reconstruction and pattern recognition via the petitot model
TRANSCRIPT
Reconstruction and pattern recognition via the Petitotmodel
J.P. Gauthier, U. Boscain, Dario Prandi
University of Toulon and Ecole Polytechnique, Paris
[01/15]January 2015
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 1 / 36
Plan
The Petitot Model
The Hypoelliptic diffusion and the semi-discrete diffusion
The lifts
Chu categories and Moore groups
The case of compact groups
The case of SE2,NPattern recognition and texture discrimination
A few results
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 2 / 36
Papers
U. Boscain, ā ,J. Duplaixā”, J.P. Gauthier, F.Rossi, Antropomorphicimage reconstruction via hypoelliptic diffusion, SIAM J. on ControlSICON, 2012.Ā¶U. Boscain, J.P. Gauthier, D. Prandi, A. Remizov, Hypoellipticdiffusion and human vision, a semi-discrete new twist, SIAM J. onImaging science, 2014.
J.P. Gauthier, J. Miteran, F. Smach, Generalized Fourier descriptorswith application to pattern recognition in SVM context, J. onmathematical imaging and vision, 30, 2008.
And the book by J. Petitot:"Vers une neurogĆ©omĆ©trie de la vision", Ed de lāecole Polytechnique, 2006.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 3 / 36
The Petitot Model
In the visual cortex V1, groups of neurons are sensitive to both positionsanddirections.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 4 / 36
Antropomorphic vision-1
the model is:
x = cos(Īø)u, y = sin(Īø)u, Īø = v , J(u, v) =ā« T
0(u(t)2+ v(t)2)dt ā min
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 5 / 36
Antropomorphic vision-2
To this model is associated a (hypoelliptic) diffusion equn:
dĪØdt
= LĪØ,
LĪØ(z , Īø) =12((cos(Īø)
ā
āx+ sin(Īø)
ā
āy)2 +
ā2
āĪø2)ĪØ(x , y , Īø),
That corresponds to go to a stochastic problem, exciting the systemby two independant Brownian motions: dxt = cos(Īø)dut ,dyt = sin(Īø)dut , dĪøt = dvt ,Geodesics can be computed using the PMP, they are given byclassical Jacobi elliptic functions,there are very close relations between SR-distance, geodesics andsmall-time asymptotics of the heat kernel (for instance,limtā0(t log(Pt (x)) = ā 14d(0, x)2),Heat kernel (fundamental solution) can be computed usingnoncommutative harmonic analysis over the group SE (2). It isgiven as a series of Mathieu functions:J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 6 / 36
Antropomorphic vision-3
Pt (g) =
+āā«0
(+ā
ān=0ea
Ī»n t < cen(Īø,
Ī»2
4),ĪŗĪ»(X , Īø)cen(Īø,
Ī»2
4) > + (1)
+ā
ān=0eb
Ī»n t < sen(Īø,
Ī»2
4),ĪŗĪ»(X , Īø)sen(Īø,
Ī»2
4) >)Ī» dĪ».
Due to the small number of pinweels (ā 20), the model is probablyin fact semi-discrete, with stochastic equation:
dzt =(cos(Īøt )sin(Īøt )
)dwt ,
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 7 / 36
Fokker-Planck with jumps-1
where Īø is a jump process and z = (x , y). SetĪN = (Ī»i ,j ), i , j = 0, ...,N ā 1, whereĪ»i ,j = limtā0 1t P [Īøt = ej |Īø0 = ei ), with ej =
2jĻN , and
Ī»i ,j = āāj 6=i Ī»i ,j .
ĪN is the infinitesimal generator of the process Īø.We assume Markov processes, where the law of the first jump time isexponential, with parameter Ī» (that will be specified later on). The jumphas probability 1
2 on each side.
Then we get a Poisson process, and the probability of k jumpsbetween 0 and t is:
P [k jumps)] =(Ī»t)k
k !eāĪ»t .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 8 / 36
Foker-Planck with jumps-2
So that:
P [Īøt = ei+1|Īø0 = ei ] =12[Ī»t +
12
Ī»2t2 + ...]eāĪ»t ,
P [Īøt = ei+2|Īø0 = ei ] =14[12
Ī»2t2 + ...]eāĪ»t ,
with the convention that ei is modulo N.
