Single image superresolution using sparsity and dictionary learning in wavelet domain

Recently sparse representation is proven to be very successful for image processing applications. This paper proposes a superresolution approach that utilizes the decorrelating and sparsifying property of discrete wavelet transform, with the signal-fitting capability of sparse representation over learned dictionaries. Two dictionaries are learned (using K-SVD algorithm) for each wavelet subband: one for the low resolution one for the high resolution images. In the training set, noisy variants of the high a resolution image obtained by interpolating its low resolution counter-part, are intentionally included. A patch based approach is employed and different patch sizes are studied. The sparse representation coefficients for the respective low and high resolution images are assumed to be the same. Experiments are conducted using the Kodak24-image set. The proposed algorithm is proven to be competitive with the leading super-resolution techniques both visually and quantitatively. With a patch size of 11×11 the proposed method is 0.89 dB better than the method proposed by Elad et.al. and 2.19 dB better than bicubic interpolation.