20.500.12556/DKUM-20518 Vgnezdene izbočene lupine kot značilnice za zaznavo dvojnikov digitalnih slik in video posnetkov Convex layers as a feature for near-duplicate detection of digital images and video clips V doktorski disertaciji obravnavamo problematiko iskanja podobnih slik in video posnetkov. Iz slike tvorimo geometrijsko značilnico: vgnezdeno izbočeno lupino, ki jo nato pretvorimo še v vektorsko značilnico. V disertaciji opišemo dva postopka za tvorbo značilnice slike. V prvem tvorimo vgnezdene izbočene lupine iz binarne slike. V drugem pa iz sivin slike tvorimo zaporedje binarnih slik. Vgnezdene izbočene lupine tvorimo na vseh binarnih slikah iz zaporedja. Vse točke iz slike se nahajajo na rastrski mreži. Za te točke smo razvili postopek tvorbe vgnezdenih izbočenih lupin, ki je linearne časovne zahtevnosti, odvisen le od števila pikslov na sliki. Za vektorske značilnice smo razvili nov postopek določanja razdalje oziroma podobnosti. Razdaljo imenujemo razdalja relacij sosedstva. Iz vektorske značilnice izračunamo štiri koeficiente, ki opisujejo monotonost funkcije. Razdalja relacij sosedstva je funkcija razlik in vsote štirih koeficientov značilnice. V disertaciji opišemo tudi razširitev značilnice na zaporedje binarnih slik in razdalje relacij sosedstva za problem iskanja dvojnikov video posnetkov. S preizkusom učinkovitosti na slikah smo ocenili kvaliteto razvitih algoritmov. Pri učinkovitosti zaznave dvojnikov se najbolje izkaže značilnica zaporednih binarnih slik v kombinaciji s korelacijskim koeficientom. Obratno, najboljše časovne rezultate dosega algoritem tvorbe značilnice iz binarne slike, vendar algoritem tvorbe značilnice iz zaporednih binarnih slik ne zaostaja mnogo. Tako pravilnostna kot časovna učinkovitost sta zelo dobri tudi pri razvitem postopku iskanja dvojnikov video posnetkov. Doctoral dissertation deals with the searching of near-duplicate images and video clips. From the picture, we form a geometric feature; convex layers, which are later converted into a feature vector. The thesis derives two algorithms for extraction the feature vector from images. In the first algorithm, convex layers are extracted from a binary image, while in the second one the feature is extracted from a sequence of binary images. In this case, the convex layers are constructed for all images in the sequence. Pixels of images are considered located on the raster. For raster points we developed an algorithm to construct convex layers in the linear time, thus its time complexity depends only on the resolution of images. For feature vectors, we developed new measure of similarity. It measures the distance called neighbourhood relations distance. From the feature vector, four coefficients are calculated. The coefficients describe the monotony of the vector function. The local area neighbourhood distance is a function of differences and sums of four coefficients. The thesis also describes the extension of serial binary images features algorithm and local area neighbourhood distance into the area of near-duplicate video clip detection. The effectiveness test of the developed algorithms was performed. The image duplicate detection efficiency is highest in serial binary images features algorithm in combination with the correlation coefficient. Lowest computational complexity was obtained by the algorithm based on binary image feature, but the algorithm of serial binary features does not lag much. Also, performance and time efficiency are also very good for the developed algorithm when detecting duplicated videos. digitalna obdelava slik digitalna obdelava video posnetkov zaznava dvojnikov mera podobnosti računalniška geometrija vgnezdene izbočene lupine tvorba značilnice podobnost značilnic digital image processing digital video processing near–duplicate detection similarity measure computational geometry convex layers feature extraction feature similarity true true false [S. Šinjur] Slovenski jezik Angleški jezik Doktorska disertacija 2011-09-21 12:38:34 2011-09-22 11:00:08 2022-04-14 07:59:57 2023-12-27 03:21:19 2011 0 Maribor 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.93:004.98(043.3) 15331094 URN:SI:UM:DK:65X9V5ZC DR_Sinjur_Smiljan_2011.pdf DR_Sinjur_Smiljan_2011.pdf 1 6BE9D7C0B1293D515F9474395089EFDD 8e2c687a991b8cc1e1595e3840c0ac581348f917c22c79ad0bcd26f52a69e887 a03cb88d-7c05-11eb-bb7a-00155d0001ca 20.500.12556/dkum/945ff20a-028b-48b6-a591-5fa9beb628c0 https://dk.um.si/Dokument.php?lang=slv&id=25473 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0