So that Ī»i ,i+1 = Ī»i ,iā1 =12Ī», and Ī»i ,i = āĪ».
Then, the infinitesimal generator of the semi-group associated with (zt , Īøt )is of the form:
LNĪØ(z , ei ) = (AĪØ)i (z) + (ĪNĪØ(z , ei ))i ,
where ĪØj (z) = ĪØ(z , ej ), and,
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 9 / 36
Foker-Planck with jumps-3
(AĪØ)i (z) = AĪØ(z , ei ) =12(cos(ei )
ā
āx+ sin(ei )
ā
āy)2ĪØ(x , y , ei ),
(ĪNĪØ(z , ei ))i =nā1āj=0
Ī»i ,jĪØj (z) =Ī»
2(ĪØiā1(z)ā 2ĪØi (z) +ĪØi+1(z)).
Then, if we set: Ī» = N 24Ļ2, we get:
(ĪNĪØ(z , ei ))i =12
ĪØiā1(z)ā 2ĪØi (z) +ĪØi+1(z)( 2ĻN )
2,
=12
ā2
āĪø2ĪØ(z , ei ) +O(
1N).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 10 / 36
Foker-Planck with jumps-4
At the limit, we get:
LĪØ(z , Īø) =12((cos(Īø)
ā
āx+ sin(Īø)
ā
āy)2 +
ā2
āĪø2)ĪØ(x , y , Īø),
which is our diffusion equation, while the exact Foker-Planck equationwith small number of angles is:
dpjdt(t, z) =
12(cos(ej )
ā
āx+ sin(ej )
ā
āy)2pj (t, z)+
Ī»
2(pjā1(t, z)ā 2pj (t, z) + pj+1(t, z))
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 11 / 36
Heat kernel via representations-1.
The group law of SE (2,N) is:
(z , ei ) ā (w , ej ) = (z + R iw , ei+j ),where R is the rotation of angle 2Ļ
N .It is a Moore group!!
Unitary irreducible reoresentations are given by the Mackeyās imprimitivitytheorem. They work on Mackeyās orbits that are all Z/NZ.They are parametrized by the orbit of the action of the discrete rotationson the plane, i.e. the dual is the "slice of camembert" SN : (With thetopology of the dual, you have to fold it in order to get the "french friescone" FN ). Let Ī», Ī½ parametrize SN . Then the unitary irreduciblerepresentations are given by:
(ĻĪ»,Ī½(z , er )) = diagk (ei<VĪ»,Ī½,R k z>)S r ,
where S is the shift modN of the components in CN . Also,VĪ»,Ī½ = (Ī» cos(Ī½),Ī» sin(Ī½)). The Plancherel measure is Ī»dĪ»dĪ½.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 12 / 36
Heat kernel via representations-2.
The GFT transforms our hypoelliptic equation into a continuous sum of(**Elliptic**) ones. In the following, MĪ»,Āµ is a N ĆN matrix:
dMĪ»,Ī½
dt= ĪNMĪ»,Ī½ ā diagk [Ī»2 cos(ek ā Ī½)]MĪ»,Ī½
= AĪ»,Ī½MĪ»,Ī½.
This is a matrix Matthieu-type diffusion.And via the inverse GFT, we get:
pt (z , er ) =ā«SNtrace[eAĪ»,Ī½t .diagk (e
i<VĪ»,Ī½,R k z>)S r ]Ī»dĪ»dĪ½.
This is the Jump heat Kernel. A much simpler formula than in the case ofSE (2).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 13 / 36
The algorithm-1.
We could start with the Kernel. What we do now is a bit less economic,but more understandable.1. First, take ordinary Fourier transform with respect to space variable z .Write w for the dual variable to z . Set also w = (Ī» cos(Īø),Ī» sin(Īø)).Diffusion becomes, at w :
dUdt= ĪNU ā diagk [Ī»2 cos(ek ā Īø)2]U.
