UPM Institutional Repository

Error concealment technique using wavelet neural network for wireless transmitted digital images


Citation

Al-Azzawi, Alaa Khamees (2012) Error concealment technique using wavelet neural network for wireless transmitted digital images. PhD thesis, Universiti Putra Malaysia.

Abstract

Continuous flow to send images via encrypted wireless channels may reduce the efficiency of transmission. This is due to the damage or loss of the numerous large-blocks from these images. Therefore, it is difficult to rebuild these blocks from the point of reception. In addition, some packets may be lost via the limited bandwidth networks when congestion occurs. However, compressing of bit-streams of images with variable length coding (VLC) will add another burden to the transmission in cases where these streams are sent via the noisy channels. Several techniques have been proposed in the past decade, particularly error concealment (EC) algorithms. In this study, attention is focused on the algorithms that have high efficiency to fill-in the damaged patches, for example, Multi-directional interpolation (MDI), neighboring matching EC, etc. Accordingly, two frameworks are proposed to tackle the following problems: (1) loss of the damaged blocks; (2) artifacts that appear after the process of filling-in (i.e., blockiness and blurring), and (3) white-Gaussian noise. The first framework is proposed to compensate for the loss of damaged patches. The compensation information includes patches that differ in size and location. Moreover, the concealment includes patches for both smooth and nonsmooth areas. This framework is implemented in three key steps. The first step is devoted to detecting invisible patches. Further, an efficient color contrast scheme is proposed. The goal behind this scheme is to detect invisible patches that are either occurred in a dusky areas of the image or have been deliberately hidden during the encryption process. Next, a multi-directional interpolation (MDI) technique is proposed. The method is used to estimate the lost coefficients in the wavelet-domain. Accurate values for the weighting coefficients were efficiently calculated to minimize the mean squared errors (MSE) at the top, bottom, left, and right of the missing blocks. The proposed EC schemes in the second step have been used to compensate for the loss of different damaged patches at the geometric information, with minimal artifacts from blurring and blockiness, as well as improving the Peak Signal-to-noise ratio (PSNR). In the meantime, a new technique called 'wavelet neural network' is proposed and implemented in the third step. This technique merges the estimated matrices' results of the proposed schemes in the second step with an artificial neural network (ANN), with the intention of obtaining results with high accuracy, as well as to overcome all the problems that may arise after the process of filling-in (blurring and blockiness artifacts). The neural network architecture that can be used to implement a nonlinear vector predictor, including a multilayer perceptron (MLP), and a radial basis function network. In the case, where the missing regions of pixels are containing textures, edges, and other image features that are not easily handled by concealment algorithms. It therefore, necessitated to use denoising rather than EC algorithms. Finally, the proposed second framework is exploited for denoising a chain of images that are affected by white-Gaussian noise to the lowest possible rates, as well as concealing these images. Experimental results demonstrate that the proposed methods simultaneously provide significant improvements in terms of both loss concealment and artifacts, especially those associated with edges. The performance efficiency has been computed in terms of the PSNR, and mean squared error (MSE) yielded 87% accuracy and tested for various images and combinations of lost blocks (CT image). The results are similar to those shown in [3], [98], and [129] with a noticeable interference pattern in the reconstruction from the uniform array. The reconstructed images by our schemes produced PSNR ranges from 33 dB to 37 dB and the lowest MSE values are obtained for percentages near to 50%.


Download File

[img]
Preview
PDF
FK 2012 148RR.pdf

Download (1MB) | Preview

Additional Metadata

Item Type: Thesis (PhD)
Subject: Error-correcting codes (Information theory)
Subject: Neural networks (Computer science)
Subject: Wavelets (Mathematics)
Call Number: FK 2012 148
Chairman Supervisor: Associate Professor M. Iqbal Saripan, PhD
Divisions: Faculty of Engineering
Depositing User: Haridan Mohd Jais
Date Deposited: 28 Mar 2017 08:43
Last Modified: 28 Mar 2017 08:43
URI: http://psasir.upm.edu.my/id/eprint/51580
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item