Title: | Fast Incremental Image Reconstruction with CNN-enhanced Poisson Interpolation |
Authors: | Erzar, Blaž Lesar, Žiga Marolt, Matija |
Citation: | WSCG 2023: full papers proceedings: 1. International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision, p. 204-212. |
Issue Date: | 2023 |
Publisher: | Václav Skala - UNION Agency |
Document type: | konferenční příspěvek conferenceObject |
URI: | http://hdl.handle.net/11025/54426 |
ISBN: | 978-80-86943-32-9 |
ISSN: | 2464–4617 (print) 2464–4625 (CD/DVD) |
Keywords: | rozpoznávání obrazu;numerická interpolace;vícesíťová metoda;konvoluční neuronové sítě;automatický kodér |
Keywords in different language: | image recognition;numerical interpolation;multigrid method;convolutional neural networks;autoencoder |
Abstract in different language: | We present a novel image reconstruction method from scattered data based on multigrid relaxation of the Poisson equation and convolutional neural networks (CNN). We first formulate the image reconstruction problem as a Poisson equation with irregular boundary conditions, then propose a fast multigrid method for solving such an equation, and finally enhance the reconstructed image with a CNN to recover the details. The method works incrementally so that additional points can be added, and the amount of points does not affect the reconstruction speed. Furthermore, the multigrid and CNN techniques ensure that the output image resolution has only minor impact on the reconstruction speed. We evaluated the method on the CompCars dataset, where it achieves up to 40% error reduction compared to a reconstruction-only approach and 9% compared to a CNN-only approach. |
Rights: | © Václav Skala - UNION Agency |
Appears in Collections: | WSCG 2023: Full Papers Proceedings |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
F47-full.pdf | Plný text | 1,97 MB | Adobe PDF | View/Open |
Please use this identifier to cite or link to this item:
http://hdl.handle.net/11025/54426
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.