uh if we think that this is a two dimensional bubble uh you meets
0:49:44	tells about the volume button roughly it goes right
0:49:47	in liquid
0:50:09	uh in this case would actually be are not and we um because these imaging
0:50:15	conditions are quite a typical bits totally inside the pipeline and some liquid is run
0:50:20	in there and we rely on that fact that at the place where you we
0:50:24	are sampling three presents the whole like with the area that might be a little
0:50:29	bit risky assumption but uh we uh we had second the possibility of having the
0:50:36	uh more images
0:50:57	uh
0:51:01	yeah
0:51:12	um
0:51:14	yeah we have a certain number of people first of all form in this uh
0:51:20	this uh a human soppy neon and then uh from their we tried to from
0:51:27	very low level attributes to build the higher level attributes
0:51:33	and the then uh koenig can make them to the physical measurements of is eight
0:51:41	samples of since we know each sample and we note no opinion of a human
0:51:46	being of from each sample and then uh then be uh for uh yeah
0:51:57	start to see the changes um actually it's a it's just send set us up
0:52:02	uh some kind of a curve but unfortunately i don't have the curve here we
0:52:08	very use it's like a rock or
0:52:10	and you check that how close you must goal that yeah this uh but utterance
0:52:15	order
0:52:17	close to the to the mean opinion score so that is the idea so it's
0:52:23	a little bit X experimental you have i understand your question but the that we
0:52:27	can say that it's a direct physical uh example exhumation which tells that very good
0:52:35	question
0:52:44	yeah this one
0:52:52	i
0:52:55	yeah
0:53:01	i
0:53:18	um
0:53:20	i actually um
0:53:24	uh we were of course we must have a normal each of these classes for
0:53:29	unsupervised and so what about what about what the band when checking the result of
0:53:34	an unsupervised uh
0:53:37	clustering but we have to do that without any knowledge of those classes
0:53:47	yeah so we have like we have about like uh database and there we have
0:53:51	exact knowledge of those categories and we at the end we check how it goes
0:53:58	so actually it's a huge the map where all these objects are clustered and then
0:54:05	we can start to see how it goes
0:54:09	and uh of course results are not so called then in a in a supervised
0:54:14	but we were quite surprised about one thing that is that the how a unique
0:54:19	the ballot the six people have used like something like uh having twenty of the
0:54:25	images and saying something about unsupervised clustering so we used hundreds of maybe six
0:54:32	for this
0:54:45	oh
0:54:51	i
0:55:01	actually we are trying to mister those things which affect the quality and actually they
0:55:10	affect the usability uh process is there
0:55:16	yeah
0:55:21	hmmm
0:55:24	well as well that's a of it that's a physical actually property of the paper
0:55:29	how we cepstral observed scene for example and that's quite the engineering style of testing
0:55:37	that there is really putting on the paper and then start into account where other
0:55:41	areas it up selves okay and we really doesn't to properly
0:55:47	yeah and before it happens that paper then yeah because actually we pretty there is
0:55:56	there are not so many technical problems anymore we have a somehow assume that
0:56:02	when the paper is okay it's not anymore so pick pretty the quality papers okay
0:56:17	ooh
0:56:21	i
0:56:24	uh excuse me what it what was
0:56:37	i mean how well you all these solutions might be what we said difficult to
0:56:42	say uh um then uh of course you can calculate the salary of those people
0:56:48	work in that club or authorities
0:56:50	and sela and the than the price of that uh
0:56:55	that solution that machine so that so quite easy to calculate but then when you
0:57:01	go to be clear peak approach to cease in paper making or impala been
0:57:07	that's a little bit more difficult to calculate but if you see that clear peak
0:57:12	lisa in the use of a raw material and water and unity i think the
0:57:17	estimates have been even up to forty percent
0:57:23	which is a lot of money

Machine Vision Solutions for the Forest Industry

Invited talks

Heikki Kälviäinen