The tv-l(1) model: theory, computation, and applications

摘要: In digital image processing, images are treated as two dimensional signals and processed using signal processing techniques. This thesis focuses on a decomposition technique that partitions an input into components with different geometric scales. multiscale is based the minimization of energy weighted sum total variation L1 distance (TV-L 1). Although L1-fidelity term penalizes intensity difference between output images, by TV-L1 model solely determined scales features (homogeneous parts) in image, independent feature intensity. important property was previously observed Alliney, Nikolova, Chan Esedoglu. Here we present Chapter 2 rigorous proof this newly established equivalence minimizations sequence energies. Furthermore, use to obtain other properties, including morphological invariance, TV-L 1 discuss new computational methods. To reveal differences TV-based models for decomposition, related ROF, Meyer, Vese-Osher formulate all four second-order cone programs (SOCPs). We used SOCP interior-point method, combined nested dissection reduce complexity, numerically solve SOCPs corresponding these obtained solutions both 1D 2D examples. Our tests show TV -L1 qualitatively from those produced and, therefore, very unique. Because its special properties shown analytically numerically, quickly finds applications medical imaging computer vision. first application, cDNA microarray inhomogeneous background have large-scale separated useful foreground spots. second multiplicative version applied face varying illumination conditions. The resulting quotient successfully recognition.