Detekcija prometnih znakov s konvolucijskimi nevronskimi mrežami

Kovačič, Andreja (author), Skočaj, Danijel (mentor)

Diplomsko delo se ukvarja z detekcijo in prepoznavanjem prometnih znakov z metodo Faster R-CNN. Razišče možnost uporabe generiranih podatkov za validacijo učenja, kot odgovor na omejeno velikost učne množice. Ker Faster R-CNN zaradi načina učenja ne dopušča običajnega učenja na težkih primerih (ang. bootstrapping), uporabimo novo metodo - sprotno iskanje težkih primerov. Na koncu iščemo optimalen način doučenja že obstoječega modela.

• Language: Slovenian
• Keywords: Faster R-CNN, klasifikacija, detekcija, sprotno iskanje težkih primerov, doučenje, prometni znaki
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-95566
• Publication date in RUL: 20.09.2017

Secondary language

• Language: English
• Title: Traffic sign detection with convolutional neural networks

Abstract

The goal of this thesis is to describe and use method Faster R-CNN for detection and recognition of traffic signs. It explores the possibility of using artificially generated images in validation set, in hopes of saving real images for train set. We tackle a real world problem of growing dataset through time. We'll try to find an optimal way to augment the already learned model with new images. Lastly, we try to apply a new method, online hard example mining, which is essentially bootstrapping for end-to-end systems.

• Keywords: Faster R-CNN, classification, detection, online hard example mining, fine-tuning, traffic signs