Národní úložiště šedé literatury Nalezeno 1 záznamů.  Hledání trvalo 0.01 vteřin. 
RAW image debayerization using deep neural network
Balušík, Peter ; Myška, Vojtěch (oponent) ; Rajmic, Pavel (vedoucí práce)
This thesis focuses on the problem of demosaicing; specifically, demosaicing using deep image prior. Deep image prior (DIP) is a concept that uses untrained convolutional neural networks to solve common reconstruction problems, with the only input information being an image degraded in some way. The aim of this thesis is to find out whether the DIP is a viable method for demosaicing problems. A new demosaicing method based on DIP is proposed and compared with common demosaicing methods. Different color filer arrays (CFAs) were tested to see the full potential of the proposed method. A numerical comparison was made using a variety of assessment methods. Based on this comparison, the proposed method proved to be similar, in some cases even better than the widely used Malvar’s demosaicing method. Visually, the proposed method displayed similar results to the finest method in the experiments – the Menon’s demosaicing method. Additionally, averaging the last few images of the optimization process proved to bring positive results in terms of numerical comparison. Even though the proposed method brought some interesting results, it turned out to be extremely computationally challenging when compared with other common demosaicing methods.

Chcete být upozorněni, pokud se objeví nové záznamy odpovídající tomuto dotazu?
Přihlásit se k odběru RSS.