Michel Desvignes

High speed 3D tomography on CPU, GPU, and FPGA

Back-projection (BP) is a costly computational step in tomography image reconstruction such as positron emission tomography (PET). To reduce the computation time, this paper presents a pipelined, prefetch, and parallelized architecture for PET BP (3PA-PET). The key feature of this architecture is its original memory access strategy, masking the high latency of the external memory. Indeed, the pattern of the memory references to the data acquired hinders the processing unit. The memory access bottleneck is overcome by an efficient use of the intrinsic temporal and spatial locality of the BP algorithm. A loop reordering allows an efficient use of general purpose processors caches, for software implementation, as well as the 3D predictive and adaptive cache (3D-AP cache), when considering hardware implementations. Parallel hardware pipelines are also efficient thanks to a hierarchical 3D-AP cache: each pipeline performs a memory reference in about one clock cycle to reach a computational throughput close to 100%. The 3PA-PET architecture is prototyped on a system on programmable chip (SoPC) to validate the system and to measure its expected performances. Time performances are compared with a desktop PC, a workstation, and a graphic processor unit (GPU).

Motion blur parameters identification from Radon transform image gradients

Motion blur is one of the most common blurs that degrades the images. Restoration of a motion blur image is highly dependent on the estimation of the parameters of the blurring kernel. The features of motion blur kernel are sharp dependent on the noise. This paper proposes a novel approach to estimate the parameters of motion blur (orientation and extension) from the observed image gradients. The image gradients enhance the periodic patterns of the motion blur kernel in the frequency space. And the proposed normalized Radon transform from the blurred image gradients could estimate the motion blur parameters in noisy image gradients. Compared to previous estimation algorithm, the results are more accurate when it comes to noisy images.

3D Meshes Registration: Application to Statistical Skull Model

In the context of computer assist surgical techniques, a new elastic registration method of 3D meshes is presented. In our applications, one mesh is a high density mesh (30000 vertexes), the second is a low density one (1000 vertexes). Registration is based upon the minimisation of a symmetric distance between both meshes, defined on the vertexes, in a multi resolution approach. Results on synthetic images are first presented. Then, thanks to this registration method, a statistical model of the skull is build from Computer Tomography exams collected for twelve patients.

Hemoglobin and melanin quantification on skin images

In this paper, we propose and compare four different approaches for quantification of hemoglobin and melanin on skin color image. The first method is to extract erythema/melanin indices based on skin absorbance theories. The second method is based on independent component analysis (ICA) assuming that hemoglobin and melanin absorbance spectra are independent. The third method is proposed based on non-negative matrix factorization (NMF) with multiplicative update algorithm. Finally, we propose a model-fitting technique based on Beer-Lambert law. Quantitative evaluation through graph-cut segmentation on 30 melanoma lesions from 10 patients indicates that model-fitting method outperforms the other three methods.

Quantification of pigmentation in human skin images

In this paper, we propose and compare four different approaches for quantification of hemoglobin and melanin in skin color images. The first method is to extract erythema/melanin indices based on skin absorbance theories. The second method is based on independent component analysis (ICA) assuming that hemoglobin and melanin absorbance spectra are independent. The third method is based on non-negative matrix factorization (NMF) with multiplicative update algorithm. The fourth method is a Beer-Lambert law based model-fitting technique. Quantitative evaluation through graph-cut segmentation on melanoma indicates that model-fitting method outperforms the other three methods.

Motion blur adaptive identification from natural image model

This paper proposes a novel approach to estimate the parameters of motion blur (orientation and extension) simultaneously from the observed image. The motion blur estimation would be used in a standard non blind deconvolution algorithm, thus yielding a blind motion deblurring scheme. Our algorithm is based on the correlation between the modified logarithm power spectrum from natural image model and the blur kernel. The local minima of the modified spectrum are closer to the horizontal line, and thus more similar to the sinc function. Compared to previous estimation algorithm, the results are more accurate in noisy images.

SUPERVISED AND UNSUPERVISED STATISTICAL MODELS FOR CEPHALOMETRY

In this paper, we present an algorithm for aligning and matching unlabeled sets of points and a method for building a statistical model composed of a mean ob- servation and associated variability. This model is used to solve the cephalomet- ric problem. The main idea of this paper consists in using a dual step strategy: estimate the pose and the correspondence alternatively. Correspondence being computed automatically, this process is applied to supervised and unsupervised model learning on unordered sets of points.