Mobilna aplikacija za razpoznavanje objektov iz več pogledov

Račič, Matej (author), Kristan, Matej (mentor)

Diplomsko delo opisuje problem klasifikacije objektov na sliki s pomočjo mobilne aplikacije. Podrobneje smo preizkusili implementacijo Inception, ki z uporabo okolja Tensorflow \cite{tensor} omogoča enostavno spremljanje in optimizacijo učenega procesa klasifikatorja. Zajeli smo specialno zbirko za realistično objektivno analizo delovanja aplikacije. Predstavili smo dobre prakse v procesu učenja in vpliv parametrov, kot sta število iteracij in učnih primerov. Vso znanje smo uspešno uporabili in klasifikator prilagodili razpoznavanju na izbrani domeni za doseganje želenih rezultatov in uspešno klasifikacijo objektov iz različnih pogledov. S pomočjo okolja Android Studio smo tudi generirani razpoznavalnik prenesli na mobilno napravo.

• Language: Slovenian
• Keywords: mobilne naprave, razpoznavanje objektov, nevronske mreže
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-95928
• Publication date in RUL: 25.09.2017

RAČIČ, Matej, 2017, Mobilna aplikacija za razpoznavanje objektov iz več pogledov [online]. Bachelor’s thesis. [Accessed 22 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=95928

Secondary language

• Language: English
• Title: A mobile application for multiview object recognition

Abstract

The thesis deals with the problem of multiview object recognition with a mobile application. To achieve this, we have used the neural network architecture called Inception, that uses the Tensorflow environment which enables simple modifications of the learning process. We used a specific collection of images, to help us objectively and realistically evaluate how well the object recognition works. We have presented good practices and we have applied that knowledge in the development of our object recognition. We have also examined various effects of different parameters such as number of iterations and training samples. We have combined it all and modified the model for recognising our own selected categories and maximising the results of the object recognition. Using Android Studio we have transferred the generated model to a mobile device.

• Keywords: mobile devices, object recognition, neural networks