Here, we write too many Matthieu equation. But this step can beimproved on.2. Integrate w.r.t. t3. Take ordinary inverse Fourier transform.If you do this, you get the exact solution of the discrete diffusion.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 14 / 36
The algorithm-2.
In fact, first, start from the discrete diffusion:dpkdt(t, z) =
12(cos(ek )
ā
āx+ sin(ek )
ā
āy)2pj (t, z)+
Ī»
2(pkā1(t, z)ā 2pki (t, z) + pk+1(t, z)).
1. Take a space discretization, and the ordinary finite differencesapproximation, to get:
dpkdt(t, zi ,j ) =
12(cos(ek )A ā
āx+ sin(ek )A ā
āy)2pk (t, z)+
Ī»
2(pkā1(t, z)ā 2pk (t, z) + pk+1(t, z)), (D)
This is a very big linear differential system: (512Ć 512ĆN). 2. Take thedouble FFT, to get a completely parallel systemd Pi ,j
dt (t) = Di ,j Pi ,j .3. Integrate with respect to time the 512Ć512 matthieu-like equations. 4.Take the inverse double FFT.
TheoremThis algorithm provides the exact solution to the space-discretized system(D).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 15 / 36
The algorithm-3.
I assume that the initial image is infinite, twice periodic.
TheoremThis algorithm provides the exact solution to the space-discretized system(D).
dpkdt(t, zi ,j ) =
12(cos(ek )A ā
āx+ sin(ek )A ā
āy)2pk (t, z)+
Ī»
2(pkā1(t, z)ā 2pk (t, z) + pk+1(t, z)), (D)
Using the Heat Kernel, one could integrate much less Matthieu-like eqs:only one for each point of the slice SN .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 16 / 36
More precisely-1.
x , y : 0āāM, ,x = kā1ā
M, y = lā1ā
M.
FFTM (u)k ;l = 1M āM
r ,s=1 ur ,seā2Ļi [ (kā1)(rā1)+(lā1)(sā1)M ];
and conversely uk ;l = 1M āM
r ,s=1 FFTM (u)r ,se2Ļi [ (kā1)(rā1)+(lā1)(sā1)M ]. Then
the discretized operator āāx is mapped to:
(āuāx)Ėk =
1āM
M
ār=1
(ur+1 ā urā1)2āM
eā2Ļi [ (kā1)(rā1)M ]
=āM
1
2āM[(M
ār=1
ur+1eā2Ļi [ (kā1)(r+1ā1)M ])e2Ļi kā1M ā
(M
ār=1
urā1eā2Ļi [ (kā1)(rā1ā1)M ])eā2Ļi kā1M ]
=
āM2uk (e
2Ļi kā1M ā eā2Ļi kā1M ) = iāMuk sin(
2Ļ(k ā 1)M
).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 17 / 36
More precisely-2.
Then, cos(Īø) āāx + sin(Īø)
āāy āk ,l
iM(sin( 2Ļ(kā1)M ) cos(Īø) + sin(Īø) sin( 2Ļ(lā1)
M )),
and:[ ā2
āĪø2+ Ī±(cos(Īø) ā
āx + sin(Īø)ā
āy )2]uāk ,l
ā2
āĪø2uk ,l ā Ī±M2(sin( 2Ļ(kā1)
M ) cos(Īø) + sin(Īø) sin( 2Ļ(lā1)M ))2uk ,l .
Finally, the diffusion becomes:
2āurk ,l
āt= Ī±
urā1k ,l ā 2urk ,l + ur+1k ,l
( ĻM )
2 ā
Ī²M2(sin(2Ļ(k ā 1)
M) cos(Īø) + sin(Īø) sin(
2Ļ(l ā 1)M
))2urk ,l .
One can see the natural projectivisation (due to the square).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 18 / 36
Relation to kernel and to dual-1.
