Refine
Year of publication
Document Type
- Diploma Thesis (451)
- Bachelor Thesis (218)
- Book (93)
- Periodicalpart (73)
- Master's Thesis (37)
- Researchpaper (28)
- Other (19)
- Article (14)
- Study Thesis (9)
- Report (7)
Language
- German (911)
- English (42)
- Multiple languages (1)
Keywords
- Bibliothek (58)
- Öffentliche Bibliothek (54)
- Internet (39)
- Deutschland (21)
- Fachhochschule Stuttgart - Hochschule für Bibliotheks- und Informationswesen ; Zeitschrift (19)
- öffentliche Bibliothek (19)
- Konzeption (16)
- Barrierefreiheit (14)
- Elektronische Zeitschrift; Stuttgart / Fachhochschule Stuttgart - Hochschule der Medien (14)
- Verlag (14)
Institute
- Bibliotheks- und Medienmanagement (Bachelor, Diplom) (305)
- Informationswirtschaft (Bachelor, Diplom) (130)
- Bibliothek (77)
- Bibliotheks- und Informationsmanagement (Bachelor, Master) (62)
- FB 1: Druck und Medien (57)
- FB 3: Information und Kommunikation (38)
- Medieninformatik (Bachelor, Diplom) (35)
- Rektorat (27)
- FB 2: Electronic Media (25)
- Audiovisuelle Medien (Bachelor, Diplom) (20)
Today’s digital cameras use a mosaic of red, green, and blue color filters to capture images in three color channels on a single sensor plane. This thesis investigates the use of convolutional neural networks (CNNs) for demosaicing – the process of reconstructing full-color images from raw mosaic sensor data. While there are existing CNNs for demosaicing raw images from the well-established regular Bayer color filter array (CFA), this thesis focuses on how they perform on alternative non-regular sampling patterns that produce less aliasing artifacts, namely the stochastic Gaussian- and the RandomQuarter sampling pattern (Backes and Fröhlich, 2020).
A basic UNet (Ronneberger et al., 2015) and the spatially adaptive SANet (T. Zhang et al., 2022) are implemented in a supervised training pipeline based on the PixelShift200 image dataset (Qian et al., 2021) to investigate their suitability for the irregular demosaicing task. The experiments indicate that the basic UNet encounters difficulties in restoring the missing color values, whereas the spatially adaptive convolutional layers help in processing the irregularly sampled raw images.
In addition, this thesis enhances SANet effectiveness by employing an alternative residual branch based on a CFA-normalized Gaussian filter, as well as a tileable modification to the Gaussian CFA pattern. The modified SANet is shown to outperform the conventional dFSR algorithm (Backes & Fröhlich, 2020) in terms of peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM).