|
Thesis
/ | ROMDOC-THESIS-2017-1003 |
Mining image symbols and semantics
Văduva, Corina
2011-10-10
Abstract: Mining Image Symbols and Semantics Ing. Corina Văduva Abstract In contextul analizei imaginilor de teledetectie de foarte mare rezolutie, aceasta teza introduce un concept nou pentru anotarea unor scene mari cu etichete ce definesc intelesul real al structurilor de pe suprafata terestra. Se doreste inlaturarea diferentelor de semantica dintre informatia continuta de date si intelesul pe care aceasta il are pentru utilizatorul uman. Metodologia descrisa isi propune sa extraga in mod automat informatie cu inteles semantic ridicat ce poate inlocui analiza manuala a utilizatorului uman in vederea intelegerii si caracterizarii continutului unei imagini satelitare. Noutatea metodologiei abordate in aceasta directie consta intr-o analiza ierarhica a continutului imaginii: pixelii sunt grupati in obiecte, iar obiectele in configuratii a caror analiza duce la definirea unui inteles la nivel de scena. Problema principala este reprezentata de complexitatea structurilor de pe suprafata terestra si a varietatii trasaturilor caracteristice. Pentru a imbunatati descrierea continutului imaginilor, a fost dezvoltat un atribut ce ilustreaza pozitia relativa a obiectelor din scena. Acest atribut este un vector-eticheta spatiala invariant (la rotatie, translatie si scalare) ce include pe langa topologia obiectelor analizate si informatii despre forma, dimensiunea si distanta dintre ele. Metodologia descrisa urmareste un lant de prelucrare a imaginii pe trei nivele: pixel, obiect, scena. Analiza incepe prin extragerea obiectelor in functie de trasaturile lor primitive folosing metode de clasificare si segmentare. Apoi sunt imbinate metode de modelare pentru text cu atributele spatiale pentru a descrie aranjamentul obiectelor in imagine. Clasificarea k-means este folosita pentru a defini paternuri de perechi similare de obiecte. Aceste paternuri sunt considerate cuvinte vizuale si sunt folosite pentru modelarea scenei cu ajutorul modelului Latent Dirichlet Allocation. Clasele semantice obtinute contin perechi de obiecte caracterizate de aceeasi informatie spectrala si au aranjament spatial similar. O proprietate specifica a metodei propuse consta in faptul ca are la baza doua abordari diferite privind procesarea si interpretarea continutului imaginilor. Prima se refera la clasificarea perechilor de obiecte, atribuind etichete cu inteles semantic scenei, iar a doua descrie un algoritm de regasire de informatie semantica in baze de imagini. Acest algoritm poate descoperi in mod latent clase semantice ce contin perechi de obiecte cu un anume aranjament spatial ce nu pot fi usor observate de catre utilizatorul uman. Eficienta metodologiei propuse si varietatea applicatiilor posibile au fost ilustrate prin diverse studii de caz si exemple. In the frame of remote sensing very high resolution imagery, this thesis introduces a new concept for the annotation of large scenes with labels expressing the real meaning of the land cover which aims to fill the semantic gap between the information contained by the data and the user’s knowledge about the scene. The methodology proposed supersedes human inductive learning and reasoning in complex scene understanding and characterization by automatic adding high-level ontology to the image. The innovation in this direction of data mining consists in a hierarchical analysis of remote sensing image content: pixels are grouped into objects according to their primitive features and objects are grouped into configurations. The analysis regarding the similarity of object groups confers semantic meaning to the image content at the scene level. The problem consists though in the complexity of the land cover structures and also in the variety of characterizing features. In order to improve the image content description, an attribute is developed to portray the relative positioning of objects to each other and also in the scene. The proposed attribute is an invariant (to rotation, translation and scaling) signature which includes the topology, as well as the information about the shape, the size and the distance between the analyzed objects. The methodology described pursues a three levels image processing chain: pixel-level, object-level and scene-level. Objects are firstly extracted based on their primitive features using classification and segmentation methods. Further, text modelling methods and the spatial attributes are combined in order to describe objects arrangement inside the scene. Through a k-means classification, new patterns of similar objects configurations could be defined. In the end, the scene is modelled according to these new patterns (visual words) using the Latent Dirichlet Allocation model into a finite mixture over an underlying set of semantic classes containing configurations of objects characterized by the same spectral and spatial information. A specific of the proposed methodology refers to the fact that it is based on two approaches regarding the image content processing and interpretation. First, it refers to a classification of pairs of objects assigning semantic labels to the scene. Second, the algorithm allows for the retrieval of semantic information inside an image database. It is also able to discover latent semantic classes containing pairs of objects characterized by a certain spatial positioning that are difficult to see for the human eye due to the complexity of the scene. Several case studies prove the efficiency and the various applications of the proposed concept and algorithm.
Keyword(s): Prelucrarea imaginii -- Topografie -- Teză de doctorat ; Teledetecţie -- Teză de doctorat ; Data mining -- Teză de doctorat
OPAC: See record in BC-UPB Web OPAC
Full Text: see files
Record created 2017-02-24, last modified 2017-02-28
Similar records
| People who viewed this page also viewed: |
|
|
Be the first to review this document.
Start a discussion about any aspect of this document.
guest ::