Comparison of Parameter reduction methods for Change Detection in Satellite  Imagery

Muliarska, Yana

Home
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Бакалаврська програма "Комп'ютерні науки"
→
2023
→
View Item

dc.contributor.author	Muliarska, Yana
dc.date.accessioned	2024-02-14T09:15:14Z
dc.date.available	2024-02-14T09:15:14Z
dc.date.issued	2023
dc.identifier.citation	Muliarska, Yana. Comparison of Parameter reduction methods for Change Detection in Satellite Imagery / Yana Muliarska; Supervisor: Petr Simanek; Ukrainian Catholic University, Department of Computer Sciences. – Lviv: 2023. – 37 p.: ill.	uk
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4410
dc.language.iso	en	uk
dc.title	Comparison of Parameter reduction methods for Change Detection in Satellite Imagery	uk
dc.type	Preprint	uk
dc.status	Опублікований і розповсюджений раніше	uk
dc.description.abstracten	Change Detection is a critical problem in Computer Vision with applications in various domains such as medical detection, satellite imagery, quality control, and traffic analysis. However, existing change detection models often have many pa- rameters, making them computationally expensive and challenging to implement in real-world applications. This study focuses on reducing the parameters set for the models designed explicitly for Change Detection in Satellite Imagery. These models typically process large-scale images, which can demand significant mem- ory resources and take considerable time to compute. As a solution, we implement three approaches, evaluate and compare their performance on a toy CNN model and an advanced SNUNet-CD model [9], designed for the Change Detection task. The highest parameter reduction rate we achieved for SNUNet-CD is 10.4% (1.25 million parameters) with only a 3.7% model accuracy drop. The experiments demonstrate that, when utilizing our methods, SNUNet-CD outperforms several SOTA models in the change detection domain. We succeeded in surpassing UNet++_MSOF [22] with respect to parameter count, while the original SNUNet-CD with 32 channels was unable to do so. The code implementation of this work is available on GitHub: https://github. com/muliarska/parameter-reduction-for-change-detection/.	uk