The Young Researchers in Imaging Seminars aims at offering the opportunity to PhD students and Post Docs working on topics related to the thematic program to present their work.
Starting from Wednesday 20 February, we organise every Wednesday (except for March 13th) in Darboux amphiteater a two-hour session from 14:00 to 16:00 dedicated to young researchers, in order give them a chance to stimulate the scientific discussion. The seminar is followed by a coffee time (with snacks) at the 2nd floor.
Image reconstruction consists in recovering an image from an indirect observation (for instance its Radon transform). In general this observation does not allow to determine a unique image and some prior (e.g. image regularity) needs to be incorporated in the reconstruction framework. I will present how one can incorporate intuitive priors about the geometric variation of the image from a reference one using the framework of deformation modules. The framework of deformation modules allows to build deformations satisfying some prior and the idea is to reconstruct an image --from some indirect observations-- as the deformation of the reference one while constraining the deformation to satisfy certain constraints. I will present this notion of deformation modules and show how it can be used to perform image reconstruction.
Many problems in machine learning and imaging can be framed as an infinite dimensional Lasso problem to estimate a sparse measure. This includes for instance regression using a continuously parameterized dictionary, mixture model estimation and super-resolution of images. To make the problem tractable, one typically sketches the observations (often called compressive-sensing in imaging) using randomized projections. In this work, we provide a comprehensive treatment of the recovery performances of this class of approaches. We show that for a large class of operators, the Fisher-Rao distance induced by the measurement process is the natural way to enforce and generalize the classical minimal separation condition appearing in the literature. We then prove that (up to log factors) a number of sketches proportional to the sparsity is enough to identify the sought after measure with robustness to noise. Finally, we show that, under additional hypothesis, exact support stability holds (the number of recovered atoms matches that of the measure of interest) when the level of noise is smaller than a specified value. This is a joint work with Clarice Poon (University of Bath) and Gabriel Peyré (ENS).
The geometry-texture decomposition of images produced by X-Ray Computed Tomography (CT) is a challenging inverse problem, which is usually performed in two steps: reconstruction and decomposition. Decomposition can be used for instance to produce an approximate segmentation of the image, but this one can be compromised by artifacts and noise arising from the acquisition and reconstruction processes. Hence, reconstruction and decomposition benefit from being performed in a joint manner. We propose a geometry-texture decomposition based on a TV-Laplacian model, well-suited for segmentation and edge detection. The problem of joint reconstruction and decomposition of CT data is then formulated as a convex constrained minimization problem, which is solved using a recently introduced proximal interior point method. Numerical experiments on realistic images of material samples illustrate the practical efficiency of the proposed approach.
Some recent denoising methods are based on a statistical modeling of the image patches. In the literature, Gaussian models or Gaussian mixture models are the most widely used priors. In this presentation, after introducing the statistical framework of patch-based image denoising, I will propose some clues to answer the following questions: Why are these Gaussian priors so widely used? What information do they encode? In the second part, I will present a mixture model for noisy patches adapted to the high dimension of the patch space. This results in a denoising algorithm only based on statistical tools, which achieves state-of-the-art performance. Finally, I will discuss the limitations and some developments of the proposed method.
Key to structured prediction is exploiting the problem structure to simplify the learning process. A major challenge arises when data exhibit a local structure (e.g., are made by "parts") that can be leveraged to better approximate the relation between (parts of) the input and (parts of) the output. Recent literature on signal processing, and in particular computer vision, has shown that capturing these aspects is indeed essential to achieve state-of-the-art performance. While such algorithms are typically derived on a case-by-case basis, in this talk we illustrate the first theoretical and algorithmic framework able (a) to deal with part-based data from a general perspective and (b) to provide generalization properties within the setting of statistical learning theory. Our analysis is novel in that it explicitly quantifies the benefits of leveraging the part-based structure of the problem with respect to the learning rates of the proposed estimator.
When considering an image with different local properties, different local regularization terms may be needed to recover the original image, or something close to it, starting from a noisy and blurred observation. I will present a new space-variant anisotropic regularization term for variational image restoration, based on the statistical assumption that the gradients of the target image distribute locally according to a bivariate generalized Gaussian distribution. The free parameters encoded in the proposed regularizer, that are automatically estimated via a maximum likelihood approach, hold the potential for faithfully modelling the local geometry in the image and describing local orientation preferences.
No seminar this week : it is the “Statistical Modeling for Shapes and Imaging” workshop.
In this seminar I will discuss methods and some results from a paper I have recently submitted. My research is focused on machine learning methods for measuring the Sustainable Development Goals (SDGs) which are a set of priorities the United Nations and World Bank have set for countries to reach in order to improve quality of life and environment globally by 2030. Free satellite images have been identified as a key resource which can be used to produce official statistics and analysis to measure progress towards SDGs, especially those concerned with the physical environment such as forest, water and crops. An issue with using satellite images, particularly in tropical areas, are missing data due to cloud cover. There are existing effective methods for filling in image gaps, however these are computationally slow and not effective at pixel scale. To address this, I use spatial implementations of machine learning algorithms; gradient boosted machine and random forest, to accurately and quickly classify observed and simulated ‘missing’ pixels in satellite images as either grassland or woodland, and predict a continuous value of Foliage Projective Cover (FPC) using the method.
I will briefly introduce the notions of generalized averages, power mean, their particular cases, analysis and level set representation. We apply these generalized averages and power mean to construct a general image data term. The properties of the general data term will also be discussed for multi-region image segmentation and handling outliers. Few test results will be exhibited. Moreover, performance of a joint segmentation and de-hazing model will also be displayed. This is a joint work with Noor Badshah, Ke Chen, Gulzar Ali Khan and Nosheen, Lavdi Rada, Awal Sher, Afzal, Haroon and Amna Shujah.
Solar flares are sudden flashes of brightness on the surface of the Sun, which can strongly affect satellite operations, aviation and communication technologies. We will recover images of solar flares from NASA RHESSI satellite by assuming that the data have been generated by a combination of few simple geometrical objects, whose parameters are estimated with a sequential Monte Carlo method. This approach recovers improved images of flares, with the additional advantage of providing uncertainty quantification of the estimated parameters. Syntetic and real data will be shown.
Visual illusions teach us that what we see is not always what it is represented in the physical world. Its special nature make them a fascinating tool to test and validate any new vision model proposed. In general, current vision models are based on the concatenation of linear convolutions and non-linear operations. In this work we get inspiration from the similarity of this structure with the operations present in Convolutional Neural Networks (CNNs). This motivated us to study if CNNs trained for typical image processing task (e.g. denoising) are deceived by visual illusions. The answer to this question opens a new bridge between human perception and CNNs: in order to obtain CNNs that better replicate human behavior, we may need to start aiming for them to better replicate visual illusions.
We will bind together and extend some recent developments regarding data-driven non-smooth regularization techniques in image processing through the means of bilevel minimization schemes. The schemes, considered in function space, take advantage of dualization frameworks and they are designed to produce spatially varying regularization parameters adapted to the data for well-known regularizers, e.g. Total Variation and Total Generalized Variation, leading to automated (monolithic), image reconstruction workflows.