IPOLIPOLhttp://www.ipol.im/feed/IPOL Articles — Latest articles published in IPOL.ikiwiki2019-09-17T08:34:48ZAn Implementation of the Mean Shift Algorithmhttp://www.ipol.im/pub/art/2019/255/Damir Demirović2019-09-17T08:34:48Z2019-09-16T22:00:00Z
In this paper, we present an implementation and analysis of the mean shift algorithm.
The mean shift is a general non-parametric mode finding/clustering procedure widely used in image processing and analysis and computer vision techniques such as image denoising, image segmentation, motion tracking, etc.
A Study of Two CNN Demosaicking Algorithmshttp://www.ipol.im/pub/art/2019/274/Thibaud Ehret,
Gabriele Facciolo2019-09-04T22:52:55Z2019-09-04T22:00:00Z
Most cameras capture the information of only one color for a given pixel. This results in a mosaicked image that must be interpolated to get three colors at each pixel. The step going from a mosaicked image to a regular RGB image is called demosaicking. This paper studies two recent demosaicking methods based on convolutional neural networks that achieve artifact-free state-of-the-art results: Deep joint demosaicking and denoising by Gharbi et al. and Color image demosaicking via deep residual learning by Tan et al. We show that these methods beat by almost two decibels the best human-crafted methods, while being faster by one order of magnitude. This, arguably, seals the destiny of human-crafted methods on this subject.
The Orthographic Projection Model for Pose Calibration of Long Focal Imageshttp://www.ipol.im/pub/art/2019/248/Laura F. Julià,
Pascal Monasse,
Marc Pierrot-Deseilligny2019-09-04T23:14:36Z2019-09-04T22:00:00Z
Most stereovision and Structure from Motion (SfM) methods rely on the pinhole camera model
based on perspective projection. From this hypothesis the fundamental matrix and the
epipolar constraints are derived, which are the milestones of pose estimation.
In this article we present a method based on the matrix factorization due to Tomasi
and Kanade that relies on a simpler camera model, resulting in orthographic projection.
This method can be used for the pose estimation of perspective cameras in configurations
where other methods fail, in particular, when using cameras with long focal length lenses.
We show this projection is an approximation of the pinhole camera model when the camera is
far away from the scene. The performance of our implementation of this pose estimation method
is compared to that given by the perspective-based methods for several configurations using
both synthetic and real data. We show through some examples and experiments that the accuracy
achieved and the robustness of this method make it worth considering in any SfM procedure.
Extraction of the Level Lines of a Bilinear Imagehttp://www.ipol.im/pub/art/2019/269/Pascal Monasse2019-08-09T23:30:33Z2019-08-09T22:00:00Z
We detail precisely an algorithm for the extraction of the level lines of a bilinear image, which is a continuous function interpolating bilinearly a discrete image. If we discard the levels of the discrete image, where topological difficulties arise, a level line is a concatenation of branches of hyperbolas. The algorithm tracks these branches and provides a sampling of the level lines in the form of closed polygons. If the level line contains a saddle point, the hyperbola degenerates to orthogonal segments, where an arbitrary but consistent choice is adopted for the tracking at the bifurcation. In any case, the extracted polygons are disjoint and enclose a bounded region. This allows to order the level lines in an enclosure tree hierarchy, which may be used for a variety of filters. Recovering this tree is a simple post-processing of the extraction algorithm.
Optimization of Image B-spline Interpolation for GPU Architectureshttp://www.ipol.im/pub/art/2019/257/Thibaud Briand,
Axel Davy2019-08-01T22:45:03Z2019-08-01T22:00:00Z
Interpolation is a vital piece of many image processing pipelines. In this document we present how to optimize the implementation of image B-spline interpolation for GPU architectures. The implementation is based on the one proposed by Briand and Monasse in 2018 and works for orders up to 11. The two main optimizations consist in: (1) transposing the B-spline coefficients before the prefiltering of the rows and (2) dividing columns into subregions in order to use more threads. We assess the impact of the floating point precision and of using high B-spline orders.
