Uporaba globokega učenja za semantično segmentacijo okolja morskih plovil

GRAH, MARCEL (avtor), Perš, Janez (mentor)

Z napredkom razvoja računalniške tehnologije se je znatno povečala tudi moč umetne inteligence. V diplomskem delu so opisani postopek izdelave, učenje in rezultat uporabljene nevronske mreže. Glavni cilj dela je bila izdelava klasifikacije posameznih objektov, zajetih s kamero GoPro. Ta je pritrjena na vrhu samovozečega čolna. Z uporabo segmentacije lahko zazna objekte v okolici. Postopek določanja klasifikacije posameznemu slikovnemu gradivu se imenuje semantična segmentacija. Zaradi pomanjkanja morskih učnih baz je bilo treba pridobiti lastne učne slike. V diplomskem delu sta opisana zgradba in delovanje konvolucijske nevronske mreže. Ta je sestavljena iz dveh delov: prvi se imenuje kodirnik in se uporablja za prepoznavo objektov, drugi pa se imenuje dekoder in podatkom kodirnika določi prostorsko lokacijo. Ob koncu diplomskega dela so predstavljeni rezultati in meritve, pridobljene z uporabo nevronske mreže. Glede na pridobljene rezultate je bilo treba ugotoviti, ali se sistem lahko zanesljivo uporablja kot eden od ladijskih senzorjev za prepoznavanje oklice. Preverilo se je tudi, kako različni dejavniki, kot so lokacija, velikost slike in kakovost slike, vplivajo na pridobljene rezultate.

• Jezik: Slovenski jezik
• Ključne besede: umetna inteligenca, semantična segmentacija, konvolucijska nevronska mreža.
• Vrsta gradiva: Diplomsko delo/naloga
• Organizacija: FE - Fakulteta za elektrotehniko
• Leto izida: 2020
• PID: 20.500.12556/RUL-116170
• Datum objave v RUL: 20.05.2020

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Using Deep Learning for Semantic Segmentation of Marine Environment

Izvleček

With the advancement of computer technology, the power of the artificial network was also increased. Thesis, describe the process of making, training and the result of using neural networks. Main goal was to classify individual objects covered by a Go pro camera. The latter is attached to the top of the self-driving boat. By using segmentation, the boat can "detect" objects in the surrounding area. The process of determining the classifications of individual images is called semantic segmentation. Due to the lack of annotated sea image bases, it was also necessary to produce the training images. The middle section described a structure of a convolutional neural network. This one consisted of two parts. The first part, called the encoder that is used to identify objects. The second part is called a decoder, which determines the spatial location of the encoder data. The last sections show the results in the measurements obtained from neural network. The obtained results were used to determine, if the camera could be used as one of the sensors for detecting surrounding area. It was also necessary to present how different factors such as location, image size in image quality affect the results.

• Ključne besede: artificial intelligence, semantic segmentation, convolutional neural network.