The (limit) diffusion is, after space-Fourier transform:
dudt=
ā2uāĪø2ā Ļ2 cos(Ļā Īø)2u, or for each fixed Ļ,Ļ :
duĻ,Ļ(Īø, t)dt
=ā2uĻ,Ļ
āĪø2ā Ļ2 cos(Ļā Īø)2uĻ,Ļ.
Set uĻ,Ļ(Ļā Īø, t) = uĻ,Ļ(Īø, t), Ļā Īø = Īø, to obtain:
duĻ,Ļ(Īø, t)dt
=ā2uĻ,Ļ(Īø, t)
āĪø2ā Ļ2 cos(Īø)2uĻ,Ļ(Īø, t).
It means that we need to compute only resolvants along the dual half-line.In the discrete case, we need to compute resolvants at each point of theslice of camembert only.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 19 / 36
Based upon this diffusion and certain heuristic complements, we get nicealgorithms for image completion:
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 20 / 36
An other point of view.
In relation with what follows (for pattern recognition), there is a differentpoint of view . There are finite dimensional subspaces [with arbitrarilylarge dimension] of the space of almost periodic functions over SE (2,N)that are invariant under the diffusion operator: any finite direct sum ofN-dimensional spaces of irreducible representations of SE (2,N).In restriction to these spaces, the diffusion (not the discretization, but theexact diffusion) can be EXACTLY integrated by the previous algorithm.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 21 / 36
The lifts 1
These considerations here work in the general context of a semi-directproduct G = N nH of a compact group N by an abelian locally compactgroup H, and the Haar measure on H is invariant under the action R of N.In that case, all linear left invariant lifts L from L2(H) to L2(G ), such thatf ā Lf (0, e) is densely defined and bounded in L2 norm are of the form:
Lf (n,X ) = [(RnĪØ) ā f ](X ),
where (RnĪØ)(X ) = ĪØ(Rnā1X ), and ĪØ ā L2(H) .Moreover such a map L is injective iffā«
N
|ĪØ(RnĻ)|2dn > 0 for a.e. Ļ ā H.
Example: we can take for ĪØ a standard orientation filter (Gabor forinstance).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 22 / 36
The lifts 2
Define, for f ā L2(H) and Ļ ā H, Ļf (Ļ) ā L2(H) byĻf (Ļ)(n) = f (Rnā1Ļ).Problem: The Fourier transform Lf (TĻ) of a left invariant lift is always arank1 operator:
Lf (TĻ) = Ļf (Ļ)āāĻĪØ(Ļ).
This will be a big problem later on.From this point of view, there are better lifts than the left invariant ones.For instance if H = R2, the "cyclic lift"
f c (n,X ) = f (RnX + Xc ),
Xc =1
f (0)
ā«H
Xf (X )dX
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 23 / 36
Chu Duality 1
Chu duality is an extension of Tannaka duality, working for certainnoncompact groups. Groups (locally compact) that have Chu Duality:Abelian, compact, Moore.Remark: not all MAP groups have Chu duality (Roedersās example).The group M2,N is Moore and then has Chu duality.
Chu dualG a topological group, RPN (G ) denotes the set of all N-dimensionalcontinuous unitary representations R of G in CN , with thecompact-open topology, and RP(G ) is the topological summ of theRPN (G ) over N ā„ 1. RPN (G ) is second countable provided that G isso.RP(G ) is called the Chu dual of G .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 24 / 36
Quasi representations
Quasi representationsA quasi representation of G is a continuous map Q from RP(G ) tothe topological sum U = āŖn1U(n) of all unitary groups, with thefollowing properties:(Q1) Q(R) ā U(n(R)), (Q2) Q(R ā R ā²) = Q(R)āQ(R ā²), (Q3)Q(R ā R ā²) = Q(R)āQ(R ā²), (Q4) Q(Uā1RU) = Uā1Q(R)U for allR,R ā² ā RP(G ) and U ā U(n(R)).Denote by RP(G ) the union of all quasi representations of Gembedded with the compact open topology. RP(G ) is called the Chuquasi dual of G .Set E (R) = Idn(R ), Q
ā1(R) = Q(Rā1). Then, RP(G ) is a Hausdorftopological group, with E as its identity.
for g ā G set g(R) = R(g). consider Ī© : G ā RP(G ), Ī©(g) = g .Ī© is a continuous homomorphism, injective provided that G is MAP.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 25 / 36
Chu duality again
Def: the group G has the Chu duality property if Ī© is a topologicalisomorphism.
