Showing papers by "Karen Egiazarian published in 2019"

The Practicality of Stochastic Optimization in Imaging Inverse Problems

Abstract: In this work we investigate the practicality of stochastic gradient descent and recently introduced variants with variance-reduction techniques in imaging inverse problems. Such algorithms have been shown in the machine learning literature to have optimal complexities in theory, and provide great improvement empirically over the deterministic gradient methods. Surprisingly, in some tasks such as image deblurring, many of such methods fail to converge faster than the accelerated deterministic gradient methods, even in terms of epoch counts. We investigate this phenomenon and propose a theory-inspired mechanism for the practitioners to efficiently characterize whether it is beneficial for an inverse problem to be solved by stochastic optimization techniques or not. Using standard tools in numerical linear algebra, we derive conditions on the spectral structure of the inverse problem for being a suitable application of stochastic gradient methods. Particularly, we show that, for an imaging inverse problem, if and only if its Hessain matrix has a fast-decaying eigenspectrum, then the stochastic gradient methods can be more advantageous than deterministic methods for solving such a problem. Our results also provide guidance on choosing appropriately the partition minibatch schemes, showing that a good minibatch scheme typically has relatively low correlation within each of the minibatches. Finally, we propose an accelerated primal-dual SGD algorithm in order to tackle another key bottleneck of stochastic optimization which is the heavy computation of proximal operators. The proposed method has fast convergence rate in practice, and is able to efficiently handle non-smooth regularization terms which are coupled with linear operators.

Hyperspectral data denoising for terahertz pulse time-domain holography

Abstract: We investigated data denoising in hyperspectral terahertz pulse time-domain holography. Using the block-matching algorithms adapted for spatio-temporal and spatio-spectral volumetric data we studied and optimized parameters of these algorithms to improve phase image reconstruction quality. We propose a sequential application of the two algorithms oriented on work in temporal and spectral domains. Experimental data demonstrate the improvement in the quality of the resultant time-domain images as well as phase images and object’s relief. The simulation results are proved by comparison with the experimental ones.

Exploiting Task Parallelism with OpenCL: A Case Study

Abstract: While data parallelism aspects of OpenCL have been of primary interest due to the massively data parallel GPUs being on focus, OpenCL also provides powerful capabilities to describe task parallelism. In this article we study the task parallel concepts available in OpenCL and find out how well the different vendor-specific implementations can exploit task parallelism when the parallelism is described in various ways utilizing the command queues. We show that the vendor implementations are not yet capable of extracting kernel-level task parallelism from in-order queues automatically. To assess the potential performance benefits of in-order queue parallelization, we implemented such capabilities to an open source implementation of OpenCL. The evaluation was conducted by means of a case study of an advanced noise reduction algorithm described as a multi-kernel OpenCL application.

Terahertz pulse time-domain holography with balance detection: complex-domain sparse imaging.

Abstract: We investigated the peculiarities of the terahertz pulse time-domain holography principle in the case of raster scanning with the balance detection system. The noise in this system represents a Skellam distribution model, which differentiates it from systems based on a photoconductive antenna. We analyzed this Skellam model and provided both numerical and experimental investigations. We found that the variance of the noise in the balance detection system does not depend on the true signal. Complex-domain images obtained in this model are filtered by block-matching algorithms adapted for spatio-temporal and spatiospectral volumetric data. We presented a new cube complex-domain filter algorithm that uses block matching in all 3D data sets simultaneously in spatial and frequency coordinates. A combination of temporal and complex-domain filters allows us to expand the dynamic range of terahertz frequencies for which we can obtain amplitude/phase information. Experimental data demonstrate an improvement in the quality of the resultant images both in the time domain and complex-spectral domain. The simulation and experimental results are in good agreement.

Spectral Object Recognition in Hyperspectral Holography with Complex-Domain Denoising

Abstract: In this paper, we have applied a recently developed complex-domain hyperspectral denoiser for the object recognition task, which is performed by the correlation analysis of investigated objects’ spectra with the fingerprint spectra from the same object. Extensive experiments carried out on noisy data from digital hyperspectral holography demonstrate a significant enhancement of the recognition accuracy of signals masked by noise, when the advanced noise suppression is applied.

