Thanks to some technical progress in filter array technologies, capturing multimodal data of a scene, such as polarization and/or spectral information, in a single acquisition is possible. Nevertheless, a reconstruction procedure referred to as demosaicing is required to produce the various full definition images in each band. The computational imaging community often needs full-reference images to assess the performance of these reconstruction algorithms. Nevertheless, these multidimensional data are increasingly complex to capture, as the number of channels increases in the image. This often leads to misalignment among channels or noise introduced by imperfect optics. In this work, we propose a study on the use of these imperfect data in the context of demosaicing. The impact of misalignment is assessed on an existing Color Polarization Filter Array database, from which we demosaic the data using three types of demosaicing algorithms and use either pre-processed or raw dataset. We found that denoising and registration do not modify the hierarchy of best performing algorithms in case of sub-pixel shifts. We also show that visual artifacts, usually attributed to drawbacks of training-based demosaicing algorithms, may instead be due to the use of unregistered images during the training stage of the algorithms.
|