Filtered Back-Projection (FBP) is a fundamental compute intense algorithm used in tomographic image reconstruction. Cone-Beam Computed Tomography (CBCT) devices use a cone-shaped X-ray beam, in comparison to the parallel beam used in older CT generations. Distributed image reconstruction of cone-beam datasets typically relies on dividing batches of images into different nodes. This simple input decomposition, however, introduces limits on input/output sizes and scalability.

Traditional tomography uses circular trajectories and here, filtered backprojection often works well. However, for objects with large aspect ratios, rotational tomography is often not feasible. In these cases, other trajectories can be more appropriate. For generic trajectories, filtered backprojection methods might not work well and full iterative reconstruction can be computationally demanding. In this study, the authors thus propose a third paradigm that combines aspects of both of these techniques. They use interpolation and backprojection techniques to generate an initial estimate of an object's internal structure using projection images taken at different orientations. Depending on the scanning geometry used to calculate the tomographic projections, this initial estimate can be understood as a blurred (filtered) approximation of the actual structure. For each scanning geometry, they specify the equivalent blurring operator that would provide the same estimate directly from a representation of the object's internal structure. They then use iterative techniques to invert this filtering operation, thus estimating the internal structure from the estimate of its blurred representation.

This paper deals with a clustering problem where feature vectors are clustered depending on the angle between feature vectors, that is, feature vectors are grouped together if they point roughly in the same direction. This directional distance measure arises in several applications, including document classification and human brain imaging. Using ideas from the field of constrained low-rank matrix factorization and sparse approximation, a novel approach is presented that differs from classical clustering methods, such as seminonnegative matrix factorization, K -EVD, or k -means clustering, yet combines some aspects of all these. As in nonnegative matrix factorization and K -EVD, the matrix decomposition is iteratively refined to optimize a data fidelity term; however, no positivity constraint is enforced directly nor do we need to explicitly compute eigenvectors. As in k -means and K -EVD, each optimization step is followed by a hard cluster assignment. This leads to an efficient algorithm that is shown here to outperform common competitors in terms of clustering performance and/or computation speed. In addition to a detailed theoretical analysis of some of the algorithm's main properties, the approach is empirically evaluated on a range of toy problems, several standard text clustering data sets, and a high-dimensional problem in brain imaging, where functional magnetic resonance imaging data are used to partition the human cerebral cortex into distinct functional regions.

In this paper we study Monte Carlo type approaches to Bayesian sparse inference under a squared error loss. This problem arises in Compressed Sensing, where sparse signals are to be estimated and where recovery performance is measured in terms of the expected sum of squared error. In this setting, it is common knowledge that the mean over the posterior is the optimal estimator. The problem is however that the posterior distribution has to be estimated, which is extremely difficult. We here contrast approaches that use a Monte Carlo estimate for the posterior mean. The randomised Iterative Hard Thresholding algorithm is compared to a new approach that is inspired by sequential importance sampling and uses a bootstrap re-sampling step based on importance weights.

Nonconvex constraints are valuable regularizers in many optimization problems. In particular, sparsity constraints have had a significant impact on sampling theory, where they are used in compressed sensing and allow structured signals to be sampled far below the rate traditionally prescribed. Nearly, all of the theory developed for compressed sensing signal recovery assumes that samples are taken using linear measurements. In this paper, we instead address the compressed sensing recovery problem in a setting where the observations are nonlinear. We show that, under conditions similar to those required in the linear setting, the iterative hard thresholding algorithm can be used to accurately recover sparse or structured signals from few nonlinear observations. Similar ideas can also be developed in a more general nonlinear optimization framework. In the second part of this paper, we therefore present related result that shows how this can be done under sparsity and union of subspaces constraints, whenever a generalization of the restricted isometry property traditionally imposed on the compressed sensing system holds.

In this paper, we study the problem of sampling and reconstructing signals which are assumed to lie on or close to one of several subspaces of a Hilbert space. Importantly, we here consider a very general setting in which we allow infinitely many subspaces in infinite dimensional Hilbert spaces. This general approach allows us to unify many results derived recently in areas such as compressed sensing, affine rank minimization, analog compressed sensing and structured matrix decompositions.

Sparse representations have proved a powerful tool in the analysis and processing of audio signals and already lie at the heart of popular coding standards such as MP3 and Dolby AAC. In this paper we give an overview of a number of current and emerging applications of sparse representations in areas from audio coding, audio enhancement and music transcription to blind source separation solutions that can solve the “cocktail party problem.” In each case we will show how the prior assumption that the audio signals are approximately sparse in some time-frequency representation allows us to address the associated signal processing task.

When sampling signals below the Nyquist rate, efficient and accurate reconstruction is nevertheless possible, whenever the sampling system is well behaved and the signal is well approximated by a sparse vector. This statement has been formalised in the recently developed theory of compressed sensing, which developed conditions on the sampling system and proved the performance of several efficient algorithms for signal reconstruction under these conditions. In this paper, we prove that a very simple and efficient algorithm, known as Iterative Hard Thresholding, has near optimal performance guarantees rivalling those derived for other state of the art approaches.

