Pattern-Invariant Unrolling for Robust Demosaicking

To acquire color images, most commercial cameras rely on color filter arrays (CFAs), which are a pattern of color filters overlaid over the sensor's focal plane. Demosaicking describes the processing techniques to reconstruct a full color image for all pixels on the focal plane array. Most demosaicking methods are tailored for a specific CFA, and tend to work poorly for others. In this work we present an algorithm for demosaicking a wide variety of CFAs. The proposed method allows to blend the knowledge of the CFA with information coming from data, employing a novel transformation and pattern-invariant loss function. The method is based on the unrolling of an algorithm based on a neural network learned on available examples. Preliminary experiments over RGB and RGBW CFAs show that the method performs well over a range of CFAs and is competitive for CFAs for which competing methods were tailored to work well on.