Implementation of a Denoising Algorithm Based on High-Order Singular Value Decomposition of Tensorshttp://www.ipol.im/pub/art/2019/226/Fabien Feschet2019-06-12T22:51:37Z2019-06-12T22:00:00Z
This article presents an implementation of a denoising algorithm based on High-Order Singular
Value Decomposition (HOSVD) of tensors. It belongs to the class of patch-based methods such
as BM3D and NL-Bayes. It exploits the grouping of similar patches in a local neighbourhood
into a 3D matrix also called a third order tensor. Instead of performing different processing
in different dimension, as in BM3D for instance, it is based on the decomposition of a tensor
simultaneously in all dimensions reducing it to a core tensor in a similar way as SVD does for
matrices in computing the diagonal matrix of singular values. The core tensor is filtered and
a tensor is reconstructed by inverting the HOSVD. As common in patch-based algorithms, all
tensors containing a pixel are then merged to produce an output image.
Local Assessment of Statokinesigram Dynamics in Time: An in-Depth Look at the Scoring Algorithmhttp://www.ipol.im/pub/art/2019/251/Ioannis Bargiotas,
Julien Audiffren,
Nicolas Vayatis,
Pierre-Paul Vidal,
Alain P. Yelnik,
Damien Ricard2019-06-10T22:14:26Z2019-06-10T22:00:00Z
In this work we discuss the multidimensional scoring approach proposed by Bargiotas and al.
[I. Bargiotas, J. Audiffren, N. Vayatis, P-P. Vidal, S. Buffat, A.P. Yelnik and Damien Ricard,
On the importance of local dynamics in statokinesigram: a multivariate approach for postural control evaluation in elderly,
PloS one, 13 (2018)]
which locally characterizes statokinesigrams - the trajectory of the center of pressure, which is highly correlated with static balance quality - on small time intervals, or blocks. This approach highlights the local dynamics of the trajectories, and we show that the resulting characterization can be used to provide a global score in order to evaluate the postural control. We evaluate our approach using the statokinesigram of 126 community-dwelling elderly, and show that it provides an efficient tool to discriminate between fallers and non-fallers.
Blind Image Deblurring using the l0 Gradient Priorhttp://www.ipol.im/pub/art/2019/243/Jérémy Anger,
Gabriele Facciolo,
Mauricio Delbracio2019-03-12T22:57:11Z2019-03-11T23:00:00Z
Many blind image deblurring methods rely on unnatural image priors that are explicitly designed to restore salient image structures, necessary to estimate the blur kernel.
In this article, we analyze the blur kernel estimation method introduced by Pan and Su in 2013 that uses an l0 prior on the gradient image.
We present deconvolution results using the estimated blur kernels. Our experiments show the effectiveness of the method as well as some of its shortcomings.
An Analysis and Speedup of the FALDOI Method for Optical Flow Estimationhttp://www.ipol.im/pub/art/2019/238/Ferran P. Gamonal,
Coloma Ballester,
Gloria Haro,
Enric Meinhardt-Llopis,
Roberto P. Palomares2019-03-12T22:57:10Z2019-03-07T23:00:00Z
We present a detailed analysis of FALDOI, a large displacement optical flow
method proposed by P.Palomares et al. This method requires a set of discrete
matches, which can be extremely sparse, and an energy functional which
locally guides the interpolation from the matches. It follows a two-step
minimization method at the finest scale which is very robust to the outliers of the
sparse matcher and can capture large displacements of small objects. The results
shown in the original paper consistently outperformed the coarse-to-fine
approaches and achieved good qualitative and quantitative performance on
the standard optical flow benchmarks. In this paper we revise the proposed
method and the changes done to significantly reduce its execution time while
reporting nearly the same accuracy. Finally, we also compare it against the
current state-of-the-art to assess its performance.
Hamiltonian Fast Marching: A Numerical Solver for Anisotropic and Non-Holonomic Eikonal PDEshttp://www.ipol.im/pub/art/2019/227/Jean-Marie Mirebeau,
Jorg Portegies2019-02-24T22:33:50Z2019-02-23T23:00:00Z
We introduce a generalized Fast-Marching algorithm, able to compute paths globally minimizing a measure of length, defined with respect to a variety of metrics in dimension two to five. Our method applies in particular to arbitrary Riemannian metrics, and implements features such as second order accuracy, sensitivity analysis, and various stopping criteria. We also address the singular metrics associated with several non-holonomic control models, related with curvature penalization, such as the Reeds-Shepp's car with or without reverse gear, the Euler-Mumford elastica curves, and the Dubins car. Applications to image processing and to motion planning are demonstrated.
