Deep learning video stabilization
Loading...
Date
2025-01-16
Authors
Σαρδέλης, Κωνσταντίνος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
With the rapid advancement of technology in portable devices (such as mobile phones, smartwatches, etc.), it is possible to capture high-quality videos without the use of heavy equipment like tripods and stabilizers. However, these videos often contain unwanted camera movements and vibrations. For this reason, digital video stabilization is a rapidly evolving and intensively studied field. The purpose of this work is to contribute to video smoothing through deep learning.
Initially, the reader is provided with an introduction to digital video stabilization through optimization and low-pass filtering, while simultaneously introducing basic concepts of computer vision. Then, a brief overview of deep learning and how convolutional networks learn and operate is presented. This is followed by detailed studies and implementations of video stabilization algorithms using artificial intelligence. These methods can be categorized into 1) camera path reconstruction/optimization without unwanted noise, 2) creation of stable frames through supervised learning, and 3) generation of new intermediate frames in the video. Through these methods, modern deep learning technologies such as transfer learning, perceptual cost functions, and adversarial learning are introduced. All algorithms are compared using a set of metrics, and their effectiveness is discussed. Finally, the issue of lack of training data is addressed by creating synthetic video data. I compare existing methods with two novel ones, which use artificial intelligence and statistical analysis to create realistically shaky videos from existing stable videos.
Description
Keywords
Computer vision, Deep learning, Video stabilization