1 kostadin n. koruchev universidad autónoma de madrid, spain e-mail: [email protected]

30
1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: [email protected]

Upload: emery-hines

Post on 22-Dec-2015

235 views

Category:

Documents


4 download

TRANSCRIPT

Page 1: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

1

Kostadin N. Koruchev

Universidad Autónoma de Madrid, Spain

e-mail: [email protected]

Page 2: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

The talk

Presentation of my university

Figure design for coding with orientation.

Brief presentation of the main themes of my research.

Page 3: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Universidad Autónoma de Madrid (UAM)

Located near Madrid. The official university of Madrid autonomous area.

www.uam.es

Page 4: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Universidad Autónoma de Madrid (UAM)

Universidad Autónoma de Madrid is one of the top-ranked Spanish Higher Education institutions.

It has 94 Ph.D. programs and 72 master’s programs

Over 32,000 students and 2,200 faculty Its campus is located 15 km (10 miles) north of Madrid’s

center and it is comfortably reachable by public transportation http://www.uam.es/presentacion/campus/

The university hospital La Paz is the biggest in the Madrid area.

www.uam.es

Page 5: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Escuela Politecnica Superior(Computer Science and Telecommunications Faculty)

New and dynamics faculty, founded 1993150 researchers from which

about 60 permanent staff (equiv. to prof. in Japan).

About 1500 undergraduate students.

150 graduated students.

www.ii.uam.es

Page 6: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Figure design for coding with orientation

The problem – to find figures suitable to code information in machine readable way, but hardly noticeable to humans.

Why? Interior design.

Machine readable orientation.Visible by humans – must be acceptable as esthetics.

Printed material Small markers, hardly visible for the humans that can

intermix with the printing (CLUSPY).

Page 7: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Previous works

S. Nashizaka, T. Tanikawa, IEEE VR’07, ACM VRST 09,

Use of markers that are selected by the user: Information coded by

the rotational angle.Very general figures.Using p-type

Fourier transform.

Page 8: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Previous work

Results: S. Nashizaka, T. Tanikawa, IEEE VR’07, ACM VRST 09, Results: Dependent on the figures. From 85% to 95% correct

determination.

Page 9: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Previous work

Kato and Kanev, 12th International Conference on Humans and Computers, December 7-10, 2009

Selecting the figures one can achieve better results. One do not need markers.

Predominant orientation is enough. Argue that L-like shape figures will work well.

Any L-like figure will work well. Recognition – work in progress.

Requires combinatorial algorithms (convex hull). Mix with CLUSPY – extremely sparse coding. This work is prolongation of this ideas.

Page 10: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

General considerations

Acceptable for the humans (artist), Machine readable (formal math criteria). Efficiency.

Characteristics that can carry information: Position – OK but only relative position. Size – Depends on the distance, uniformity. Form – well known problems of image recognition. Orientation – Yes! Also mirror symmetry (left/right variants).

Not every figure can carry this information: Symmetries. We ought to cofactor any symmetry.

Least symmetric better.

},{},,,,,,{ RLZZMMLnS yx

},,{},,,,,,,{ mod NanRLZZMMLnS yx

Page 11: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Proposed solution:

Use the moments of the figure. They are easy to compute, reliable up to order 4-5 for usual figure sizes and are noise resistant.

-- scale and translation invariants.Easy transformable by rotation (tensors).

2/)2(

,

,

,

,

/),()()(

/),(

/),(

),(

mk

yx

my

kxkm

yxy

yxx

yx

AyxfMyMxM

AyxfyM

AyxfxM

yxfA

2mk

Page 12: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Rotation

Fixing the rotation:Rotating the figure among its axes of the ellipse

, makingThe angle of rotation is

Conditions the angle to be defined:

Practical requirement:

021120 ,, MMM .011 M

0220 MM

).36.0(/, 120020220 CMMMM

2/),2(atan 0220112 MMM

Page 13: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Further requirements

Recall:

(determines but

Determines uniquely.

Determines Z uniquely.

.actually,0 23030 CMM

403321 ||or|| CMCM

./ 12002 CMM },,{},,,,,,,{ mod NanRLZZMMLnS yx

0mod 180

Page 14: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Classes of figures:

Too round:

Too symmetric by X

Too symmetric by Y:

All criteria satisfied:L shapes are in.

./ 12002 CMM

.230 CM

403321 and CMCM

Page 15: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Recognition

The reverse problem must be solved.Precision of the discrete parameters –

absolute.Precision of the angle:

Less then 1 deg. error by pictures of size 200 x 200 pixels.N=50.

Size 40x40 less then 1.3 deg. N=400. We can encode >6 bits per figure.

Page 16: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Biomedical Applications and Biomedical on-line Processing. The importance of the problem:

According to IBM GTO 2010 Bioinformatics is one of the main technological areas for the next 5 years.

The state of art of the automation in the hospitals. Disconnected autonomous devices. Disjoint databases. (patients, hospitals, health insurance providers)

A lot of potential in the integration of these data. A lot of value for the patients.

