Odstranjevanje pegastega šuma iz slik SAR z uporabo globokega učenja

magistrsko delo

Tadej Habjanič (Author), Dušan Gleich (Mentor)

Abstract

Postopek odstranjevanja pegastega šuma je neizogiben pri obdelavi slik z radarjem s sintetično odprtino (SAR). Obstaja več različnih metod za odstranjevanje pegastega šuma, vendar se je postopek s konvolucijsko nevronsko mrežo (CNN) izkazal kot zelo učinkovita metoda. Pri preprosti strukturi CNN se še vedno izgubi precejšnje število podrobnosti na sliki. Za rešitev tega problema je bila uporabljena arhitektura kodirnika – dekoderja. Model se uči s pristopom, ki temelji na veliki količini podatkov, z uporabo algoritma gradientnega spuščanja s kombinacijo spreminjanja ojačanja pri odstranjevanju šuma in funkcije izgube celotne variacije. Poskusi, izvedeni na realnih slikah, kažejo, da ta metoda dosega pomembne izboljšave v primerjavi z ostalimi metodami.

Keywords

pegasti šum;radar s sintetično odprtino;konvolucijska nevronska mreža;arhitektura kodirnik – dekodirnik;magistrske naloge;

Data

Language:	Slovenian
Year of publishing:	2023
Typology:	2.09 - Master's Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[T. Habjanič]
UDC:	004.85:[004.932:528.8.044.2](043.2)
COBISS:	189928707
Views:	89
Downloads:	11
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Despeckling of SAR image using deep learning
Secondary abstract:	Despecklink process is inevitable in synthetic aperture radar (SAR) images processing. There are several different methods for removing speckled noise, but the Convolutional Neural Network (CNN) approach has proven to be a very effective method. With a simple CNN structure, a considerable amount of detail in the image is still lost. An encoder – decoder architecture was used to solve this problem. The model is trained using an approach based on a large amount of data, utilizing the gradient descent algorithm in combination with adaptive gain tuning for speckle removal and the total variation loss function. Experiments performed on real images show that this method achieves significant improvements compared to other methods.
Secondary keywords:	descpeckling;synthetic aperture radar;convolutional neural network;encoder – decoder architecture;
Type (COBISS):	Master's thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Elektrotehnika
Pages:	1 spletni vir (1 datoteka PDF (X, 62 f.))
ID:	21796425