TriplClust: An Algorithm for Curve Detection in 3D Point Cloudshttp://www.ipol.im/pub/art/2019/234/Christoph Dalitz,
Jens Wilberg,
Lukas Aymans2019-04-02T21:29:32Z2019-01-18T23:00:00Z
In this article, we describe an algorithm for detecting and separating curves in 3D point clouds without making a priori assumptions about their parametric shape. The algorithm is called 'TriplClust' because it is based on the idea of clustering point triplets instead of the original points. We define a distance measure on point triplets and then apply a single-link hierarchical clustering on the triplets. The clustering process can be controlled by several parameters, which are described in detail, and suggestions for reasonable choices for these parameters based on the input data are made. Moreover, we suggest a simple criterion for stopping the single link clustering automatically.
An Analysis and Implementation of the FFDNet Image Denoising Methodhttp://www.ipol.im/pub/art/2019/231/Matias Tassano,
Julie Delon,
Thomas Veit2019-09-12T14:42:33Z2019-01-05T23:00:00Z
FFDNet is a recent image denoising method based on a convolutional neural network
architecture. In contrast to other existing neural network denoisers, FFDNet exhibits several
desirable properties such as faster execution time and smaller memory footprint, and the ability
to handle a wide range of noise levels effectively with a single network model. The combination
between its denoising performance and lower computational load makes this algorithm attractive
for practical denoising applications. In this paper we propose an open-source implementation
of the method based on PyTorch, a popular machine learning library for Python. Code for the
training of the network is also provided. We also discuss the characteristics of the architecture
of this algorithm and we compare it to other similar methods.
An Affine Invariant Patch Similarityhttp://www.ipol.im/pub/art/2018/202/Vadim Fedorov,
Coloma Ballester2018-12-24T18:16:24Z2018-12-23T23:00:00Z
Image and video comparison is often approached by comparing patches of visual information. In this work we present a detailed description and implementation of an affine invariant patch similarity measure that performs an appropriate patch comparison by automatically and intrinsically adapting the size and shape of the patches. We also describe the complete implementation of the proposed iterative algorithm for computation of those shape-adaptive patches around each point in the image domain.
EPLL: An Image Denoising Method Using a Gaussian Mixture Model Learned on a Large Set of Patcheshttp://www.ipol.im/pub/art/2018/242/Samuel Hurault,
Thibaud Ehret,
Pablo Arias2018-12-23T19:13:44Z2018-12-22T23:00:00Z
The Expected Patch Log-Likelihood method, introduced by Zoran and Weiss, allows for whole image restoration using a patch-based prior (in the likelihood sense) for which a maximum a-posteriori (MAP) estimate can be calculated. The prior used is a Gaussian mixture model whose parameters are learned from a dataset of natural images. This article presents a detailed implementation of the algorithm in the context of denoising of images contaminated with white additive Gaussian noise. In addition, two possible extensions of the algorithm to handle color images are compared.
Improvements of the Inverse Compositional Algorithm for Parametric Motion Estimationhttp://www.ipol.im/pub/art/2018/222/Thibaud Briand,
Gabriele Facciolo,
Javier Sánchez2018-12-18T22:57:43Z2018-12-17T23:00:00Z
In this work, we propose several improvements of the inverse compositional algorithm for parametric registration. We propose an improved handling of boundary pixels, a different color handling and gradient estimation, and the possibility to skip scales in the multiscale coarse-to-fine scheme. In an experimental part, we analyze the influence of the modifications. The estimation accuracy is at least improved by a factor 1.3 while the computation time is at least reduced by a factor 2.2 for color images.
An Analysis and Implementation of the Shape Preserving Local Histogram
Modification Algorithmhttp://www.ipol.im/pub/art/2018/236/Jose-Luis Lisani2018-12-06T23:02:26Z2018-12-06T23:00:00Z
In this paper we describe the implementation of the
algorithm for local contrast enhancement published by Caselles et al. in 1999.
This algorithm was the first designed explicitly to increase the contrast
while preserving the so-called 'shape structure' of the image, that is, its
set of level sets. According to the mathematical morphology school,
artifacts are created when this structure is modified.
The original algorithm is described and also two alternative implementations
are proposed, which limit the over-enhancement of noise.
