CATADIOPTRIC IMAGE DENOISING: A SPATIALLY VARIANT APPROACH

A catadioptric camera uses a conventional camera in conjunction with a quadratic mirror for capturing an omnidirectional field of view in real-time. The resolution of catadiop- tric images, however, is non-uniform due to the mirror curvature. A widely used approach to processing catadioptric images is to apply classical methods to them directly or via a trans- formed domain. The aim of this work is to demonstrate that for the task of image denoising, an appropriate approach is to modify classical methods so that they become spatially adaptive with the non-uniform resolution of catadioptric images. To this end, we modify the famous Rudin-Osher-Fatemi (ROF) denoising model by introducing a space-variant regularizer. The proposed model comprises a spatially varying total variation term, which adjusts the edge- preservation and the noise reduction abilities in the whole image domain. We carry out an empirical evaluation of the performance of the proposed model compared with the widely used methods for processing catadioptric images. The results reveal that, despite its simplic- ity, our model improves the performance of the original method in terms of both quantitative and qualitative aspects.