Otkrivanje ljudi na slikama upotrebom konvolucijske neuronske mreže

undergraduate thesis

• Title (english): Human Detection in Images Using Convolutional Neural Networks
• Author: Dorian Ivanković
• Mentor: Goran Delač (mentor)
• Committee member: Goran Delač (predsjednik povjerenstva)
• Committee member: Marin Šilić (član povjerenstva)
• Committee member: Klemo Vladimir (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2018-07-13, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U ovom radu pokrivene su osnovne tehnike strojnog učenja relevantne za razvijena programska rješenje i za temu rada općenito. Opisani su postupci optimizacije funkcija gubitka za različite probleme, kao i nekoliko metoda kojima se one mogu optimizirati.
Navedene su osnovne značajke i opisane su neuronske mreže, kao i jedan od učinkovitih algoritama treniranja neuronskih mreža, a to je algoritam unazadnog ulančavanja.
Zatim su proučene specifičnosti i način rada konvolucijskih neuronskih mreža... More koje su posebno pogodne za probleme računalnog vida.
Opisan je implementirani klasifikator slika ovisno o prisutnosti ljudi na slici korištenjem biblioteke TensorFlow. Pokazano je kako je model treniran, korišteni podatci i rezultati pomoću F1 mjere. Postignut je rezultat s F1 mjerom od 0.9 i gotovo jednakim vrijednostima preciznosti i sjećanja.
Nakon toga dan je uvid u najpoznatije metode detekcije objekata na slikama s posebnim naglaskom na YOLO algoritam koji je korišten u implementaciji pa je zato malo detaljnije opisan od ostalih metoda.
Opisano je programsko rješenje detekcije ljudi na slikama korištenjem biblioteke otvorenog koda darkflow, koje postiže iznimno dobre rezultat s mjerom mAP (mean average precision od 90.42%).
Daljnji rad mogao bi se usmjeriti prema kompleksnijim modelima detekcije objekata i prošiti problem na više vrsta objekata. Takav sustav koji dobro radi u stvarnom vremenu mogao bi doprijenjiti razvoju različitih inteligentih i sigurnosnih sustavima.
Less

Abstract (english)

This paper covers the basic machine learning techiques relevant for developed solutions and for the topic of this thesis in general. Different loss functions have been described, as well as several methods for optimizing them.
Neural networks and their basic features have been explained as well as one of the effective algorithms used to train them, the backpropagation algorithm.
Then, the specifics and advantages of convolutional neural networks have been studied, which are particularly... More suitable for computer vision problems.
Classifying images depending on the presence of people using TensorFlow has been developed and described. It is shown how is the model trained, used datasets and results using the F1 measure.
Results with 0.9 score of F1 measure and precision and recall of similiar values have been achieved.
This is followed by an introduction to the most popular object detection methods in images with a special emphasis on the YOLO algorithm used in the implementation which is, therefore, described in more detail that the other methods.
A software solution for people detection in images using open source dark flow library has been developed and described, achieving very good results with mAP (mean average precision) of 90.42%.
Further work could focus on more complex object detection models and extend the problem to more types of objects. Such a system that works well in real time could assist in development of various intelligent and security systems.
Less

• Keywords: Strojno učenje; ADAM optimizator; Konvolucijske neuronske mreže; Algoritam unazadnog ulančavanja; Klasifikacija; Detekcija; YOLO; INRIA; TensorFlow
• Keywords (english): Machine learning; ADAM optimizer; Convolutional neural networks; Backpropagation algorithm; Classification; Detection; YOLO; INRIA; TensorFlow
• Language: croatian
• URN:NBN: urn:nbn:hr:168:293412
• 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: 2019-03-19 14:30:11