Novel Filtering Methods for Image and Video Processing Applications

Abstract

During the last few years, digital filtering methods for image/video processing applications have reached a satisfactory level. However, their performance degrades in the presence of noise, trend, motion, shape deformation, intensity inhomogeneity, shadows, or low image quality, to name a few. To cope with these challenges, this dissertation presents novel filtering methods for image/video processing applications that outperform the existing and state-of-the-art methods. The dissertation starts by introducing a novel trend filtering method that transforms the inter-frame registration problem into low complexity trend filtering problem. In the proposed method, Laplacian eigenmaps in conjunction with the modified empirical mode decomposition has been used to suppress the noise artifacts and the trend term. In multi-dimensional signals, the trend term is often referred to as non-uniform illumination or global intensity inhomogeneity. This dissertation presents a new filtering method for estimating the global intensity inhomogeneity in two dimensional and volume images. Global intensity inhomogeneity often arises due to the imperfections of data acquisition device, direction of source light, and properties of the subject under study. The proposed method generates a high-pass filter based on the grey-weighted distance transform of the frequency content of an image/volume. It provides an accurate estimation of global intensity inhomogeneity without any parameter tweaking, which makes it applicable to many imaging modalities. The dissertation also presents a filtering methodology to cope with local intensity inhomogeneity that gives rise to shadow artifacts. These artifacts appear as sharp discontinuities and are often corrected at different scales and orientations. The proposed method makes use of decimation-free directional filter bank to suppress the local intensity inhomogeneity and shadow artifacts irrespective of scale and orientation. In addition to intensity inhomogeneity correction, the dissertation also presents a filtering method that utilizes the Gabor filter bank to generate rotation invariant feature codes. The effectiveness of the proposed method has been evaluated in both identification and verification modes for fingerprint recognition. The uniqueness of the presented filtering methods lies in the fact that they are essentially parameter free and can easily be scaled to higher dimensions. This makes them applicable to many different image/video processing applications with least of effort from the end user, i.e., eliminating the user biases.