Project Name: Machine Learning Methods for Demosaicing and Super-resolution

Related Work: For the traditional methods of demosaicing we mentioned the low-pass-filtered chrominance method in the lecture. We can use Gaussian filter or bilateral filter [1] as our low-pass filter. Another approach is edge-directed interpolation, which is proposed by Gunturk et al [2], it tries to interpolate the image based on the direction of edges, to preserve the edge sharpness. We also explored the linear interpolation method proposed by Malvar et al [3], it uses only linear kernels to perform the demosaicing by optimizing the kernel gain parameters on Kodak photo dataset. There are many machine learning methods for image super-resolution[4], for example, k-Nearest Neighbors, Support Vector Regression and Super-Resolution Convolutional Neural Network. Machine learning techniques are used to predict the missing color/texture information. In the general image optimization problem, we model the problem as an optimization problem[5][6]. As mentioned in [5], image optimization problems contain (1) a variable which represents the target image to be reconstructed, (2) a linear operation matrix which represents the downsampling/demosaicing process, (3) a penalty measure which represents the difference of the results of downsampling/demosaicing from the measured data, and (4) the priors and constraints on the the variables. Usually we encode the prior information as a penalty (regularization) term in the objective function, to manage the ill-conditionedness of the original reconstruction problem. Once we define the constrained optimization problem, iterative update methods are applied to solve the best estimate of the higher resolution image.