Search

National audience

Determining the optical flow of a video is a compute-intensive task essential for computer vision. For achieving this processing in real time, the whole algorithm deployment chain must be thought of for efficiency first. The development is usually divided into two parts: first, designing an algorithm that meets precision constraints, then, implementing and optimizing its execution on the targeted platform. We argue that unifying those operations enhances performance on the embedded processor. This paper is based on an industrial use case of computer vision. The objective is to determine dense optical flow in real time on an embedded GPU platform: the Nvidia AGX Xavier. The CLG (combined local–global) optical flow method, initially chosen, is analyzed to understand the convergence speed of its underlying optimization problem. The Jacobi solver is selected for implementation because of its parallel nature. The whole multi-level processing is then ported to the GPU, using several specific optimization strategies. In particular, we analyze the impact of fusing the solver’s iterations with the roofline model. As a result, with a 30 W power budget, our implementation runs at 60FPS, on $$640 \times 512$$ images, with a four-level processing. Hopefully, this example should provide feedback on the issues that arise when trying to port a method to a parallel platform and serve for further implementations of computer vision algorithms on specialized hardware.

International audience; In order to enhance image quality for controlling the interior of a volume in industry, model-based iterative reconstruction (MBIR) methods in three-dimensional (3-D) X-ray computed tomography (CT) have shown their efficiency compared to analytical reconstruction methods. MBIR methods enforce a prior model on the volume to reconstruct and make fusion of the information contained in the prior model and the projection data. The projections have many uncertainties that have very different origins in 3-D CT. They are taken into account in MBIR methods but, despite of their different origins, they are in general gathered in only one term in the forward model. In this paper, we propose to derive a new forward model by adding a further error term in the Poisson statistics of photon-count, corresponding to the deviation of the monochromatic model with respect to the actual polychromacy of the rays. A Taylor expansion of the Poisson log-likelihood leads us to a new algebraic forward model accounting for two terms of uncertainties: we call it the error-splitting forward model. Different prior models are assigned to each of distinguished uncertainties: measurement uncertainties are modeled as Gaussian, while linear model uncertainties are modeled as heavy-tailed to bring robustness to the reconstruction process. We give strategies to fix the parameters of the error-splitting forward model. Next, we use it in a full MBIR method with Gauss-Markov-Potts prior model on the volume, in order to reconstruct piecewise-constant objects for non-destructive testing in industry. Compared to the conventional forward model, we show in our experiments that the use of the error-splitting forward model with Gauss-Markov-Potts prior combines better robustness and accuracy.

International audience; This article tackles the entire lifecycle of an algorithm: from its design to its implementation. It exhibits a method for making efficient choices at algorithm design time knowing the characteristics of the underlying hardware target. As of today , computing the optical flow of a stream of images is still a demanding task. In the meantime, the use of Graphics Processing Units (GPU) has become mainstream and allows substantial gains in processing frame rate. In this paper, we focus on a specific variational method (CLG [1]) where linear systems have to be solved. They depend on two parameters α and ρ. To efficiently solve the problem, we look at convergence speed with respect to the model's parameters. We benchmark usual linear solvers with preconditioners to identify the fastest in terms of convergence per iteration. We then show that once implemented on GPUs, the most efficient solver changes depending on the model parameters. For 640 × 480 images, with the right choice of solver and parameters, our implementation can solve the system with relative 10e −7 accuracy in 0.25 ms on a Titan V GPU. All the results are aggregated on a 30-image set to increase confidence in their extendability.

International audience; This article gives a methodological approach to accelerating an environment of a RADAR (RAdio Detecting And Ranging) simulation, from a single-core CPU implementation to a multi-core GPU implementation. We focus our attention on the most common tools for GPU programming like CUDA [2], but more specifically on OpenACC [6], a directive based parallel programming language. One of its promises is, with minimal modifications, to transform a CPU code to take advantage of many-core architectures, CPUs or GPUs alike. Radar systems rely on many layers of testing, one of them being software validation. As technology moves forward, systems become increasingly complex, thus increasing the required processing power to simulate those systems. With CPU performance stalling, it is imperative to switch to alternative architectures. Our contribution is providing key steps for accelerating a software simulation of a radar algorithm on a GPU, with a particular focus on performance but also on the ease of programming. Maximum achieved execution time speedup on GPU architecture for our typical use case of radar processing is of 8.2 for CUDA and of 4.56 for OpenACC compared to the reference implementation on CPU.