TheoremMoore groups have the Chu duality property.
Not all MAP groups have Chu duality property
SE2,N has the Chu duality property while SE2 has not.
Chu duality is a (topological) generalization of Tannaka duality forcompact groups (which is itself an analog of Pontryaginās duality forabelian groups).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 26 / 36
Pattern recognition: the bispectral principle
Let G be a locally compact group. Let G denote as usual the dual ofG , i.e. the set of (equivalence classes of) unitary irreduciblerepresentations of G . For f ā L2(G ), Define If : G Ć G ā C,
If (Ī»1,Ī»2) = f (Ī»1)ā f (Ī»2) f (Ī»1 ā Ī»2)ā, in which f (Ī») is the Fouriertransform of f , and Ī»1 ā Ī»2 is the tensor product representation.By the properties of the Fourier transform, If is invariant under the leftaction of G on L2(G ).
Bispectral principle:
TheoremLet G be separable, abelian or compact, or Moore (with certainrestriction). Then, there is a residual Subset R of L2(G ) such thatIf = Ih implies that h = g0.f for some g0 ā G .
That is, over the very big subset R of L2(G ), functions are separatedmodulo translations by the invariants If .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 27 / 36
Pattern recognition: the bispectral principle
Let G be a locally compact group. Let G denote as usual the dual ofG , i.e. the set of (equivalence classes of) unitary irreduciblerepresentations of G . For f ā L2(G ), Define If : G Ć G ā C,
If (Ī»1,Ī»2) = f (Ī»1)ā f (Ī»2) f (Ī»1 ā Ī»2)ā, in which f (Ī») is the Fouriertransform of f , and Ī»1 ā Ī»2 is the tensor product representation.By the properties of the Fourier transform, If is invariant under the leftaction of G on L2(G ).
Bispectral principle:
TheoremLet G be separable, abelian or compact, or Moore (with certainrestriction). Then, there is a residual Subset R of L2(G ) such thatIf = Ih implies that h = g0.f for some g0 ā G .
That is, over the very big subset R of L2(G ), functions are separatedmodulo translations by the invariants If .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 27 / 36
Proof of the bispectral principle (compact case, sketch)
The generic set R is the subset of L2(G ) such that the Fourier transformf (Ī») is invertible for all Ī» ā G .Assume If = Ih. Apply it to Ī»1 = Ī»,Ī»2 = T , the trivial representation,to get that f (Ī») f (Ī»)ā = h(Ī») h(Ī»)ā. It follows thath(Ī») = f (Ī»)U(Ī»), for some unitary operator U(Ī»).The map U extends uniquely to a map RP(G )ā U by requiring Q2(commutation with ā). Due to the definition of the Fourier transform, themap U meets Q4 (commutation with unitary equivalences). PropertyQ3 (commutation with tensor product) is obtained from the equalityIf = Ih and the definition of the generic set R.It follows that U is a quasi representation, and by Chu (Tannaka) duality,there is g ā G such that U = g . Then, h(Ī») = f (Ī») Ī»(g), and by theelementary property of the Fourier transform, h = g .f .
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 28 / 36
In the case of 1-dimensional signals f (t), the bispectral invariants are just
B(Ī»1,Ī»2) = f (Ī»1)f (Ī»2)f (Ī»1 + Ī»2)
Note that B (Ī»1,0)f (0)
is just the "power spectral density" of the signal, and
B(Ī»1,Ī»2) contains the missing phase informations.
The B(Ī»1,Ī»2) are used in several areas of signal processing. For instance:Dubnov S, Tishby N and Cohen D. (1997). "Polyspectra as Measures ofSound Texture and Timbre". Journal of New Music Research 26: 277ā314.
