Journal of Applied Mathematics and Computational Mechanics 2014, 13(1), 33-38 THREE-DIMENSIONAL OCTONION WAVELET TRANSFORM Andrzej Katunin Institute of Fundamentals of Machinery Design, Silesian University of Technology Gliwice, Poland [email protected] Abstract. The necessities of processing of spatial data require developing new feature- -sensitive tools such as extensions of the wavelet transform. Considering the advantages of the application of complex wavelets and their extension to quaternionic wavelets for two- dimensional data structures, the new octonion discrete wavelet transform for the analysis of three-dimensional data structures was introduced in this paper. The construction of the wavelet pyramid as well as octonionic wavelets were presented. Keywords: octonion wavelet transform, spatial data processing, octonionic wavelets Introduction The modern demands of spatial data processing require efficient tools for data analysis and feature extraction. One of the intensively developed branches of tools for data processing is the wavelet transform, which allows for a great variation of scaling and wavelet functions, which could be applied during the transform in order to fulfill the requirements of the specific tasks. However, despite the great localization properties in both spatial and frequency domain, the real-valued wavelets lack the shift-invariance and directional selectivi- ty properties. Such properties have the complex-valued wavelets [1], because the independent filterbanks of real and imaginary wavelets constitute the Hilbert trans- form pair [2]. This explains why great attention is paid to the complex wavelets, complex wavelet transforms and their generalizations. Many authors developed the complex extensions of real-valued wavelets [3, 4] and new complex wavelets [5,6]. The natural extension of the complex wavelets and complex wavelet transform is the hypercomplex transform with appropriate wave- lets. The next generalization of a complex wavelet transform is the discrete quater- nion wavelet transform (QWT), which operates on the quaternion (hypercomplex) numbers of 4D vector space. The QWT was introduced in [7] and extended in [8] for the application to 2D image processing. The continuous version of QWT was introduced in [9]. Further, several researchers used QWT for various problems: Soulard and Carré [10] used QWT for texture classification, Liu et al. [2] presented a study of measuring of image sharpness based on quaternion wavelets, while the authors of [11, 12] applied QWT for texture and banknote classification. 34 A. Katunin The goal of the following study is to extend the QWT algorithm applicable to 2D datasets to octonion wavelet transform (OWT), which could be applied for the analysis of 3D datasets and ensures the shift-invariance and directional selectivity properties. The formulation of OWT including the multiresolution pyramid as well as formulation of octonionic wavelets structure were under consideration in this paper. 1. Octonion wavelet transform 1.1. Octonion algebra The octonions (known also as Cayley numbers) are the hypercomplex numbers of 8-dimensional vector space, which could be presented as: O = {o = xo + e1 x 1 + e2 x 2 + e3 x 3 + e4 x 4 + e5 x 5 + e6 x 6 + e7 x 7}, (1) where O is the octonion algebra, xi ∈R are the real coefficients and ei are the 2 orthogonal imaginary numbers, i.e. ei = −1. Considering the fact that octonions are neither commutative e e = −e e ≠ e e ≠ i j j i j i for i, j 0 (2) nor associative e e e = −e e e ≠ e e e ≠ ( i j) k i( j k ) i( j k ) for i,, j k 0 , (3) the multiplicative rules for them differed with respect to the complex algebra C or quaternionic algebra H. The description of the operations on octonions could be found in [13]. The conjugate of o is defined as: {o* = xo − e1 x 1 − e2 x 2 − e3 x 3 − e4 x 4 − e5 x 5 − e6 x 6 − e7 x 7}, (4) while the norm is defined as: o = o*⋅o , (5) which satisfies the following property: o1 o 2 = o1 ⋅ o2 . Extending the polar decomposition of H [12], the octonion o could be decom- posed in the same manner: O = O − }0{ ≅ S7 × R + , where S7 is not a