Study of the Principal Component Analysis Method for the
Correction of Images Degraded by Turbulencehttp://www.ipol.im/pub/art/2018/47/Tristan Dagobert,
Yohann Tendero,
Stéphane Landeau2018-12-15T22:35:42Z2018-11-22T23:00:00Z
This article analyzes and discusses a well-known paper
[D. Li, R.M. Mersereau and S. Simske, IEEE Letters on Geoscience and Remote Sensing, 3:4 (2007), pp. 340-344]
that applies principal component analysis
in order to restore image sequences degraded by atmospheric turbulence.
We propose a variant of this
method and its ANSI C implementation.
The proposed variant applies to image sequences acquired with short as well
as long exposure times.
Examples of restored images
using sequences of real atmospheric turbulence are presented.
The acquisition of a dataset of image sequences with real atmospheric turbulence is described
and the dataset is made available for download.
An Implementation of the Exposure Fusion Algorithmhttp://www.ipol.im/pub/art/2018/230/Charles Hessel2018-11-17T20:35:04Z2018-11-16T23:00:00Z
Exposure Fusion is a high dynamic range imaging technique to fuse a bracketed
exposure sequence into a high quality image, introduced in 2009 by Mertens et al.
Contrarily to most HDR imaging methods, exposure fusion does not construct an
intermediate HDR image but directly constructs the final LDR one by seamlessly
fusing the best regions of the input sequence, using the Laplacian pyramid.
Since its publication, this method received considerable attention, being both
effective and efficient. We propose in this paper a precise description of the method
and an analysis of its main limitation, an out-of-range artifact.
Structural Similarity Metrics for Quality Image Fusion Assessment: Algorithmshttp://www.ipol.im/pub/art/2018/196/Silvina Pistonesi,
Jorge Martinez,
Silvia Maria Ojeda,
Ronny Vallejos2018-10-25T22:16:44Z2018-10-25T22:00:00Z
The wide use of image fusion techniques in different fields such as medical diagnostics, digital camera vision, military and surveillance applications, among others, has motivated the development of various image quality fusion metrics, in order to evaluate them. In this paper, we study and implement the algorithms of non-reference image structural similarity based metrics for fusion assessment: Piella's metric, Cvejic's metric, Yang's metric, and Codispersion Fusion Quality metric. We conduct the comparative experiment of the selected image fusion metrics over four multiresolution image fusion algorithms, performed on different pairs of images used in different applications.
Interpolation of Missing Samples in Sound Signals Based on Autoregressive Modelinghttp://www.ipol.im/pub/art/2018/23/Laurent Oudre2018-10-17T21:56:33Z2018-10-16T22:00:00Z
This article proposes an implementation and a study of the paper 'Adaptive Interpolation of Discrete-Time Signals That Can Be Modeled as Autoregressive Processes' by Janssen et al. The algorithm presented in this paper allows one to reconstruct an audio signal which presents localized degradations by interpolating the missing samples. This method assumes that the signal can locally be modeled as a realization of an autoregressive process and iteratively estimates the model parameters and the interpolated samples by minimizing a quadratic criterion. We investigate the limits and the algorithmic aspects of this method on several audio examples.
An Analysis and Implementation of the Harris Corner Detectorhttp://www.ipol.im/pub/art/2018/229/Javier Sánchez,
Nelson Monzón,
Agustín Salgado2018-10-03T10:26:21Z2018-10-02T22:00:00Z
In this work, we present an implementation and thorough study of the Harris corner detector. This feature detector relies on the analysis of the eigenvalues of the autocorrelation matrix. The algorithm comprises seven steps, including several measures for the classification of corners, a generic non-maximum suppression method for selecting interest points, and the possibility to obtain the corners position with subpixel accuracy. We study each step in detail and propose several alternatives for improving the precision and speed. The experiments analyze the repeatability rate of the detector using different types of transformations.
Estimating an Image's Blur Kernel Using Natural Image Statistics, and Deblurring it: An Analysis of the Goldstein-Fattal Methodhttp://www.ipol.im/pub/art/2018/211/Jérémy Anger,
Gabriele Facciolo,
Mauricio Delbracio2018-09-26T21:47:21Z2018-09-25T22:00:00Z
Despite the significant improvement in image quality resulting from improvement in optical sensors and general electronics, camera shake blur significantly undermines the quality of hand-held photographs. In this work, we present a detailed description and implementation of the blur kernel estimation algorithm introduced by Goldstein and Fattal in 2012. Unlike most methods that attempt to solve an inverse problem through a variational formulation (e.g. through a Maximum A Posteriori estimation), this method directly estimates the blur kernel by modeling statistical irregularities in the power spectrum of blurred natural images. The adopted mathematical model extends the well-known power-law by contemplating the presence of dominant strong edges in particular directions. The blur kernel is retrieved from an estimation of its power spectrum, by solving a phase retrieval problem using additional constraints associated with the particular nature of camera shake blur kernels (e.g. non-negativity and small spatial support). Although the algorithm is conceptually simple, its numerical implementation presents several challenges. This work contributes to a detailed anatomy of the Goldstein and Fattal method, its algorithmic description, and its parameters.