The If (Ī»1,Ī»2) (the Bispectral Invariants) are the generalization of these,to 2-D signals (more precisely to their lifts). It is expected that theycontain all information modulo motions.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 29 / 36
The case of SE(2,N) 1
Define I Ī»1,Ī»2,k2 (f ) =< Ļf (Ī»1 + k.Ī»2),Ļf (Ī»1)Ļf (k.Ī»2) >L2(N ), where
is the pointwise product.Then, we have the following relation:
[Lf (Ī»1 ā Ī»2) Lf (Ī»1)ā ā Lf (Ī»1)ā].F (u, ukā1) =I Ī»1,Ī»2,k2 (f )ĪØ(uā1(Ī»1 + k.Ī»2)) (ĻĪØ(Ī»1)
ā āĻĪØ(Ī»2)ā.F )
From the bispectral principle, we could expect that the I Ī»1,Ī»2,k2 form a
(weakly) complete set of invariants of the action of SE (2,N) on L2(H),but it doesnāt work since the Fourier transform operators have all rank 1.This is still a conjecture. We have arguments to conjecture that this couldbe false.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 30 / 36
The case of SE(2,N) 2
Define I Ī»1,Ī»2,k2 (f ) = [Lf (Ī»1 ā Ī»2) Lf (Ī»1)ā ā Lf (RākĪ»2)ā].
We say that Ļf (Ī») is cyclic if SnĻf (Ī») is a basis of L2(N). A functionf ā L2(H) is said weakly cyclic if Ļf (Ī») is cyclic for almost all Ī» ā HNote that I Ī»1,Ī»2,k
2 are not invariant under translations (work for centeredfunctions only). We can prove:
Theorem
(T) Assume that ĪØ is weakly cyclic and ĪØ(Ī») 6= 0 for almost all Ī» ā H.Let f , g , be weakly cyclic functions with compact support having the sameinvariants. Then, there is n ā N such that f = Rng .
The sketch of the proof is the same as the general proof for compactgroups: we construct a quasi-representation of G . But a lot of complicateddetails come. A key point is the induction-reduction theorem (an analogof the Clebsch-Gordan decomposition for compact groups).
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 31 / 36
Texture discrimination 1
Let K be a (finite or countable) subset of R2, which is stable under theaction of ZN .It is natural to consider images whose lift are now almost-periodicfunctions f over SE (2,N). Those are called "texture images".
f (n, x) = ān
hāK
a(n, h)e i<Rnh,x>
and that are in the B2 Besikovich class. The theory above can be adaptedto these spaces of functions, and an analog of theorem (T) above can beproved.Moreover,
If K is finite, this space can be used (as was pointed out) to solveexactly the Diffusion equation (first part of the talk). It is a finitedimensional invariant subspace of the action of SE (2,N), andnaturally a sum of spaces of irreducible representations.The (analog of) conjecture of completeness of the invariants I Ī»1,Ī»2,k
2is false (we have a counterexample)
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 32 / 36
A few results 1
Bispectral invariants have good properties w.r.t. scale: letfĪ±(x) := f (Ī±x), then:B(Ī»1,Ī»2)(fĪ±) = 1
Ī±6B(Ī»1
Ī± ,Ī»2Ī± )(f ). We can use this
relation to eliminate the scale effects.At this point, we use a standard strategyBispectral invariants are used together with a learning machine (SVM), torealize "pseudo 3D" pattern recognition.Vapnik, Vladimir N.; The Nature of Statistical Learning Theory,Springer-Verlag, 1995Also, for texture discrimination, wit is very natural to use thealmost-periodic context.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 33 / 36
A few results 2
A long time ago, we got a series of very nice results on standard academicdata bases. We get nice results w.r.t. standard strategies.
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 34 / 36
A few results 3
Face detection on the ORL data base
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 35 / 36
Work on texture discrimation is going on
We thank you for your attention
J.P. Gauthier, U. Boscain, Dario Prandi (University of Toulon and Ecole Polytechnique, Paris)Reconstruction and pattern recognition via the Petitot model[01/15]January 2015 36 / 36