Duboka konvolucijska arhitektura EfficientNetV2

undergraduate thesis

• Title (english): Deep convolutional architecture EfficientNetV2
• Author: Karlo Frankola
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Matej Grcić (član povjerenstva)
• Committee member: Petra Bevandić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Klasifikacija slika važan je problem u računalnom vidu s mnogim zanimljivim primjenama. Trenutno stanje tehnike koristi duboke konvolucijske mreže. Velika računalna složenost tih modela dovodi do problema u treniranju i kasnijoj evaluaciji. To je dovelo da se posebno traže konvolucijske arhitekture koje kroz puno manji broj parametara svejedno mogu ostvariti komparabilnu točnost. U okviru rada odabran je okvir za automatsku diferencijaciju te biblioteke za rad s tenzorima i slikama. Ukratko je opisana teorijska podloga strojnog i dubokog učenja te neke klasične konvolucijske arhitekture. Istražene su efikasne arhitekture EfficientNet i EfficientNetV2 te uspoređeni njihovi rezultati s rezidualnim mrežama.

Abstract (english)

Image classification is an important problem in computer vision with many interesting applications. Current state of the art techniques use deep convolutional networks. The large computational complexity of such models leads to problems in training and later evaluation of them. This lead to a special search for convolutional architectures that even with significantly less parameters are able to achieve comparable accuracy. In this paper, an automatic differentiation framework was chosen, and libraries for working with tensors and images. The theoretical basis for machine and deep learning was introduced and some classic convolutional architectures were explained. Efficient architectures of EfficientNet and EfficientNetV2 were researched and their results were compared to residual networks.

• Keywords: duboko učenje; konvolucijske arhitekture; računalni vid; ResNet; EfficientNet; EfficientNetV2
• Keywords (english): deep learning; convolutional architectures; computer vision; ResNet; EfficientNet; EfficientNetV2
• Language: croatian
• URN:NBN: urn:nbn:hr:168:968729
• 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: 2023-01-15 10:08:31