Wavelets are often used to analyse piece-wise smooth signals. Wavelet coefficients can efficiently represent a signal which has led to data compression algorithms using wavelets. Wavelet analysis is extended for multidimensional signal processing as well. This article introduces a few methods for wavelet synthesis and analysis for multidimensional signals. There also occur challenges such as directivity in multidimensional case.
Multidimensional separable discrete wavelet transform (DWT) The discrete wavelet transform is extended to the multidimensional case using the tensor product of well known 1-D wavelets. In 2-D for example, the tensor product space for 2-D is decomposed into four tensor product vector spaces as (φ(x) ⨁ ψ(x)) ⊗ (φ(y) ⨁ ψ(y)) = { φ(x)φ(y), φ(x)ψ(y), ψ(x)φ(y), ψ(x)ψ(y) } This leads to the concept of multidimensional separable DWT similar in principle to the multidimensional DFT. φ(x)φ(y) gives the approximation coefficients and other subbands: φ(x)ψ(y) low-high (LH) subband, ψ(x)φ(y) high-low (HL) subband, ψ(x)ψ(y) high-high (HH) subband, give detail coefficients.
Implementation of multidimensional separable DWT Wavelet coefficients can be computed by passing the signal to be decomposed though a series of filters. In the case of 1-D, there are two filters at every level-one low pass for approximation and one high pass for the details. In the multidimensional case, the number of filters at each level depends on the number of tensor product vector spaces. For M-D, 2M filters are necessary at every level. Each of these is called a subband. The subband with all low pass (LLL...) gives the approximation coefficients and all the rest give the detail coefficients at that level. For example, for M=3 and a signal of size N1 × N2 × N3 , a separable DWT can be implemented as follows:
Applying the 1-D DWT analysis filterbank in dimension N1, it is now split into two chunks of size N1⁄2 × N2 × N3. Applying 1-D DWT in N2 dimension, each of these chunks is split into two more chunks of N1⁄2 × N2⁄2 × N3. This repeated in 3-D gives a total of 8 chunks of size N1⁄2 × N2⁄2 × N3⁄2.
Disadvantages of M-D separable DWT The wavelets generated by the separable DWT procedure are highly shift variant. A small shift in the input signal changes the wavelet coefficients to a large extent. Also, these wavelets are almost equal in their magnitude in all directions and thus do not reflect the orientation or directivity that could be present in the multidimensional signal. For example, there could be an edge discontinuity in an image or an object moving smoothly along a straight line in the space-time 4D dimension. A separable DWT does not fully capture the same. In order to overcome these difficulties, a method of wavelet transform called Complex wavelet transform (CWT) was developed.
Multidimensional complex wavelet transform Similar to the 1-D complex wavelet transform, tensor products of complex wavelets are considered to produce complex wavelets for multidimensional signal analysis. With further analysis it is seen that these complex wavelets are oriented. This sort of orientation helps to resolve the directional ambiguity of the signal.
Implementation of multidimensional (M-D) dual tree CWT Dual tree CWT in 1-D uses 2 real DWTs, where the first one gives the real part of CWT and the second DWT gives the imaginary part of the CWT. M-D dual tree CWT is analyzed in terms of tensor products. However, it is possible to implement M-D CWTs efficiently using separable M-D DWTs and considering sum and difference of subbands obtained. Additionally, these wavelets tend to be oriented in specific directions. Two types of oriented M-D CWTs can be implemented. Considering only the real part of the tensor product of wavelets, real coefficients are obtained. All wavelets are oriented in different directions. This is 2m times as expansive where m is the dimensions. If both real and imaginary parts of the tensor products of complex wavelets are considered, complex oriented dual tree CWT which is 2 times more expansive than real oriented dual tree CWT is obtained. So there are two wavelets oriented in each of the directions. Although implementing complex oriented dual tree structure takes more resources, it is used in order to ensure an approximate shift invariance property that a complex analytical wavelet can provide in 1-D. In the 1-D case, it is required that the real part of the wavelet and the imaginary part are Hilbert transform pairs for the wavelet to be analytical and to exhibit shift invariance. Similarly in the M-D case, the real and imaginary parts of tensor products are made to be approximate Hilbert transform pairs in order to be analytic and shift invariant. Consider an example for 2-D dual tree real oriented CWT: Let ψ(x) and ψ(y) be complex wavelets: ψ(x) = ψ(x)h + j ψ(x)g and ψ(y) = ψ(y)h + j ψ(y)g. ψ(x,y) = [ψ(x)h + j ψ(x)g][ ψ(y)h + j ψ(y)g] = ψ(x)hψ(y)h - ψ(x)gψ(x)g + j [ψ(x)hψ(y)g - ψ(x)hψ(x)g] The support of the Fourier spectrum of the wavelet above resides in the first quadrant. When just the real part is considered, Real(ψ(x,y)) = ψ(x)hψ(y)h - ψ(x)gψ(x)g has support on opposite quadrants (see (a) in figure). Both ψ(x)hψ(y)h and ψ(x)gψ(y)g correspond to the HH subband of two different separable 2-D DWTs. This wavelet is oriented at -45o. Similarly, by considering ψ2(x,y) = ψ(x)ψ(y)*, a wavelet oriented at 45o is obtained. To obtain 4 more oriented real wavelets, φ(x)ψ(y), ψ(x)φ(y), φ(x)ψ(y)* and ψ(x)φ(y)* are considered. The implementation of complex oriented dual tree structure is done as follows: Two separable 2-D DWTs are implemented in parallel using the filterbank structure as in the previous section. Then, the appropriate sum and difference of different subbands (LL, LH, HL, HH) give oriented wavelets, a total of 6 in all.
Similarly, in 3-D, 4 separable 3-D DWTs in parallel are needed and a total of 28 oriented wavelets are obtained.
… excerpt ends here. Continue reading the full article.