Sparse modeling of signals has recently received a lot of attention. Often, a linear under-determined generative model for the signals of interest is proposed and a sparsity constraint imposed on the representation. When the generative model is not given, choosing an appropriate generative model is important, so that the given class of signals has approximate sparse representations. In this paper we introduce a new scheme for dictionary learning and impose an additional constraint to reduce the dictionary size. Small dictionaries are desired for coding applications and more likely to ldquoworkrdquo with suboptimal algorithms such as Basis Pursuit. Another benefit of small dictionaries is their faster implementation, e.g. a reduced number of multiplication/addition in each matrix vector multiplication, which is the bottleneck in sparse approximation algorithms.

Compressed sensing is an emerging signal acquisition technique that enables signals to be sampled well below the Nyquist rate, given that the signal has a sparse representation in an orthonormal basis. In fact, sparsity in an orthonormal basis is only one possible signal model that allows for sampling strategies below the Nyquist rate. In this paper, we consider a more general signal model and assume signals that live on or close to the union of linear subspaces of low dimension. We present sampling theorems for this model that are in the same spirit as the Nyquist-Shannon sampling theorem in that they connect the number of required samples to certain model parameters. Contrary to the Nyquist-Shannon sampling theorem, which gives a necessary and sufficient condition for the number of required samples as well as a simple linear algorithm for signal reconstruction, the model studied here is more complex. We therefore concentrate on two aspects of the signal model, the existence ofonetoonemaps to lower dimensional observation spaces and the smoothness of the inverse map. We show thatalmostalllinear maps areonetoonewhen the observation space is at least of the same dimension as the largest dimension of the convex hull of the union ofanytwosubspaces in the model. However, we also show that in order for the inverse map to have certain smoothness properties such as a given finite Lipschitz constant, the required observation dimension necessarily depends logarithmically on the number of subspaces in the signal model. In other words, while unique linear sampling schemes require a small number of samples depending only on the dimension of the subspaces involved, in order to have stable sampling methods, the number of samples depends necessarily logarithmically on the number of subspaces in the model. These results are then applied to two examples, the standard compressed sensing signal model in which the signal has a sparse representation in an orthonormal basis and to a sparse signal model with additional tree structure.

Sparse signal approximations are approximations that use only a small number of elementary waveforms to describe a signal. In this paper we proof the convergence of an iterative hard thresholding algorithm and show, that the fixed points of that algorithm are local minima of the sparse approximation cost function, which measures both, the reconstruction error and the number of elements in the representation. Simulation results suggest that the algorithm is comparable in performance to a commonly used alternative method.

Sparse coding is a new field in signal processing with possible applications to source coding. In this paper we present a new method that combines the problems of sparse signal approximation with coefficient quantization. This method uses overcomplete dictionaries and exploits signal redundancy. The proposed method will be derived as an extension of a recently presented method (iterative thresholding) to find sparse representations of signals. Because in digital communication and storage we need a quantized representation of the signal, instead of quantization of sparse representations a posteriori, we propose a refined method that combines sparse approximation and quantization. To compare the proposed method to a posteriori quantization, we present an audio example.

Redundancy reduction has been proposed as the main computational process in the primary sensory pathways in the mammalian brain. This idea has led to the development of sparse coding techniques, which are exploited in this article to extract salient structure from musical signals. In particular, we use a sparse coding formulation within a generative model that explicitly enforces shift-invariance. Previous work has applied these methods to relatively small problem sizes. In this paper, we present a subset selection step to reduce the computational complexity of these methods, which then enables us to use the sparse coding approach for many real world applications. We demonstrate the algorithm's potential on two tasks in music analysis: the extraction of individual notes from polyphonic piano music and single-channel blind source separation.

We use Bayesian statistics to study the dictionary learning problem in which an over-complete generative signal model has to be adapted for optimally sparse signal representations. With such a formulation we develop a stochastic gradient learning algorithm based on importance sampling techniques to minimise the negative marginal log-likelihood. As this likelihood is not available analytically, approximations have to be utilised. The importance sampling Monte Carlo marginalisation proposed here improves on previous methods and addresses three main issues: (1) bias of the gradient estimate; (2) multi-modality of the distribution to be approximated; and (3) computational efficiency. Experimental results show the advantages of the new method when compared to previous techniques. The gained efficiency allows the treatment of large scale problems in a statistically sound framework as demonstrated here by the extraction of individual piano notes from a polyphonic piano recording.

Many time-series in engineering arise from a sparse mixture of individual components. Sparse coding can be used to decompose such signals into a set of functions. Most sparse coding algorithms divide the signal into blocks. The functions learned from these blocks are, however, not independent of the temporal alignment of the blocks. We present a fast algorithm for sparse coding that does not depend on the block location. To reduce the dimensionality of the problem, a subspace selection step is used during signal decomposition. Due to this reduction, an iterative reweighted least squares method can be used for the constrained optimisation. We demonstrate the algorithm's abilities by learning functions from a polyphonic piano recording. The found functions represent individual notes and a sparse signal decomposition leads to a transcription of the piano signal.