doktorska disertacija
Simon Gangl (Author), Borut Žalik (Mentor)

Abstract

V doktorski disertaciji obravnavamo algoritem stiskanja domenskih zaporedij slik. Pojem domenskega zaporedja slik pri tem predstavlja opis za vsebinsko povezana, urejena zaporedja slik, ki opisujejo bodisi časovni bodisi prostorski potek spremembe poljubne domene. Teoretični opis v praksi združuje dve, za naš algoritem, sorodni nalogi: stiskanje časovnih zaporedij slik, torej videov, ter stiskanje prostorskih zaporedij slik, na primer naborov medicinskih slik, zajetih s tehnologijo CT ali MRI. V disertaciji opišemo strukturo in delovanje algoritma, ki omenjeni problem rešuje s projekcijo v prostor osnovnih komponent. Najprej predstavimo matematično ozadje, ki je osnova za, v statistiki pogosto uporabljeno, metodo analize osnovnih komponent. Prav ta je izhodišče za izračun projekcijskih prostorov, v katerih je možno predstaviti slike dane domene, pri čemer ni pomembno, za kakšno vrsto zaporedja slik gre. Da bi razširili neodvisnost od domene, ki jo zagotavlja izhodiščna matematična metoda, na nivo algoritma stiskanja, je prvi pomemben korak izbira podzaporedja slik, ki so osnova za izračun projekcijskih prostorov. Za to nalogo uporabimo dvokriterijski algoritem, ki izbira slike - imenujemo jih bazne slike - glede na medsebojno odstopanje in oddaljenost v vhodnem zaporedju. Iz izbranega zaporedja baznih slik določimo zaporedje projekcijskih prostorov glede na v disertaciji uveden koncept, po katerem sledeče si projekcijske prostore določamo na osnovi množic baznih slik, ki imajo vsaj en skupni element. Kot analogijo konceptu drsečega okna tak pristop opišemo kot ''drseč lasten prostor''. Vzporedno uvedemo način izračuna projekcijskih prostorov, ki omogoča kasnejšo rekonstrukcijo vhodnih podatkov z bistveno manjšim računskim bremenom. To dosežemo z vključitvijo vmesnih računskih rezultatov v stisnjeno predstavitev podatkov, pri čemer je vpliv na stopnjo stiskanja zanemarljiv. V eksperimentalni analizi podamo primerjavo med razvitim algoritmom, do sedaj najpogosteje uporabljeno metodo s projekcijo v prostor osnovnih komponent, in standardom H.264. Tako dokažemo, da algoritem po vizualni kakovosti ne presega le prejšnje metode, ampak se v njej, kakor tudi v stopnji stiskanja, lahko primerja celo s H.264. Rezultate eksperimentov nadalje potrdimo v teoretični analizi, kjer formalno dokažemo prednosti razvitega algoritma in ocenimo tudi vpliv kontrolnih parametrov metode na učinkovitost stiskanja.

Keywords

algoritmi;izgubno stiskanje;stiskanje zaporednih slik;stiskanje videoposnetkov;slike DICOM;PCA;analiza osnovnih komponent;lasten prostor;projekcija v lasten prostor;lasten vektor;

Data

Language: Slovenian
Year of publishing:
Typology: 2.08 - Doctoral Dissertation
Organization: UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher: [S. Gangl]
UDC: 004.932:004.627(043.3)
COBISS: 17863190 Link will open in a new window
Views: 1779
Downloads: 153
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: DOMAIN-BOUND IMAGE SEQUENCE COMPRESSION ALGORITHM BASED ON PROJECTION TO PRINCIPAL COMPONENT SPACE
Secondary abstract: The scope of this doctoral dissertation is an algorithm for the compression of domain-bound image sequences. The notion of a domain-bound image sequence is used as a description for ordered image sequences of a linked content, which represent either the temporal, or the spatial course of change of an arbitrary domain. The theoretical description translates in practice, within the scope of our algorithm, to two similar tasks: the compression of temporal image sequences, thus videos, and the compression of spatial image sequences, for example medical image sets acquired by CT or MRI technologies. In this dissertation we describe the structure and functioning of the algorithm, which tackles the mentioned problem by projection to principal component space. Initially we introduce the mathematical background, which is the foundation for principal component analysis, a method frequently used in statistics. It is this method, which is the origin for the computation of projection spaces that can be used to represent images from a given domain, whereby the type of the image sequence is irrelevant. In order to extend the domain independence, provided by the underlying mathematical method, to the scope of the compression algorithm, selecting the subsequence of images that are the origin for the projection space computation is the first important step. For this task we use a double-criterion algorithm, which selects the images - we name them base images - based on their mutual deviation and distance within the input sequence. A projection space sequence is computed from the selected sequence of base images using a concept introduced in this dissertation, which defines that adjacent projection spaces are computed from overlapping sets of base images, which have thus at least one element in common. In analogy to the sliding window concept we describe this approach as a ''sliding eigenspace''. In parallel we introduce a method of computing projection spaces that allows for a later reconstruction of the input data at a significantly reduced computational cost. This is achieved by including intermediary computational results into the compressed data representation, while the influence on the compression ratio is insignificant. In the experimental analysis we provide a comparison between the developed algorithm, the previously generally used method based on projection to principal component space, and the H.264 standard. Thereby we prove that in terms of visual quality the algorithm does not only outperform the previously used method, but is actually capable of competing in it, as well as in compression ratio with H.264. The experimental results are further confirmed by a theoretical analysis, where we formally prove the advantages of the developed algorithm and evaluate the impact of the methods control parameters on compression efficiency.
Secondary keywords: algorithms;lossy compression;image sequence compression;video compression;DICOM images;PCA;principal component analysis;eigenspace projection;eigenvector;Slike;Disertacije;Algoritmi stiskanja;
URN: URN:SI:UM:
Type (COBISS): Doctoral dissertation
Thesis comment: Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko
Pages: XII, 111 str.
ID: 8729439