Image sequence restoration by median filtering

Median filters are non-linear filters that fit in the generic category of order-statistic filters. Median filters are widely used for reducing random defects, commonly characterized by impulse or salt and pepper noise in a single image. Motion estimation is the process of estimating the displacement vector between like pixels in the current frame and the reference frame. When dealing with a motion sequence, the motion vectors are the key for operating on corresponding pixels in several frames. This work explores the use of various motion estimation algorithms in combination with various median filter algorithms to provide noise suppression. The results are compared using two sets of metrics: performance-based and objective image quality-based. These results are used to determine the best motion estimation / median filter combination for image sequence restoration. The primary goals of this work are to implement a motion estimation and median filter algorithm in hardware and develop and benchmark a flexible software alternative restoration process. There are two unique median filter algorithms to this work. The first filter is a modification to a single frame adaptive median filter. The modification applied motion compensation and temporal concepts. The other is an adaptive extension to the multi-level (ML3D) filter, called adaptive multi-level (AML3D) filter. The extension provides adaptable filter window sizes to the multiple filter sets that comprise the ML3D filter. The adaptive median filter is capable of filtering an image in 26.88 seconds per frame and results in a PSNR improvement of 5.452dB. The AML3D is capable of filtering an image in 14.73 seconds per frame and results in a PSNR improvement of 6.273dB. The AML3D is a suitable alternative to the other median filters