Kornkamol Thakulsukanant

An experimental performance analysis of image reconstruction techniques under both Gaussian and non-Gaussian noise models

Recently, the images reconstruction approaches are very essential in digital image processing (DIP), especially in terms of removing the noise contaminations and recovering the content of images. Each image reconstruction approach has different mathematical models. Therefore a performance of individual reconstruction approach is varied depending on several factors such as image characteristic, reconstruction mathematical model, noise model and noise intensity. Thus, this paper presents comprehensive experiments based on the comparisons of various reconstruction approaches under Gaussian and non-Gaussian noise models. The employing reconstruction approaches in this experiment are Inverse Filter, Wiener Filter, Regularized approach, Lucy-Richardson (L-R) approach, and Bayesian approach applied on mean, median, myriad, meridian filters together with several regularization techniques (such as non-regularization, Laplacian regularized, Markov Random Field (MRF) regularization, and one-side Bi-Total Variation (OS-BTV) regularization). Three standard images of Lena, Resolution Chart, and Susie (40th) are used for testing in this experiment. Noise models of Additive White Gaussian Noise (AWGN), Poisson, Salt&Pepper, and Speckle of various intensities are used to contaminate all these images. The comparison is done by varying the parameters of each approach until the best peak-signal-to-noise ratio (PSNR) is obtained. Therefore, PSNR plays a vital parameter for comparisons all the results of individual approaches.

An alternative single-image super resolution framework employing high frequency prediction using a robust huber rational function

In general prospective, SI-SR or Single-Image Super-Resolution, which is one of the most useful algorithms of Super Resolution-Reconstruction (SRR) algorithms, is a mathematical procedure for acquiring a high-resolution image from only one coarse-resolution image, which is usually computed by Digital Image Processing (DIP). Even thought there have been substantially researched during the last decade, Single - Image Super-Resolution for applying on real implementations still keeps throw down the gauntlet. One of the practical Single-Image Super-Resolution is the resolution enhancement using prediction of the high-frequency image because of its high performance and its less complexity however the rational function C(x, y) of high-frequency image prediction process of this technique is depend upon three parameters (b, h, k) therefore the parameter turning is difficult for maximizing its performance. From this problem prospective, this paper presents the alternative SI-SR framework employing robust rational function based on Huber function, which is depend upon only one parameter (T), instead of three parameters like the rational function C(x, y). Using up to fourteen standard images, which are crooked by varied noise models, in analysis testing section, the proposed SI-SR is demonstrated to be somewhat simper than the original SI-SR with equivalent efficiency because the saving in parameter turning time will be very important for SI-SR in real implementations.

An Alternative Robust SL0 Based on Recursive Huber Stochastic Estimation Technique in Frequency Domain

In this paper, we propose a novel robust algorithm based on Huber norm estimation to recover the signal under several noise models by using only a few components. The proposed algorithm is used with a matrix that the number of row is obviously fewer than the column. It is stated that if the signal or the image is sufficiently sparse, we can recover it from a small number of linear measurements by solving a convex program of the measurement vector. For most of the image, the signal information tends to be concentrated in the low frequency components of the frequency domain. The performance of the proposed algorithm is compared with other classical algorithms such as L1 and SL0 norm estimations. Finally, the experimental results are presented on both synthesis and real image under various kinds of noises (AWGN, Salt & Pepper Noise and Speckle Noise) with different powers. The effects of different noise models are compared in order to show the improvement of Huber over SL0 Norm estimation.

A Robust Video Super-Resolution Using a Recursive Leclerc Bayesian Approach with an OFOM (Optical Flow Observation Model)

Due to the inaccuracy of image registration between each frame of observed sequence, especially in a very complex motion frame, almost Video Super Resolution Reconstruction (SRR) frameworks found in review literatures cannot be worked well to real sequences with arbitrary scene content and/or arbitrary motion. Moreover, the observed system noise is typically assumed to be a Gaussian distribution thus the performance of SRR algorithm is usually degraded when the real system noise is non-Gaussian distribution. This paper proposes the alternative SRR framework for applying in the complex sequences and for against non-Gaussian noise models. First, the proposed SRR framework is based on classical stochastic ML (Minimization Likelihood) framework using L1, L2 and Leclerc norm estimations in order to measure the difference between the projected estimation of the reconstructed image and each observed images and to remove noise in the observed images. Later, the proposed algorithm is used an Optical Flow Observation Model (OFOM) based on 2D optical flow Block-Based Full (BOF) search algorithm for coping with complex motion between two frames of observed sequences. Finally, the experimental section shows that the proposed framework can be well effectively worked on real sequences such as Susie and Foreman sequences under several Gaussian and Non-Gaussian noise models (such as AWGN, Poisson, Salt & Pepper noise and Speckle) at different noise powers.

Experimental Investigation for Practical Sparsity Number for Image Reconstruction Based on SL0 Algorithm in Discrete Frequency Domain

Compressive Sensing (CS) is known as a new sampling theory that use small number of basis elements for constructing of signals (or images) and these basis elements (so called sparsity number) are very important parameters that approximate how sparsify the image is. Due to several characteristics of each image groups, the sparsity number could be varied and there is unfortunately very little research for this issue. This paper presents two main contributions: First, this paper proposes a practical sparsity number estimation technique using for an image reconstruction for SL0 algorithm based on Discrete Cosine Transform domain (DCT). Second the practical sparsity number of difference image groups is the experiment based on over 2000 images. The DCT is exclusively applied for a sparse representation of images because it is proven as a useful instrument for image analysis and processing. In general, images can be represented by a linear superposition of small number of wavelet elements selected from a suitable filter. The proposed models process the image with Smoothed norm algorithm. This algorithm stated that if signal or image is sufficiently sparse, we can reconstruct it from small amount of none zero basis components. The experiment is comprehensively tested under 2000 sampling images that are categorized in 18 groups by their characteristics. Moreover this sparsity number is practically used for any CS algorithm based on DCT domain.

A Performance Comparison of Single Image Reconstruction Techniques under Several Noisy Environments

Due to noise contamination on the image during the observation process, digital image reconstruction is an essential in terms of recovering the information of the contents (e.g. document and image) and utilized in many applications such as digital image forensic, medical image processing, machine vision, and etc. Therefore, this paper is concerned with the performance comparisons of single image employing various reconstruction approaches. These are Inverse filter, Wiener filter, Regularized technique, Lucy-Richardson technique, and Bayesian technique based on median, mean, myriad, and meridian filters. The experiments test on the three standard pictures (Lena, Resolution chart, and Susie (40th)) under the same noise conditions. Four types of noise models consider in this paper are AWGN, Poisson, Salt&Pepper, and Speckle noises. The performance of evaluations is done by varying parameters of individual technique. Peak-signal-to-noise-ratio (PSNR) is a key indicator on the performance comparison results.