A data dependent systems approach to image deblurring and edge detection
Mo Liu · 1991
The dissertation presents a Data Dependent Systems (DDS) approach to image processing. It develops a new DDS deconvolution methodology that provides solutions for boundary detection, image enhancement and image restoration. A new edge enhancement technique has also been developed for the deconvolution of non-causal image blur by two dimensional Fourier transform. This development complements the DDS deconvolution which is limited to causal systems. New image enhancement and curve detection techniques based on an AR(3,3) model were developed. The image enhancement was achieved by edge enhancement. A single parameter $\phi$ which controls the enhancement can be easily adjusted. The curve detection was performed by evaluating the differences of the adjacent pixels. The newly developed curve detection technique can reduce noise. A uniform motion blurred TV image was restored by DDS deconvolution. The newly developed DDS deconvolution decomposes data into the Green's function and residuals. The DDS deconvolution not only allows flexible modeling constraints but also allows each problem to be solved in its unique way when some prior knowledge is available. Although DDS deconvolution employs the modeling strategy of DDS, its modeling constraints vary according to the individual problem. The residuals of the system are not restricted to random noise, they are the information remaining in the data after the Green's function has been extracted. The goal of the research was to develop practical methods to solve image processing problems. A new deblurring technique was developed for the restoration of Gaussian blurred images. It was demonstrated that the Fourier transform deblurring method cannot work effectively due to the loss of the edge structure of the blurred image. The new deblurring technique provides a method of reconstructing the edges so that the image can be restored by two-dimensional Fourier transform. The method can also be applied to kernels other than Gaussian as long as the type of the deblurring kernel is known. The effects of overdeblurring and underdeblurring have been investigated. A criterion has been established so that a closed loop computer program can be developed for deblurring.