High-performance systems for robust image enhancement and restoration

The following dissertation addresses the problem of extracting useful information from images which may have been degraded by noise. The proposed solutions encompass and integrate two areas of research in Electrical Engineering: signal processing algorithm design and VLSI architecture. The image enhancement and restoration algorithms devised are robust and are capable of suppressing noise with minimal distortion. These algorithms are suitable for hardware implementation. Parallel VLSI architectures are developed to perform the enhancement/restoration task at video-rate.;The image enhancement filters (LUM filters) originate from the theory of order statistics. While maintaining noise rejection capability similar to the median, the LUM filters blur the processed image much less. The filter incorporates a tuning parameter through which its nonlinear characteristic can be adjusted. An adaptive scheme is devised for adjusting the filter's parameter dynamically based on the local image characteristics. The resultant space-variant filter is quite effective for removal of impulsive noise while maintaining the signal features intact.;Implementation of the LUM filtering algorithm is based on sorting all the sample points in a 2-D window and multistage order statistic operations on the sorted samples. The computational steps are similar to those used in many 2-D generalized order statistic filters; nevertheless, the majority of VLSI design suggestions are limited to 1-D and single order statistic finding cases. To alleviate these gaps, a recursive class of VLSI architectures for generalized multi-dimensional order statistic filtering is developed. The architectural concepts are exploited for efficient hardware implementation of the LUM filters.;For the LUM filtering algorithm, only the nonlinear characteristic can be adjusted. In order to produce detail-preserving filters with parameters for adjusting both linear and nonlinear characteristics independently, a new class of restoration algorithms is introduced. The filters of this class incorporate a state-space model and originate from the theory of robust regression. Through the state-space model which tracks the dynamics of an image, a parameter is incorporated to adjust the linear behavior. The robust loss function used to solve the regression provides a robustness parameter that limits the influence of gross errors and can be used to adjust the nonlinear behavior. The filters introduced are capable of removing Gaussian as well as impulsive noise from the images while preserving fine image details.;The proposed class of restoration algorithms is suitable for massively parallel implementation; VLSI algorithms for hardware implementation are developed.