International audience; 3D X-ray Computed Tomography (CT) is used in medicine and non-destructive testing 1 (NDT) for industry to visualize the interior of a volume and control its healthiness. Compared to 2 analytical reconstruction methods, model-based iterative reconstruction (MBIR) methods obtain 3 high-quality reconstructions while reducing the dose. Nevertheless, usual Maximum-A-Posteriori 4 (MAP) estimation does not enable to quantify the uncertainties on the reconstruction, which can 5 be useful for the control performed afterwards. Herein, we propose to estimate these uncertainties 6 jointly with the reconstruction by computing Posterior Mean (PM) thanks to Variational Bayesian 7 Approach (VBA). We present our reconstruction algorithm using a Gauss-Markov-Potts prior model 8 on the volume to reconstruct. For PM calculation in VBA, the uncertainties on the reconstruction are 9 given by the variances of the posterior distribution of the volume. To estimate these variances in our 10 algorithm, we need to compute diagonal coefficients of the posterior covariance matrix. Since this 11 matrix is not available in 3D X-ray CT, we propose an efficient solution to tackle this difficulty, based 12 on the use of a matched pair of projector and backprojector. In our simulations using the Separable 13 Footprint (SF) pair, we compare our PM estimation with MAP estimation. Perspectives for this work 14 are applications to real data as improvement of our GPU implementation of SF pair. 15

International audience; Iterative reconstruction methods are used in X-ray Computed Tomography in order to improve the quality of reconstruction compared to filtered backprojection methods. However, these methods are computationally expensive due to repeated projection and backprojection operations. Among the possible pairs of projector and backprojector, the Separable Footprint (SF) pair has the advantage to be matched in order to ensure the convergence of the reconstruction algorithm. Nevertheless, this pair implies more computations compared to unmatched pairs commonly used in order to reduce the computation time. In order to speed up this pair, the projector and the backprojector can be parallelized on GPU. Following one of our previous work, in this paper, we propose a new implementation which takes benefits from the factorized calculations of the SF pair in order to increase the number of data handled by each thread. We also describe the adaptation of this implementation for multi-streaming computations. The method is tested on large volumes of size 1024 3 and 2048 3 voxels.

Extracting the multiphase structure of the neutral interstellar medium (ISM) is key to understand the star formation in galaxies. The radiative condensation of the diffuse warm neutral medium producing a thermally unstable lukewarm medium and a dense cold medium is closely related to the initial step which leads the HI-to-H2 transition and the formation of molecular clouds. Up to now the mapping of these phases out of 21 cm emission hyper-spectral cubes has remained elusive mostly due to the velocity blending of individual cold structures present on a given line of sight. To address this problem, we developed a new Gaussian decomposition algorithm named ROHSA (Regularized Optimization for Hyper-Spectral Analysis) based on a multi-resolution process from coarse to fine grid. ROHSA uses a regularized non-linear least-square criterion to take into account simultaneously the spatial coherence of the emission and the multiphase nature of the gas. In order to obtain a solution with spatially smooth parameters, the optimization is performed on the whole data cube at once. The performances of ROHSA were tested on a synthetic observation computed from numerical simulations of thermally bi-stable turbulence. An application on a 21 cm observation of a high Galactic latitude region from the GHIGLS survey is presented. The evaluation of ROHSA on synthetic 21 cm observations shows that it is able to recover the multiphase nature of the HI and physically meaningful information about the underlying three-dimensional fields (density, velocity and temperature). The application on a real 21 cm observation of a high Galactic latitude field produces a picture of the multiphase HI, with isolated, filamentary and narrow structures and wider, diffuse and space filling components. The test-case field used here contains significant intermediate-velocity clouds that were well mapped out by the algorithm., 20 pages, 21 figures; Accepted for publication in A&A

International audience; The use of IEEE 754-2008 half-precision floating-point numbers is an emerging trend in Graphical Processing Units' architecture. Being such a compact way of representing data, its use may speed up programs by reducing the memory bandwidth usage and allowing hardware designers to fit more computing units within the same die space. In this paper, we highlight the acceleration offered by the use of half floating-point numbers over different implementations of the same operation, a 2D convolution. We show that even though it may lead up to a significant speed-up, the degradation brought by this new format is not always negligible. Then, we choose a deconvolution problem inspired by the SKA radio-telescope processing pipeline to show how half floats behave in a more complex application.

