Research Progress Report * on Directional Image Representations

Summary of 1st year of research from November 2003 to October 2004,

carried out by Mr. Pancham Shukla under the supervision of  Dr P L Dragotti

Communications and Signal Processing Research Group

Department of Electrical and Electronic Engineering

Imperial College London

 

1.            Introduction

 

According to Shannon’s sampling theory, it is possible to sample and reconstruct the original continuous signal from a finite number of samples (from its low resolution subspace) using a Sinc sampling kernel, provided that the signal is bandlimited. In many situations, this ‘bandlimited-sinc’ constraint is too restrictive to abide by the available acquisition devices and processing algorithms [1].

In recent past, Unser et al. extended the conventional ‘bandlimited-sinc’ scenario to classes of signals such as uniform Splines, which are not bandlimited but reside in a subspace spanned by a generating function and its shifted versions [1,2]. Very recently, Vetterli et al. showed that it is possible to develop sampling schemes for a larger class of signals that are neither bandlimited nor reside in a vector subspace but subsist in a nonlinear space of finite dimension. Such signals are classified as signals with Finite Rate of Innovation (FRI). Streams of Diracs and piecewise polynomial signals are the examples of such signals. Vetterli et al. addressed the sampling of FRI signals using Sinc and Gaussian kernels in Fourier domain [3,4].

For many natural images, often, the region of interest may be reasonably modelled with the FRI structures, and these structures may follow random distributions and orientations. Following the above hypothesis, our objective is to investigate how efficiently we can sample and reconstruct the set of FRI signals from only a few samples employing a more general set of sampling kernels that can closely match the characteristics of acquisition devices. For 1-D signals, Dragotti et al. showed that wavelets can be an effective choice as sampling kernel [5]. This may be attributed to its improved time-frequency localization and compact support. Initial results on wavelet sampling of 1-D FRI signals are promising but the problem becomes more involved for the multidimensional case as usual. In the first phase of our research, we are addressing wavelet sampling and reconstruction of 2-D FRI structures, in particular, streams of Diracs and piecewise constants (planar polygons).

 

2.            Materials and Methods

 

Sampling setup

We use a 2-D generic sampling setup as shown in Figure 1, where the 2-D FRI signal to be sampled is filtered with smoothing kernel and a uniform set of samples  is available from the filtered version with the inner product:

                

where  are sampling intervals in  and directions respectively over a uniform rectangular lattice.

 

 

 

Figure 1: 2-D generic sampling setup.

 

FRI signals

The signal  is composed of some simple FRI structures such as streams of 2-D Diracs and planar polygons (piecewise constants). For simplicity, the planar polygons can be classified in two classes; parallel boundary and skewed boundary depending on the orientation (angle) complexity of the boundaries. The selected FRI structures are depicted in Figure 2.

                                                                                                                                   

   

 

Figure 2: (a) left: streams of Diracs, (b) centre: parallel boundary polygons, (c) right: skewed boundary polygons.

 

Sampling kernel

The selected sampling kernel is a compactly supported 2-D separable scaling function  obtained from any valid 1-D scaling function . For example, 2-D Linear Spline and Daubechies sampling kernels are shown in Figure 3.

            

Figure 3: (a) left: liner Spline kernel (b) right: 4-tap Daubechies kernel.

 

Important properties of sampling kernel

At the heart of the research reclines the partition of unity and polynomial approximation properties of the sampling kernel  as highlighted in Figure 4. These crucial properties of the kernel play a major role in sampling and reconstruction of the selected classes of FRI structures.

 

  

 

Figure 4: (a) left: partition of unity, responsible for reconstruction of the amplitude of the Dirac, (b) centre: reproduction of polynomial of degree 1 in direction, responsible for locating the-coordinate of the Dirac (c) right: reproduction of polynomial of degree 1 in direction, responsible for locating the-coordinate of the Dirac.

 

Related tools and techniques

We also adopt the following techniques in order to generalize our algorithms in reconstructing the FRI structures from their finite number of samples.

1.      Annihilating filter methods [3]: We use them for the cases of densely packed Diracs and for locating (and reconstructing) the complex-valued vertices of planar polygons (in particular, skewed boundary ones) from their complex-valued moments [7,9].

2.      Radon transform projections [6,7]: A close connection between the moment property (line integrals) of radon projections and partition of unity of wavelet sampling kernel is exploited in determining the orientations of the boundaries of planar polygons.

