Description
To produce a color image, the “color-blind” sensor in a digital camera is attached to a grid of colored filters, called Bayer filter. This design produces a sparse raw image, whose pixels have only the value of a single (position-dependent) channel. The missing values are interpolated from the measured values by demosaicing, i.e., removing the mosaic. Many demosaicing methods are known: simple linear interpolation of each channel independently, gradient-based demosaicing, i.e., interpolation of R and B with a help of gradient of G, or most recently neural demosaicing that employs neural networks. The choice of a demosaicing method influences the information flow between the color channels during the demosaicing, and may have an effect on performance of forensic methods or steganographic detectors.
The goal of the thesis is to study the spill-over of information for different demosaicing methods. This is measured by comparing the color channels of images, produced with existing demosaicing methods. Common design choices should also be considered, e.g., the presence of median filter or color correction. The impact of the demosaicing method should be further demonstrated by measuring the performance of a down-stream task, e.g., PRNU estimation.