International audience; —Model-based iterative reconstruction methods enable to improve the quality of reconstruction in 3D X-ray Computed Tomography (CT). The main computational burden of these methods lies in successive projection and backprojection operations. Among existing pairs of projector and backprojector, Separable Footprint (SF) pair combines computational efficiency and accurate modelling of X-rays passing through the volume to image. In order to accelerate these operators, implementations on Graphical Processor Units (GPUs) for parallel-computing have been proposed for SF pair. Due to a CPU-loop, these implementations involve many memory transfers between CPU and GPU which are known to be the main bottleneck for GPU computing. In this paper, we investigate a new GPU implementation of SF projector and backprojector in order to minimize these memory transfers. Our proposed GPU SF projector and backprojector have no CPU-loop, and use two ray-driven kernels for the projection and one voxel-driven kernel for the backprojection. After having described their implementations, we study these operators as single modules and validate it in a MBIR method. Perspectives for this work are GPU optimizations and comparisons with the other existing implementations of SF pair.

International audience; This article deals with the evaluation of FPGAs resurgence for hardware acceleration applied to computed tomography on the back-projection operator used in iterative reconstruction algorithms. We focus our attention on the tools developed by FPGAs manufacturers, in particular the Intel FPGA SDK for OpenCL, that promises a new level of hardware abstraction from the developers perspective, allowing a software-like programming of FPGAs. Our first contribution is to propose an accurate memory benchmark, and we follow with an evaluation of different custom OpenCL implementations of the back- projection algorithm. With some clues on memory fetching and coalescing, we then further tune designs to improve performance. Finally, a comparison is made with GPU implementations, and a preliminary conclusion is drawn on FPGAs future for computed tomography.

International audience; The piecewise constant or homogeneous image reconstruction in the context of X-ray Computed Tomography is considered within a Bayesian approach. More precisely, the sparse transformation of such images is modelled with heavy tailed distributions expressed as Normal variance mixtures marginals. The derived iterative algorithms (via Joint Maximum A Posteriori) have identical updating expressions, except for the estimated variances. We show that the behaviour of the each algorithm is different in terms of sensibility to the model selection and reconstruction performances when applied in Computed Tomography.

National audience; Dans cet article, on considère le problème de reconstruction en tomographie 3D à rayons X en utilisant une approche bayésienne avec un modèle a priori hiérarchique (HHBM: Haar based Hierarchical Bayesian Method). La loi de Student généralisée est utilisée pour imposer la structure parcimonieuse des coefficients de la transformation multi-niveaux de Haar sur le volume. Des comparaisons avec l'état de l'art sont présentées, montrant que la méthode proposée donne des résultats de reconstruction plus précis et une convergence plus rapide. Des résultats de simulation sont également fournis pour montrer l'efficacité du modèle proposé pour une reconstruction avec un nombre limité de projections.

International audience; In this paper, a Bayesian method with a hierarchical sparsity enforcing prior model for Dual-Tree Complex Wavelet Transform (DT-CWT) coefficients is proposed. This model is used for X-ray Computed Tomography (CT) image reconstruction. A generalized Student-t distributed prior model is used to enforce the sparse structure of the DT-CWT coefficient of the image. The joint Maximum A Posterior algorithm (JMAP) is used in this Bayesian context. Comparisons with the conventional and other state-of-the-art methods are presented, showing that the proposed method gives more accurate and robust reconstruction results while the dataset is insufficient.

The Bayesian approach is considered for inverse problems with a typical forward model accounting for errors and a priori sparse solutions. Solutions with sparse structure are enforced using heavy-tailed prior distributions. The particular case of such prior expressed via normal variance mixtures with conjugate laws for the mixing distribution is the main interest of this paper. Such a prior is considered in this paper, namely, the Student-t distribution. Iterative algorithms are derived via posterior mean estimation. The mixing distribution parameters appear in updating equations and are also used for the initialization. For the choice of mixing distribution parameters, three model selection strategies are considered: (i) parameters approximating the mixing distribution with Jeffrey law, i.e., keeping the mixing distribution well defined but as close as possible to the Jeffreys priors, (ii) based on the prior distribution form, fixing the parameters corresponding to the form inducing the most sparse solution and (iii) based on the sparsity mechanism, fixing the hyperparameters using the statistical measures of the mixing and prior distribution. For each strategy of model selection, the theoretical advantages and drawbacks are discussed and the corresponding simulations are reported for a 1D direct sparsity application in a biomedical context. We show that the third strategy seems to provide the best parameter selection strategy for this context.