Lensless broadband diffractive imaging with improved depth of focus: wavefront modulation by multilevel phase masks

Abstract: This paper introduces a novel lensless full colour diffractive computational imaging system with a planar Multilevel Phase Mask (MPM) as a diffractive optical element (DOE). The novelty concerns: a methodology of MPM design for improved depth of focus (DoF); design of PSFs for RGB imaging and an inverse imaging algorithm with sparse colour image modelling simultaneous for all RGB channels. MPMs are step-wise invariant. The cubic wavefront coding (WFC) is incorporated in MPMs with optimization of number of levels and width of invariant steps. This design of MPM makes the system robust with respect to defocus (improves DoF) and diminish chromatic aberrations typical for DOEs. Broadband multichannel test-images are exploited for design and testing of the lensless system. We consider two alternative optical setups: Wavelength Multiplexing (WM) and Wavelength Division (WD). In WM, the light beam is broadband multichannel with light sources radiating all wavelengths simultaneously and a CMOS sensor is equipped ...

Deep Convolutional Autoencoder for Estimation of Nonstationary Noise in Images

Abstract: A precise estimation of noise parameters is very important in many image processing applications, such as denoising, deblurring, compression, etc. This problem is well studied for the case of stationary noise in images, and much less studied for the case of nonstationary noise. In this paper, we develop an efficient method of nonstationary noise variance estimation in image regions, based on specially designed deep convolutional autoencoder (DCAE) with a small dimensionality reduction. Training of the proposed DCAE is carried out for a large set of image blocks, including fragments of noise free textures, faces and texts. In the numerical analysis, we compare the proposed method and method of blind estimation of nonstationary noise, based on block matching (BM). Additionally, we analyze efficiency of the proposed DCAE in comparison with the conventional autoencoder (AE). We show that usage of the proposed DCAE provides an error of noise variance estimation about 2 times smaller, that the error when the standard AE is used, and 4 times smaller than the variance estimation error of the BM method.

The Limitation and Practical Acceleration of Stochastic Gradient Algorithms in Inverse Problems

Abstract: In this work we investigate the practicability of stochastic gradient descent and recently introduced variants with variance-reduction techniques in imaging inverse problems, such as space-varying image deblurring. Such algorithms have been shown in machine learning literature to have optimal complexities in theory, and provide great improvement empirically over the full gradient methods. Surprisingly, in some tasks such as image deblurring, many of such methods fail to converge faster than the accelerated full gradient method (FISTA), even in terms of epoch counts. We investigate this phenomenon and propose a theory-inspired mechanism to characterize whether a given inverse problem should be preferred to be solved by stochastic optimization technique with a known sampling pattern. Furthermore, to overcome another key bottleneck of stochastic optimization which is the heavy computation of proximal operators while maintaining fast convergence, we propose an accelerated primal-dual SGD algorithm and demonstrate the effectiveness of our approach in image deblurring experiments.

Enhancement of Component Images of Multispectral Data by Denoising with Reference

Abstract: Multispectral remote sensing data may contain component images that are heavily corrupted by noise and the pre-filtering (denoising) procedure is often applied to enhance these component images. To do this, one can use reference images—component images having relatively high quality and that are similar to the image subject to pre-filtering. Here, we study the following problems: how to select component images that can be used as references (e.g., for the Sentinel multispectral remote sensing data) and how to perform the actual denoising. We demonstrate that component images of the same resolution as well as component images of a better resolution can be used as references. To provide high efficiency of denoising, reference images have to be transformed using linear or nonlinear transformations. This paper proposes a practical approach to doing this. Examples of denoising tests and real-life images demonstrate high efficiency of the proposed approach.

Hyperspectral holography and spectroscopy: computational features of inverse discrete cosine transform

Abstract: Broadband hyperspectral digital holography and Fourier transform spectroscopy are important instruments in various science and application fields. In the digital hyperspectral holography and spectroscopy the variable of interest are obtained as inverse discrete cosine transforms of observed diffractive intensity patterns. In these notes, we provide a variety of algorithms for the inverse cosine transform with the proofs of perfect spectrum reconstruction, as well as we discuss and illustrate some nontrivial features of these algorithms.

