Image Reconstruction by Filtered, Stochastic and Transform Based Techniques using Parallel Beam Transmission Tomography

Abstract

The image reconstruction by parallel ray transmission tomography has been carried out by using filtered, stochastic and transform-based techniques. Image quality has been analyzed by various error measures using Filtered Back-Projection (FBP) technique with various composite filters, namely Shepp-Logan, Cosine and Hamming filters. The effect of projections on filtered-based image reconstruction has also been thoroughly investigated. A stochastic-based, Hybrid Simulated Annealing (HSA) algorithm has been proposed for the image reconstruction using lesser number of projections or when the data is missing or corrupt. Various parameters have been tuned including generic parameters, like initial and final temperatures, annealing schedule, slab-thickness with constant temperature, etc. along with problem specific parameters, like size and structure complexities, number of projections, etc. The image quality has been compared with other techniques, namely FBP and Algebraic Reconstruction Technique (ART). Another stochastic algorithm, the Hybrid Continuous Genetic Algorithm (HCGA), has also been introduced for image reconstruction. It has been used with preprocessed image data or even the actual human data for image reconstruction. The templates used were of two types: the FBP-based and the multiscale Wavelet Transform based. Various operators, namely Mixed Selection Scheme, Image Row Crossover operator, Offset Based Mutation and Hybrid Diversification operators have been introduced. The sensitivity analysis of various selection schemes, like truncation, roulette wheel, tournament selection, etc., crossover schemes, like single point, multi point, uniform crossover schemes, etc., and mutation schemes, namely standard- and gradient-based schemes, have been investigated in depth for tomographic image reconstruction. The number of projections has been optimized for HCGA reconstruction to p = 30. The HCGA-based image reconstruction for 8 × 8 to 128 × 128 sized images have been found to be better than FBP and ART techniques, and even better relatively for larger image sizes than the HSA. The fitness function analysis has also been carried out for various functions like Root Mean Squared error (RMSE), Mean Squared error (MSE), Mean Absolute error (MAE), Relative Squared error (RSE), Root Relative Squared error (RRSE), etc. The MAE and RMSE have been generally found to converge relatively in shorter times in comparison to the rest of the functions under consideration. The HDO introduction in the Evolution cycle, using the SA and the decreasing law of mutation probability, culminated to relatively more promising results. The essence of this thesis is the relative comparison of various stochastic techniques by using filtered and transform-based techniques for the first time in the field of parallel ray transmission tomography. The work is exhaustive, multi-dimensional with conclusive results.