Pakistan Research Repository Home

Title of Thesis

Tanweer Ahmed Cheema
Institute/University/Department Details
Department of Electronics Engineering/ Mohammad Ali Jinnah University
Electronics Engineering
Number of Pages
Keywords (Extracted from title, table of contents and abstract of thesis)
blind disconsolation, still images, hybrid computing, blind image restoration, moving average (ma) process, artificial neural networks, genetic algorithms, learning algorithms, improvement in signal to noise ratio (isnr), normalized mean square error (nmse)

Restoration of images degraded by unknown blur is a difficult problem. It is called blind image restoration. The unknown blur can be of linear or nonlinear nature. It can also be space invariant or space variant making the blind image restoration problem all the more challenging. Sometimes it is necessary to simultaneously identify the blur and restore the image, a problem that we have addressed in this thesis. The assumption is that coefficients of image model are controlled by autoregressive(AR) process while that of blurring model are controlled by moving average (MA) process. Moreover, a natural assumption about most of the images in practical applications is that they are smooth in nature.

We have used three layered artificial neural networks (ANN) to embed naturally the AR process between its first two layers and MA process between its last two layers. The genetic algorithms (GA) have been used to avoid getting stuck in local minima. The first major work has been the extension of the network to handle nonlinear space-invariant degradations in the images by incorporating nonlinear ARMA model using the concept of Volterra filters. This approach can cater for the sharp contrasts which may come in the degraded images as well. The second major work has been the adaptation of the ANN to handle the space variant blur. The image and hence forth, the layers of ANN are divided into blocks and each block is categorized according to the level of activity. Thus the weights between the layers are no more universal and space invariant but they can vary from block to block according to the activity. This kind of freedom results in better results in case of space variant blurs and even space invariant blurs. The third major work has been the extension of the cost function. The two extra terms in cost function take into account the human visual perception system. They match the second order statistics (local variances) of the images at different layers and give improvement in the visual quality of the images.

Gradient based learning algorithms have been developed for the fast convergence to the solution. Several degraded images have been restored. The results have been compared with some of the important current techniques in the literature using improvement in signal to noise ratio (ISNR) of restored images and normalized mean square error (NMSE) of estimated blur as figures of merit.

Download Full Thesis
1750.26 KB
S. No. Chapter Title of the Chapters Page Size (KB)
1 0 Contents
157.7 KB
2 1 Introduction 1
59.75 KB
  1.1 Statement Of Problem 1
  1.2 Contributions Of The Dissertation 1
  1.3 Organization Of The Dissertation 5
3 2 Digital Image Restoration 8
268.44 KB
  2.1 Image Formation 8
  2.2 Blur Models 8
  2.3 Types Of Noises In Images 19
  2.4 Images Restoration Methods 21
4 3 Hybird Computing Tools- Artificial Neural Network And Genetic Algroithms 39
106 KB
  3.1 Feedforward Multiplayer Perception 39
  3.2 Genetic Algorithms 46
5 4 Blind Deconvolution Of Linearly Degraded Images Using Neural Networks 52
407.83 KB
  4.1 Image And Degradation Models 52
  4.2 Blind Image Deconvolution Of Linearly Degraded Images 21
6 5 Blind Deconvolution Of Non-Linearly Degraded Images Using Neural Networks 74
375.16 KB
  5.1 Neural Networks For Blind Deconvolution Of Nonlinearly Degraded Images 75
  5.2 Neural Networks For Blind Disconsolation Of Nonlinearly Degraded Images Using Nonlinear Arma Model 83
7 6 A New Space-Variant Neural Network Approach To Blind Image Deconvolution 91
435.69 KB
  6.1 Image And Degradation Models 93
  6.2 Neural Networks For Blurred Image Representation 94
  6.3 Simulation Studies 104
8 7 Conclusion 114
35.63 KB
  7.1 Summary Of Results 114
  7.2 Futher Directions 116
9 8 Reference 118
86.84 KB