International audience; The 3D X-ray Computed Tomography (CT) is used in many domains. In medical imaging and industrial Non Destructive Testing (NDT) applications, this technique becomes of great interest. In these applications, very often, we need not only to reconstruct the image, but also to detect the contours between the homogeneous regions of the piecewise continuous image. Generally, contours are obtained by a post processing from the reconstructed image. In this paper, we propose a method to estimate image and contour simultaneously. For this we use the Bayesian approach with a prior model in which the relationship between the image and its contour is considered by using a hierarchical Markovian model, and use a sparsity enforcing prior model for the contours. This proposed method can be used for reconstructions when the image is piecewise continuous. The simulation results are compared with some state of the art methods, and they show the efficiency of simultaneously reconstructing and edge detecting by using proposed method.

International audience; Gauss-Markov-Potts models for images and its use in manyimage restoration, super-resolution and Computed Tomography(CT) have shown their effective use for Non DestructiveTesting (NDT) applications. In this paper, we propose a 3DGauss-Markov-Potts model for 3D CT for NDT applications.Thanks to this model, we are able to perform a joint reconstructionand segmentation of the object to control, which isvery useful in industrial applications. First, we describe ourprior models for each unknown of the problem. Then, wepresent results on simulated data and compare them to thoseof Total Variation (TV) minimization algorithm. Two qualityindicators exploiting the segmentation are also proposed.

International audience; Unsupervised iterative reconstruction algorithms based on a Bayesian approach for piecewise constant images are presented in this paper. Such images can be expressed via a sparse representation and the reconstruction problem can be addressed using sparsity enforcing priors. We focus on sparsity enforcing priors expressed as Normal variance mixture, considering three mixing distributions: Inverse Gamma distribution, corresponding to Student-t prior, general inverse Gaussian distribution with the real parameter fixed, corresponding to Normal-inverse Gaussian prior and Gamma distribution corresponding to Variance-Gamma prior. We present and discuss the corresponding iterative algorithms considering the Joint Maximum A Posteriori estimation showing simulations results for 3D X-ray Computed Tomography.

International audience; In this paper, we consider the 3D X-ray CT reconstruction problem by using the Bayesian approach with a hierarchical prior model. A generalized Student-t distributed prior model is used to enforce the sparse structure of the multilevel Haar Transformation of the image. Comparisons with some state of the art methods are presented, showing that the proposed method gives more accurate reconstruction results and a faster convergence. Simulation results are also provided to show the effectiveness of the proposed hierarchical model for a reconstruction with more limited projections.

International audience; Iterative reconstruction methods in Computed Tomography (CT) are known to provide better image quality than analytical methods but they are not still applied in many fields because of their computational cost. In the last years, Graphical Processor Units (GPU) have emerged as powerful devices in order to parallelize calculations, but the efficiency of their use is conditionned on applying algorithms that can be massively parallelizable. Moreover, in non-destructive testing (NDT) applications, a segmentation of the reconstructed volume is often needed in order to have an accurate diagnosis on the material health, but performing a segmentation after the reconstruction introduces uncertainties in the diagnosis from both the reconstruction and the segmentation algorithms. In this paper, we propose an iterative reconstruction method for 3D CT that performs a joint reconstruction and segmentation of the controlled object in NDT for industrial applications. The method is based on a 3D Gauss-Markov-Potts prior model in Bayesian framework, which has shown its effective use in many image restoration and super-resolution problems. First, we briefly describe this model, before deriving the expression of the joint posterior distribution of all the unknowns. Next, an effective maximization of this distribution is presented. We use a ray-driven projector and a voxel-driven backprojector implemented on GPU. The algorithm is developed so it can be massively parallelized. * The authors are grateful to Lionel Gay and Nicolas Cochennec for having provided the real IQI phantom used to test the method. They would also like to thank Thomas Boulay for his contribution to the implementation of the projector and the backprojector on GPU. 2 C. Chapdelaine, A. Mohammad-Djafari, N. Gac, E. Parra / 3D Bayesian CT with Gauss-Markov-Potts prior model Finally, we present our results on simulated and real phantoms. In addition, we investigate further reconstruction quality indicators in order to compare our results with other methods.

