Razpoznavanje dreves s pomočjo računalniškega vida

ŠVAB, MATIC (author), Kristan, Matej (mentor)

V diplomski nalogi je predstavljen postopek razpoznavanja dreves s pomočjo računalniškega vida, ki analizira drevesno lubje. Iz posameznih slik lubja postopek izlušči značilnice LBP, ki jih SVM uporabi za učenje in testiranje. Ker prosto dostopna zbirka slik drevesnega lubja ne obstaja, je bilo potrebno zajeti večjo anotirano zbirko slik lubja, ki je tudi prva javno dostopna zbirka. Pri razpoznavanju se pojavi še problem določitev skale, saj različne naprave zajamejo slike različnih velikosti, različnih razmerjih širine/višine slike predvsem pa ljudje ne slikajo enako oddaljeni od dreves. V diplomski nalogi je tudi predlagan postopek, ki s pomočjo značilnic, pridobljenih s detektorjem DoG, samodejno določi skalo slike, s katero se vhodna slika pred izračunom LBP-ja vedno preskalira v referenčno velikost in s tem teži k normalizirani velikosti pomembnih struktur v slikah. Končni eksperiment je na zbirki 12 dreves dosegel 84.62 % natančnost.

• Language: Slovenian
• Keywords: lokalni binarni vzorci, metoda podpornih vektorjev, klasifikacija dreves, samodejno določanje skale
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2014
• PID: 20.500.12556/RUL-29559
• Publication date in RUL: 23.09.2014

Secondary language

• Language: English
• Title: Computer-vision-based tree trunk recognition

Abstract

This thesis presents a process of a tree recognition by means of the computer vision, which analyses the tree bark. The procedure extracts LBP features from individual pictures of bark, which are used for training and testing by SVM. Since freely accessible collection of tree bark pictures does not exist, it was necessary to create a larger annotated collection which is also the first database publicly available. In recognition there is also a problem with scale or picture size, because different devices take pictures of different sizes, in different width/height proportions and mostly people do not take photographs from the same distance. The thesis also proposes a procedure that by means of the features gained by DoG detector, automatically determines the picture scale, by means of which the input picture is always rescaled in the reference size before the calculation of LBP. In the final experiment the 84.62 % accuracy was achieved on the collection of 12 trees.

• Keywords: local binary patterns, support vector machine, tree classification, automatic scale determination