Određivanje okvira objekata na slikama s pomoću dubokog učenja

undergraduate thesis

• Title (english): Bounding Box Detection Using Deep Learning
• Author: Dorjan Štrbac
• Mentor: Krešimir Križanović (mentor)
• Committee member: Krešimir Križanović (predsjednik povjerenstva)
• Committee member: Josipa Lipovac (član povjerenstva)
• Committee member: Rafael Josip Penić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-09-11, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Objašnjene su osnove strojnog i dubokog učenja, unaprijednih potpuno povezanih neuronskih mreža i konvolucijskih neuronskih mreža za računalni vid. U detalje se objašnjavaju potpuno povezani slojevi te 2D konvolucijski slojevi i 2D slojevi združivanja i njihove uloge u konvolucijskim neuronskim mrežama za računalni vid. Objašnjen je koncept prijenosnog učenja u računalnom vidu. Predstavljen je podatkovni skup korišten za rješavanje zadatka uokvirivanja objekta na slici u što uži okvir sa stranicama paralelnim s vertikalnom i horizontalnom osi slike. Objašnjena je IoU metrika te neke funkcije gubitka i ostale korištene metrike temeljene na IoU. Opisana je arhitektura baza VGG16 i ResNet50 mreža i modeli s tim pretreniranim bazama koriste se za prijenosno učenje na prethodno spomenutom zadatku te se uči i relativno jednostavna konvolucijska mreža i prikazuju i uspoređuju rezultati. Također se objašnjavaju neki problemi dubokih neuronskih mreža i njihova rješenja.

Abstract (english)

The basics of machine learning, deep learning, feedforward fully connected neural networks, and convolutional neural networks for computer vision are explained. The fully connected layers, 2D convolutional layers, and 2D pooling layers, as well as their roles in convolutional neural networks for computer vision, are detailed. The concept of transfer learning in computer vision is explained. The dataset used for solving the task of framing an object in the image within the tightest possible box, with sides parallel to the vertical and horizontal axes of the image, is presented. The IoU metric, along with some loss functions and other IoU-based metrics, are explained. The architecture of the VGG16 and ResNet50 network backbones is described, and models with these pretrained backbones are used for transfer learning on the previously mentioned task. A relatively simple convolutional network is also trained, and the results are presented and compared. Additionally, some issues with deep neural networks and their solutions are discussed.

• Keywords: duboko učenje; konvolucijska neuronska mreža; detekcija objekata
• Keywords (english): deep learning; convolutional neural network; object detection
• Language: croatian
• URN:NBN: urn:nbn:hr:168:541311
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 22:18:20