International audience; X-ray Computed Tomography (CT) has become a hot topic in both medical and industrial applications in recent decades. Reconstruction by using a limited number of projections is a significant research domain. In this paper, we propose to solve the X-ray CT reconstruction problem by using the Bayesian approach with a hierarchical structured prior model basing on the multilevel Haar transformation. In the proposed model, the multilevel Haar transformation is used as the sparse representation of a piecewise continuous image, and a generalized Student-t distribution is used to enforce its sparsity. The simulation results compare the performance of the proposed method with some state-of-the-art methods.

International audience; Acoustic imaging is a standard technique for mapping acoustic source powers and positions from limited observations on microphone sensors, which often causes an ill-conditioned inverse problem. In this article, we firstly improve the forward model of acoustic power propagation by considering background noises at the sensor array, and the propagation uncertainty caused by wind tunnel effects. We then propose a robust super-resolution approach via sparsity constraint for acoustic imaging in strong background noises. The sparsity parameter is adaptively derived from the sparse distribution of source powers. The proposed approach can jointly reconstruct source powers and positions, as well as the background noise power. Our approach is compared with the conventional beamforming, deconvolution and sparse regularization methods by simulated, wind tunnel data and hybrid data respectively. It is feasible to apply the proposed approach for effectively mapping monopole sources in wind tunnel tests.

In this paper, a hierarchical prior model based on the Haar transformation and an appropriate Bayesian computational method for X-ray CT reconstruction are presented. Given the piece-wise continuous property of the object, a multilevel Haar transformation is used to associate a sparse representation for the object. The sparse structure is enforced via a generalized Student-t distribution ( S t g ), expressed as the marginal of a normal-inverse Gamma distribution. The proposed model and corresponding algorithm are designed to adapt to specific 3D data sizes and to be used in both medical and industrial Non-Destructive Testing (NDT) applications. In the proposed Bayesian method, a hierarchical structured prior model is proposed, and the parameters are iteratively estimated. The initialization of the iterative algorithm uses the parameters of the prior distributions. A novel strategy for the initialization is presented and proven experimentally. We compare the proposed method with two state-of-the-art approaches, showing that our method has better reconstruction performance when fewer projections are considered and when projections are acquired from limited angles.

International audience; In many imaging systems and in particular in X rayComputed Tomography (CT) the reconstruction problem can bewritten as a linear inverse problem. In these problems, one propertywhich can often be exploited is sparsity of the solution in anappropriate basis. In this work we consider the Student-t model inits hierarchical Normal-Inverse Gamma with an appropriate dictionarybased coefficient. Then, thanks to the hierarchical generativemodel of the observation, we derive the expression of the jointposterior law of all the unknowns and an alternate optimisation algorithmfor obtaining the joint MAP solution. We then detail theimplementation issues of this algorithms for parallel computationand show the results on real size 2D and 3D phantoms.

International audience; In recent decades X-ray Computed Tomography (CT) image reconstruction has been largely developed in both medical and industrial domain. In this paper, we propose using the Bayesian inference approach with a new hierarchical prior model. In the proposed model, a generalised Student-t distribution is used to enforce the Haar transformation of images to be sparse. Comparisons with some state of the art methods are presented. It is shown that by using the proposed model, the sparsity of the sparse representation of images is enforced, so that edges of images are preserved. Simulation results are also provided to demonstrate the effectiveness of the new hierarchical model for reconstruction with fewer projections.

International audience; In order to improve the quality of X-ray Computed Tomography (CT) reconstruction for Non Destructive Testing (NDT), we propose a hierarchical prior modeling with a Bayesian approach. In this paper we present a new hierarchical structure for the inverse problem of CT by using a multivariate Student-t prior which enforces sparsity and preserves edges. This model can be adapted to the piecewise continuous image reconstruction problems. We demonstrate the feasibility of this method by comparing with some other state of the art methods. In this paper, we show simulation results in 2D where the image is the middle slice of the Shepp-Logan object but the algorithms are adapted to the big data size problem, which is one of the principal difficulties in the 3D CT reconstruction problem.

