Non-Local Sparse Representation Method for Demosaicing of Single DoFP Polarimetric Image

The images in different polarization directions collected from division-of-focal-plane (DoFP) imaging system are under-sampled. To solve the demosaicing problem of DoFP imaging, this paper presents a learning model based on sparse representation to optimize the interpolation result of DoFP images. Firstly, image blocks rich in edge or texture information are selected according to the local gradient, and these blocks are clustered based on non-local similarity to learn a sub-dictionary from each class adaptively. The model uses local similarity and sparsity of coding coefficients as regularization terms to minimize coding errors, and then the algorithm iteratively optimizes dictionary atoms and coding coefficients alternately to obtain enhanced images. The experiment takes 8 composed DoFP images as reference and compares the interpolation results of the proposed algorithm with different methods. The proposed method obtains smaller interpolation error than other methods at every image in the experiment.