Blind Prediction of Original Image Quality for Sentinel Sar Data

Abstract: Synthetic aperture radar (SAR) images are often subject to visual inspection and analysis. Many factors impact on visual quality of SAR images, such as properties of speckle, dynamic range of data, etc. Thus, the corresponding metrics have to be applied and it is worth predicting their values before one starts analyzing images. Using a set of input parameters (both statistical and spectral) and a trained neural network (NN), we show that full-reference visual quality metrics can be predicted for images acquired by modern SAR Sentinel-1. A prediction accuracy is studied and verified on real-life examples. The source codes and datasets will be made publicly available at https://github.com/asrubel/EUVIP2019.

Use of Modified BM3D Filter and CNN Classifier for SAR Data to Improve Crop Classification Accuracy

Abstract: Monitoring of agricultural regions is an important task. Recent trends to solve it are based on applying multi-temporal remote sensing in order to obtain reliable crop classification maps. If a radar remote sensing is used, speckle presence in the original data reduces a classification accuracy. A negative impact of speckle can be reduced by image prefiltering procedure. Recent studies carried out for Sentinel-1 imagery have shown that more efficient pre-filtering usually results in better classification. Thus, here we propose a modification of block matching three-dimensional (BM3D) filter adapted to the properties of Sentinel-1 radar data. We demonstrate that its use leads to improved classification of crops in agricultural region of Ukraine compared to earlier methods which are based on other filters. Additionally, a convolutional neural network approach for crop mapping is investigated in this paper for decreasing noise impact and is compared with a traditional pixel-based multi-layer perceptron.

Single exposure lensless subpixel phase imaging

Abstract: Lensless phase-retrieval system with phase modulation of free propagation wavefront is proposed. Contrary to the traditional super-resolution phase-retrieval, the method in this paper requires a single observation only and uses advanced SR-SPAR iterative technique. Successful object imaging relies on modulation of the object wavefront with a random phase-mask, which generates enlarged intensity patterns, allowing us to extract more information than it is possible without such a mask. The achieved high-quality super-resolution phase-imaging is demonstrated by simulation-tests produced with the parameters corresponding to the physical prototype of the considered optical system.

Hyperspectral phase imaging with denoising in SVD image subspace

Abstract: We propose a modified denoising algorithm for hyperspectral data. The algorithm is based on a complex domain block-matching 3D filter, on estimation of the noise correlation matrix and on dimension reduction of the Singular Value Decomposition (SVD) eigenspace.

Hyperspectral terahertz pulse time-domain holography: noise filtering

Abstract: Using the block-matching denoising algorithms adapted to spatio-temporal and spatio-spectral volumetric data, we studied and optimized the parameters of these algorithms to improve the phase/amplitude image reconstruction in hyperspectral terahertz pulse time-domain holography. We propose a sequential application of two different algorithms oriented on work in temporal and spectral domains. Experimental data demonstrates essential improvement in the quality of the resulting phase imaging.

Hyperspectral complex domain denoising

Abstract: We consider hyperspectral complex domain imaging from hyperspectral complex-valued noisy observations. The proposed algorithm is based on singular value decomposition (SVD) of observations and complex domain block-matching 3D (CDBM3D) filtering in optimized SVD eigenspace. Simulation experiments demonstrate high efficiency of the proposed complex domain joint filtering of hyperspectral data in comparison with CDBM3D filtering of separate 2D slices of hyperspectral cubes as well as with respect to joint real domain independent phase/amplitude filtering this kind of data.

Near Lossless JPEG Compression Based on Masking Effect of Non-predictable Energy of Image Regions

Abstract: This paper studies near lossless JPEG image compression. A method of estimation of image regions masking ability (maximal level of distortions invisible for human visual system) using non-predictable energy of image regions is described. A novel method of zeroing quantized DCT coefficients of JPEG images to increase their compression ratio without introducing visible distortions is proposed. A numerical analysis of effectiveness of the proposed near lossless compression method using 300 noise free test images of TAMPERE17 database is carried out. It is shown that the proposed method provides an increase of compression ratio of JPEG images without visible distortions at about 1.35 times in average. Additionally, the proposed method results in decreasing of variability of compression ratio values for different images. It is shown that the proposed method increases minimal compression ratio for highly textured JPEG images from 1.1…1.5 times to 2 times. Carried out experiments demonstrated once again that the traditional PSNR metric does not correspond to human perception for this task.