Especially integration and distributed processing of the on-line data can give significant advantages to the patients.

Page 17: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Specific area

Biomedical monitoring application. Problems – the different modalities of the monitoring are

not integrated in the automatic systems Neurology – epilepsy.

Selected because the video data is the most demanding data-stream that we found.

Normally the physiological data are observed in periods of several seconds.

ECU, Preoperative observation, Pseudo epilepsy.

Page 18: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

The Problem

Page 19: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

State of art

Build a model and detect the pattern37% (at most) of the seizures are detectable. In practice some 15%.40% of the cases not detected in ICU have fatal

exit. It is clear that better detection can help.The problem – the model is not complete.Our approach – find the part that do not

conform the “normal” model.

Page 20: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Proposed Solution

Main problem – having just EEG there are many false alarms (3-4 times more).

Analysis of the problemHuman experts use the variety of signals – EEG,

Video, EMG, etc. to detect the situation of epileptic seizure.

The detection should be multimodal on-line and independent of details.Novelty detection.Specific seizure detection.

Page 21: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

The novelty detector – the most peculiar partIt is that carries the maximum

informationIt is that do not confirm any

previously known modelUnique (or rare).

[K.K.&E.K. IWCIA 08], images

[K.K et al. ISVC 09] – EEG, video

Page 22: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Novelty detector The most peculiar part can be mathematically defined. The exact solution is combinatorial problem and the time is

not affordable. It is a problem defined in space with dimension several 1000.

In probabilistic terms it can be solved in time proportional to the data volume. We use random projection in order to solve it. Close to PCA. The probabilistic solution is feasible for all signals with exponential

decay distribution longer tail.

IMAGE-VIDEO EEG/EMG but… ECG

Page 23: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Solved Problems Epileptic patients in the ECU.

This is live saving technique. Epileptic patients in preoperative observation (holter).

The efficiency is much higher. Search in vary large databases of images. [KK. Pat.Rec. 08]

A single Rx unit produce some 5 images per minute, some 1200000 per year.

The most peculiar part can be conditioned – most peculiar regarding these samples – the search is very efficient.

Trying to mix these techniques with the modern bag of words. Currently bag of words can achieve good performance in the range of

up to 3000 images. The mix can solve a lot of problems.

Page 24: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Novelty detector – Gaussian projections

The signals with autonomous regulating systems (cardio activity, blood pressure, corporal temperature, glycemia level can not be treated that way.

Compression codes – compress the signal with for example wavelets and use the compressing components. They have exponential tail distribution.

These signals are important – the deviation in its complexity has high predictive value for different pathologies. Example – body temperature, glycemia level. To appear [M. Varela, K.K., BioSignals]

Page 25: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Open problems

Evoked potentials by epilepsy observation. The method of EP:

There is a signal provoked by some stimulus. The signal s(t) is smaller than the rest of EEG, that is regarded as noise n(t). The observation

u(t)=n(t)+s(t). The signal is extracted by averaging various instances of the observation.

EP(t)=Si ui(t)/N.

Example: The problem – the patient can not say his name. We do not know where the fault is.

Reception (auditory) Recognition (associative, auditory). Conscience MotorEP can give this information.EP standard procedure is unacceptable inEpileptic patients.

Solution – use the “natural” stimuli. At the moment – only sound due to the

time resolution. Observables – before and after the crisis

P1, N1, P2.

Page 26: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

Thanks

ありがとう

Page 27: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

6 The problem

Orientation – the main component. 10 deg. precision more than 5 bits.

Useful for a wide range of figures. The features are generic.

Easy to decode. (segmentation, extract the features, calculate the code). HUD – up to 0.6s, Computer vision – 1/30s.

Find the figures that has detectable components of S. Find formal criteria to distinguish these figures.

Page 28: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

8 Rotation invariants

Hu invariants. Useful to detect the figure up to 8-9 figures. No information precisely about S.

Successive approximations.First moments – dot or disk.Second moments – ellipse.Next -- Legendre Polynomials – like quantum

orbital moments.

Page 29: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

13 Can it be decoded?

CLUSPY like encoding with no marker. 120 quantization of the angle. Random angles. Save the angles for the decoding.

Segment (image processing). Calculate the moments. Calculate: ,,, yx MMA

Page 30: 1 Kostadin N. Koruchev Universidad Autónoma de Madrid, Spain e-mail: k.koroutchev@uam.es

14 Can it be decoded?

Calculate the rotation angle of the grid. Take the central element as a reference. Calculate all angle relative to that angle. Go trough the points forming spiral. Write the closest approximation to 12 deg. step.

84.0423 142.8450 347.2797 153.7838 203.5967 192.0838 …

0, 2,16, 26, 5, 22, 28,18, 22,16, -1, 3, 28

Decode by looking in the tablegenerated while printing.

Necessary number of figures 5. Complexity – except segmentation

proportional to the number of pixels.