Fast Affine Invariant Image Matchinghttp://www.ipol.im/pub/art/2018/225/Mariano Rodríguez,
Julie Delon,
Jean-Michel Morel2018-09-30T21:43:24Z2018-09-23T22:00:00Z
Methods performing Image Matching by Affine Simulation (IMAS) attain affine invariance by applying a finite set of affine transforms to the images before comparing them with a Scale Invariant Image Matching (SIIM) method like SIFT or SURF. We describe here how to optimize IMAS methods. First, we detail an algorithm computing a minimal discrete set of affine transforms to be applied to each image before comparison. It yields a full practical affine invariance at the lowest computational cost. The matching complexity of current IMAS algorithms is divided by about 4. Our approach also associates to each image an affine invariant set of descriptors, which is twice smaller than the set of descriptors usually used in IMAS methods, and only 6.4 times larger than the set of similarity invariant descriptors of SIIM methods. In order to reduce the number of false matches, which are inherently more frequent in IMAS approaches than in SIIM, we introduce the notion of hyper-descriptor, which groups descriptors whose keypoints are spatially close. Finally, we also propose a matching criterion allowing each keypoint of the query image to be matched with several keypoints of the target image, in order to deal with situations where an object is repeated several times in the target image.
Numerical Simulation of Landscape Evolution Modelshttp://www.ipol.im/pub/art/2018/205/Marc Lebrun,
Miguel Colom,
Jérôme Darbon,
Jean-Michel Morel2018-09-03T19:51:45Z2018-09-02T22:00:00Z
This paper gives the complete numerical schemes implementing the main physical laws proposed in landscape evolution models (LEMs). These laws can be modeled by a system of three partial differential equations governing water runoff, stream incision, hill slope evolution and sedimentation. The goal of the presented algorithm, code and online demo is to be able to test these equations on digital elevation models (DEMs) of any resolution, and to illustrate its potential to simulate the fine structure of the river network, and to understand the landscape morphology and its causes. The equations simulate plausible evolutions. We illustrate experiments on DEMs of several sites, including one site, La Réunion where the DEM is given at three different resolutions: the SRTM resolution (90m), and then 12m and 4m on DEMs derived from several Pléiades pairs. Other many DEMs are proposed in the online demo, which allows to upload and tests other DEMs.
An Analysis and Implementation of Multigrid Poisson Solvers With Verified Linear
Complexityhttp://www.ipol.im/pub/art/2018/228/Matías Di Martino,
Gabriele Facciolo2018-09-04T21:52:26Z2018-07-25T22:00:00Z
The Poisson equation is the most studied partial differential equation, and it
allows to formulate many useful image processing methods in an elegant and efficient
mathematical framework. Using different variations of data terms and boundary conditions,
Poisson-like problems can be developed, e.g., for local contrast enhancement, inpainting, or
image seamless cloning among many other applications. Multigrid solvers are among the most
efficient numerical solvers for discrete Poisson-like equations. However, their correct
implementation relies on: (i) the proper definition of the discrete problem, (ii) the right
choice of interpolation and restriction operators, and (iii) the adequate
formulation of the problem across different scales and layers. In the present work we address
these aspects, and we provide a mathematical and practical description of multigrid methods. In
addition, we present an alternative to the extended formulation of Poisson equation
proposed in 2011 by Mainberger et al. The proposed formulation of the problem suits better
multigrid methods, in particular, because it has important mathematical properties that can be
exploited to define the problem at different scales in a intuitive and natural way. In
addition, common iterative solvers and Poisson-like problems are empirically analyzed and
compared. For example, the complexity of problems is compared when the topology of Dirichlet
boundary conditions changes in the interior of the regular domain of the image. The main
contribution of this work is the development and detailed description of an implementation of a
multigrid numerical solver which converges in linear time.