Naslov
Segmentacija i klasifikacija neprozirnih i poluprozirnih regija na slikama prirodnog krajolika
(Segmentation and classification of non-transparent and semi-transparent regions on natural landscape images)

Autori
Braović, Maja

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, doktorska disertacija

Fakultet
Fakultet elektrotehnike, strojarstva i brodogradnje u Splitu

Mjesto
Split

Datum
09.12

Godina
2015

Stranica
135

Mentor
Stipaničev, Darko

Ključne riječi
klasifikacija slika mediteranskog krajolika; prostori boje; lokalne i globalne znacˇajke slike
(Keywords: Mediterranean landscape image classification; color spaces; local and global image features.)

Sažetak
This doctoral dissertation presents research on automatic segmentation and clas- sification of regions (that can be as small as 1 pixel) that can be found in images of natural Mediterranean landscape. This research encompasses the following: the search for optimal color spaces, optimal local image features and optimal global image features for the classi- fication of individual Mediterranean landscape classes, proposition of a novel classifier for the classification of Mediterranean landscape images, proposition of the FESB Mediterranean Landscape Image Dataset (FESB MLID), and the search for the changes that smoke causes when it appears in the image and that could possibly improve automatic smoke detection in video sequences. This doctoral dissertation consists of a number of steps. First, it reviews the theories that seek to explain human visual perception, and it also gives a review of methods used in artificial visual perception (such as the analysis of color, texture, shape, and local and global features of the image). Second, it presents the FESB MLID dataset. This dataset consists of 400 images of natural Mediterranean landscape and another 400 images of their manual segmentations, and it is also available on the Internet so that it could be of use to other researchers that deal with the automatic segmentation and classification of images. Third, an evaluation of different color spaces and the selection of those optimal for the classification of predefined classes that can be found in the natural Mediterranean landscape is presented. Fourth, the optimal local and global image features for the classification of predefined classes that can be found in natural Mediterranean landscape are determined. Fifth, a proposition of a novel classifier that can be used for the classification of natural Mediterranean landscape images is given and compared with different classifiers often used in image processing. Although the classification results obtained by its use are sometimes better than those obtained for other classifiers, sometimes approximately the same and sometimes worse, its execution time is always lower. Sixth, an analysis revolving around the impact of the appearance of smoke on the Mediterranean land- scape image background is given, and the changes that smoke causes when it appears in the image are discovered. These changes are given as a set of rules that could help in the more accurate detection of smoke in video sequences. Finally, a conclusion that briefly discusses the contributions of this doctoral dissertation and possible further research is given.

Izvorni jezik
Hrvatski

Znanstvena područja
Računarstvo

POVEZANOST RADA