Deep Joint Demosaicking and Denoising

摘要

Demosaicking and denoising are the key first stages of the digitalrnimaging pipeline but they are also a severely ill-posed problem thatrninfers three color values per pixel from a single noisy measurement.rnEarlier methods rely on hand-crafted filters or priors and still exhibitrndisturbing visual artifacts in hard cases such as moir´e or thin edges.rnWe introduce a new data-driven approach for these challenges: werntrain a deep neural network on a large corpus of images insteadrnof using hand-tuned filters. While deep learning has shown greatrnsuccess, its naive application using existing training datasets doesrnnot give satisfactory results for our problem because these datasetsrnlack hard cases. To create a better training set, we present metrics tornidentify difficult patches and techniques for mining community photographsrnfor such patches. Our experiments show that this networkrnand training procedure outperform state-of-the-art both on noisy andrnnoise-free data. Furthermore, our algorithm is an order of magnitudernfaster than the previous best performing techniques.