diplomsko delo
Oskar Čokl (Author), Luka Čehovin (Mentor)

Abstract

Z digitalizacijo zajema slik se je njihova koliˇcina izjemno poveˇcala, obenem pa je iskanje po zbirkah slik postalo zelo teˇzavno. V diplomi obravnavamo problem poizvedovanja po zbirki slik na podlagi referenˇcnega primera. Upo- rabljamo sodoben pristop na podlagi opisnikov, pridobljenih iz globokega modela na osnovi konvolucijske nevronske mreˇze. Taki opisniki niso redki in nad njimi ne moremo zgraditi obrnjenih indeksov. V naˇsem pristopu se za poizvedovanje posluˇzimo hierarhiˇcnega gruˇcenja z drevesom na podlagi pogojnih verjetnosti. Okoli drevesne strukture zgradimo prototip storitve za poizvedovanje po zbirki slik, ki lahko odzivno streˇze veˇc uporabnikov hkrati. Reˇsitev ovrednotimo napram poizvedovanju s surovo silo in ugotovimo, da je predlagani pristop bolj ustrezen pri veliki koliˇcini slik, saj porabi manj pomnilnika in manj ˇcasa za poizvedovanje.

Keywords

poizvedovanje;slike;poizvedovanje s primerom;skalabilnost;CD-drevo;vrsta opravil;univerzitetni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [O. Čokl]
UDC: 004(043.2)
COBISS: 82086659 Link will open in a new window
Views: 158
Downloads: 20
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: Efficient content-based image retrieval
Secondary abstract: With the digitalization of capturing images, the amount of images drasti- cally increased and searching through a collection of images became very hard. This dissertation deals with querying a collection of images based on a reference image. We use a modern approach based on features obtained from a deep model based on a convolutional neural network. Such features are not sparse and we cannot build inverted indexes with them. In our approach we use hierarchical clustering with a conditional density tree for querying. We build a prototype of an image search service with the tree structure which is able to responsively serve multiple users at the same time. We test the solution against a brute force approach and find that the suggested method is more suited for large collections of images, as it consumes less memory and needs less time for queries.
Secondary keywords: querying;images;query by example;scalability;CD-tree;task queue;computer and information science;diploma;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000468
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 38 str.
ID: 13682514