Adaptive Demosaicking using Multiple Neural Networks

Demosaicking is one of the important tasks in the image-processing pipeline in digital cameras using a single electronic sensor overlaid with a color filter array. We quantitatively shows that, demosaicking algorithms perform better in low-gradient flat areas than in high-gradient steep areas. Based on this, an adaptive scheme is proposed that uses more complex neural networks to tackle steep areas in larger sizes of neighborhoods. And interpolation is edge-directed with different networks for different chosen directions. Thus networks are specialized in learning and depicting non-linear spatial inter-pixel correlations at respective gradients and directions. Its performance surpasses Go's neural-network method greatly. Compared with 2 recent state-of-the-art methods, our method provides an excellent trade-off between computational expense and PSNR, and well preserves image edge information. As an extension, we compare the performance of these algorithms with and without Lukac's postprocessing.