group as for quaternions, S3 = SU )2( , it has a structure called a loop [14]. Similarly as in [12], the octonion could be decomposed as follows: o = o exp(e1α )exp(e2β )exp(e3γ )exp(e4δ )exp(e5ε )exp(e6ζ )exp(e7η) , (6) where o is the modulus of o and (α,,,,,,β γ δ ε ζ η) are the phase angles. Three-dimensional octonion wavelet transform 35 1.2. Octonionic wavelet pyramid Following Mallat’s multiresolution algorithm the 3D data structure f(x,y,z) could be presented in the decomposed form as follows: o f( x,, y z) = An f + n , (7) + o + o + o + o + o + o + o ∑(Di,1 f Di,2 f Di,3 f Di,4 f Di,5 f Di,6 f Di,7 f ) i=1 where o in the superscript was used for indicating octonion 3D analytic signal o o = defined e.g. in [15], An f( x,, y z) and Di, j f(,,) x y z , j 1,...,7, are the approximation and details sets of coefficients of f( x,, y z), respectively. The two latter sets could be then presented in terms of scaling function Φo (x,, y z) and wavelet functions Ψo j (x,, y z) as follows ∞ ∞ ∞ o = Φo Ai f(,,) x y z ∑ ∑ ∑ai,,, k l m i,,, k l m ()x,, y z , (8) k=−∞ l=−∞ m=−∞ ∞ ∞ ∞ o = Ψo Di, j f(,,) x y z ∑ ∑ ∑di,,,, j k l m i,,,, j k l m ()x,, y z , (9) k=−∞ l=−∞ m=−∞ where Φo = −iΦ o −i − −i − −i − i,,, k l m (x,, y z) 2 [2 (x k),2 (y l),2 (z m)], (10) Ψo = −i Ψo −i − −i − −i − i,,,, j k l m (x,, y z) 2 j [2 (x k ),2 (y l),2 (z m)], (11) and = Φo ai,,, k l m f( x,,, y z) i,,, k l m (x,, y z) , (12) = Ψo di,,,, j k l m f( x,,, y z) i,,,, j k l m (x,, y z) , (13) which could be obtained by the iterative determination of impulse responses ho o Φo Ψo and g of the filters (x,, y z) and j (x,, y z) following Mallat’s algorithm. Φo Ψo In the case, when (x,, y z) and j (x,, y z) are separable, they could be decomposed in the following way: o o Φ (x,, y z) = φ(x)φ(y)φ(z), Ψ4 (x,, y z) =ψ (x)ψ (y)φ(z), o o Ψ1 ()()()()x,, y z = φ x φ y ψ z , Ψ5 ()()()()x,, y z =ψ x φ y ψ z , o o (14) Ψ2 ()()()()x,, y z = φ x ψ y φ z , Ψ6 ()()()()x,, y z = φ x ψ y ψ z , o o Ψ3 ()()()()x,, y z =ψ x φ y φ z , Ψ7 ()()()()x,, y z =ψ x ψ y ψ z , 36 A. Katunin where φ (•) and ψ (•) are the 1D complex filters applied along the directions x, y and z. Considering the filter representation of OWT the octonionic scaling and wavelet functions (14) could be presented in the following way: Φo = + + (x,, y z) φh (x)φh (y)φh (z) e1φh (x)φh (y)φg (z) + + + e2φh ()()()()()()()()()x φg y φh z e3φh x φg y φg z e4φg x φh y φh z (15a) + + + e5φg ()()()()()()()()()x φh y φg z e6φg x φg y φh z e7φg x φg y φg z , Ψo = + + 1 (x,, y z) φh (x)φh (y)ψ h (z) e1φh (x)φh (y)ψ g (z) + + + + e2φh ()()()()()()()()()x φg y ψ h z e3φh x φg y ψ g z e4φg x φh y ψ h z (15b) + + + e5φg ()()()()()()()()()x φh y ψ g z e6φg x φg y ψ h z e7φg x φg y ψ g z , Ψo = + + 2 (x,, y z) φh (x)ψ h (y)φh (z) e1φh (x)ψ h (y)φg (z) + + + + e2φh ()()()()()()()()()x ψ g y φh z e3φh x ψ g y φg z e4φg x ψ h y φh z (15c) + + + e5φg ()()()()()()()()()x ψ h y φg z e6φg x ψ g y φh z e7φg x ψ g y φg z , Ψo = + + 3 (x,, y z) ψ h (x)φh (y)φh (z) e1ψ h (x)φh (y)φg (z) + + + + e2ψ h ()()()()()()()()()x φg y φh z e3ψ h x φg y φg z e4ψ g x φh y φh z (15d) + + + e5ψ g ()()()()()()()()()x φh y φg z e6ψ g x φg y φh z e7ψ g x φg y φg z , Ψo = + + 4 (x,, y z) ψ h (x)ψ h (y)φh (z) e1ψ h (x)ψ h (y)φg (z) + + + + e2ψ h ()()()()()()()()()x ψ g y φh z e3ψ h x ψ g y φg z e4ψ g x ψ h y φh z (15e) + + + e5ψ g ()()()()()()()()()x ψ h y φg z e6ψ g x ψ g y φh z e7ψ g x ψ g y φg z , Ψo = + + 5 (x,, y z) ψ h (x)φh (y)ψ h (z) e1ψ h (x)φh (y)ψ g (z) + + + + e2ψ h ()()()()()()()()()x φg y ψ h z e3ψ h x φg y ψ g z e4ψ g x φh y ψ h z (15f) + + + e5ψ g ()()()()()()()()()x φh y ψ g z e6ψ g x φg y ψ h z e7ψ g x φg y ψ g z , Ψo = + + 6 (x,, y z) φh (x)ψ h (y)ψ h (z) e1φh (x)ψ h (y)ψ g (z) + + + + e2φh ()()()()()()()()()x ψ g y ψ h z e3φh x ψ g y ψ g z e4φg x ψ h y ψ h z (15g) + + + e5φg ()()()()()()()()()x ψ h y ψ g z e6φg x ψ g y ψ h z e7φg x ψ g y ψ g z , Ψo = + + 7 (x,, y z) ψ h (x)ψ h (y)ψ h (z) e1ψ h (x)ψ h (y)ψ g (z) + + + + e2ψ h ()()()()()()()()()x ψ g y ψ h z e3ψ h x ψ g y ψ g z e4ψ g x ψ h y ψ h z (15h) + + + e5ψ g ()()()()()()()()()x ψ h y ψ g z e6ψ g x ψ g y ψ h z e7ψ g x ψ g y ψ g z .