3.      Directional derivatives [8]: This procedure facilitates to decompose the problem of sampling a range of planar polygons into a generalized formulation of sampling the streams of Diracs with any valid sampling kernel. It leads to a local and less complex solution for reconstruction of planar polygons with the significant flexibility in the orientation of their boundaries.

4.      Complex moments [7,9]: This approach provides a global solution for the reconstruction of planar polygons (in particular, skewed boundary ones) form their complex-valued moments. However, it poses concerns on the complexity and stability in reconstruction for the polygons with many vertices.

 

3.            Results and future work

We have extended several ideas of continuous domain to our discrete setup, i.e. a set of a finite number of samples. The properties shown in Figure 4, and the directional derivatives along the sampling axis allowed us to sample and reconstruct simple cases of FRI structures such as a single Dirac per area and parallel boundary polygons. The promising results encouraged us to generalize our model for more complex cases, i.e. more number of Diracs per area and skewed boundary polygons.

         By discovering the relevant connections of the related tools and techniques (as mentioned above) to our wavelet sampling, we are successful in sampling and reconstructing the streams of 2-D Diracs and planar polygons with different structural complexities (more than one Diracs per area, and a large range of permissible orientations for the boundaries of planar polygons). For both types of FRI structures, we have derived a couple of novel algorithms that can be employed locally or globally depending on the structural complexity of a given FRI signal. Finally, we should accentuate that our implementation allows a wide choice of wavelets as sampling kernel and requires only a finite number of samples for reconstruction algorithms. This has a direct practical relevance in the design of acquisition devices and processing algorithms for the state-of-the-art digital cameras.

            We are currently working on extending the notion of wavelet footprints (a multiresolution framework for resolution enhancement [10]) in 2-D by incorporating our sampling results on FRI structures. Efficient ways of presenting 2-D directional structures with footprints should improve the performance of many commercial applications. We anticipate that the future work might offer some novel schemes to the current state of image-resolution enhancement (or super-resolution imaging).

The image-resolution enhancement has many applications such as, the corner point detection in blurred satellite images, position detection in robotics and machine vision, and diagnostics in bio-medical imaging. To judge the current and future potential of this subject, we refer to the issue of IEEE Signal Processing Magazine [11].

 

References

 

[1]. M Unser. Sampling- 50 Years after Shannon. Proceedings of the IEEE, 88(4):569-587, April 2000.

[2]. A  Aldroubi and K Grochenig. Non-uniform sampling in shift-invariant spaces. SIAM Review, 43:585-620, 2001.

[3]. Martin Vetterli, Pina Marziliano, and Thierry Blu. Sampling signals with finite rate of innovation. IEEE Transactions on Signal Processing, 50(6):1417-1428. June 2002.

[4] Irena Maravic and Marin Vetterli. Exact sampling results for some classes of parametric nonbandlimited 2-D signals. IEEE Transactions on Signal Processing, 52(1):175-189, January 2004.

[5]. P L Dragotti and M Vetterli. Wavelet and footprint sampling of signals with a finite rate of innovation. In Proceedings of IEEE International conference on Acoustics, Speech and Signal Processing (ICASSP), Montreal, Canada, May 2004.

[6]. I Maravic and M Vetterli. A sampling theorem for the Radon transform of finite complexity objects. In Proceedings of  IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Orlando, Florida, May 2002.

[7]. P Milanfar, G Verghese, W Karl , and A Willsky. Reconstructing polygons from moments with connections to array processing. IEEE Transactions on Signal Processing, 43:432-443, February 1995.

[8]. V Velisavljevic, B Beferull-Lozano, M Vetterli and P L Dragotti. Discrete multi-directional wavelet bases, In Proceedings of IEEE International Conference on Image Processing (ICIP), Barcelona, Spain, September 2003.

[9]. M Elad, P Milanfar, and G H Golub. Shape from moments- an estimation theory perspective. IEEE Transactions on Signal Processing, 52(7):1814-1829, July 2004.

[10]. P L Dragotti and M Vetterli. Wavelet footprints: theory, algorithms and applications, IEEE Transactions on Signal Processing, 51(5):1306-1323, May 2003.

[11]. -----, IEEE Signal Processing Magazine, 20(3), May 2003.

* The time scale of the first year research is illustrated through the attached Gant chart.