Here, we have investigated the consequences of the loss of proof-reading exonuclease function on the ability of the replicative T4 DNA polymerase (gp43) to elongate past a single abasic site located on model DNA substrates. Our results show that wild-type T4 DNA polymerase stopped at the base preceding the lesion on two linear substrates having different sequences, whereas the gp43 D219A exonuclease-deficient mutant was capable of efficient bypass when replicating the same substrates. The structure of the DNA template did not influence the behavior of the exonuclease-proficient or deficient T4 DNA polymerases. In fact, when replicating a damaged "minicircle" DNA substrate constructed by circularizing one of the linear DNA, elongation by wild-type enzyme was still completely blocked by the abasic site, while the D219A mutant was capable of bypass. During DNA replication, the T4 DNA polymerase associates with accessory factors whose combined action increases the polymerase-binding capacity and processivity, and could modulate the behavior of the enzyme towards an abasic site. We thus performed experiments measuring the ability of wild-type and exonuclease-deficient T4 DNA polymerases, in conjunction with these replicative accessory proteins, to perform translesion DNA replication on linear or circular damaged DNA substrates. We found no evidence of either stimulation or inhibition of the bypass activities of the wild-type and exonuclease-deficient forms of T4 DNA polymerase following addition of the accessory factors, indicating that the presence or absence of the proof-reading activity is the major determinant in dictating translesion synthesis of an abasic site by T4 DNA polymerase.

In vitro bypass of damaged DNA by replicative DNA polymerases is usually blocked by helix-distorting or bulky DNA lesions. In this study, we report that substitution of the divalent metal ion Mg2+ with Mn2+ promotes quantitative replication of model DNA substrates containing the major cisplatin or N-2-acetylaminofluorene adducts by the catalytic subunit (UL30) of the replicative DNA polymerase of herpes simplex virus. The ability of Mn2+ ions to confer bypass of bulky lesions was not observed with other replicative DNA polymerases of the B family, such as bacteriophage T4 or delta polymerases. However, for these enzymes, manganese induced the incorporation of one nucleotide opposite the first (3') guanine of the d(GpG) intrastrand cisplatin lesion. Translesion replication of the cisplatin adduct by UL30 led to the incorporation of mismatched bases, with the preferential incorporation of dAMP opposite the 3' guanine of the lesion. Furthermore, substitution of MgCl2 with MnCl2 greatly inhibited the 3' to 5' exonuclease of UL30 but had a far lesser effect on that of T4 DNA polymerase. Finally, manganese induced a conformational change in the structure of UL30 bound to the platinated substrate. Taken together, the latter findings suggest a mechanism by which manganese might allow UL30 to efficiently promote translesion DNA synthesis in vitro.

978-0-7354-1280-4; International audience; In order to improve quality of 3D X-ray tomography reconstruction for Non Destructive Testing (NDT), we investigate in this paper hierarchical Bayesian methods. In NDT, useful prior information on the volume like the limited number of materials or the presence of homogeneous area can be included in the iterative reconstruction algorithms. In hierarchical Bayesian methods, not only the volume is estimated thanks to the prior model of the volume but also the hyper parameters of this prior. This additional complexity in the reconstruction methods when applied to large volumes (from 5123 to 81923 voxels) results in an increasing computational cost. To reduce it, the hierarchical Bayesian methods investigated in this paper lead to an algorithm acceleration by Variational Bayesian Approximation (VBA) [1] and hardware acceleration thanks to projection and back-projection operators paralleled on many core processors like GPU [2].In this paper, we will consider a Student-t prior on the gradient of the image implemented in a hierarchical way [3, 4, 1]. Operators H (forward or projection) and Ht (adjoint or back-projection) implanted in multi-GPU [2] have been used in this study. Different methods will be evalued on synthetic volume "Shepp and Logan" in terms of quality and time of reconstruction. We used several simple regularizations of order 1 and order 2. Other prior models also exists [5]. Sometimes for a discrete image, we can do the segmentation and reconstruction at the same time, then the reconstruction can be done with less projections.