An improved image inpainting algorithm based on multi-scale dictionary learning in wavelet domain

The image inpainting method based on the single scale often leads to the details part inpainting deficiency. To solve this problem, we propose an image inpainting method based on the multi-scale dictionary learning. First, we select some fine images for dictionary learning, and then the sample images do the wavelet transform one by one. Second, for each sub-band after the images are transformed into the wavelet domain, a large number of blocks of samples are selected in a superimposed manner to make up the training set, the K-means singular value decomposition (K-SVD) using wavelets approach presented here applies dictionary learning in the analysis domain, sub-dictionaries at different data scales, consisting of small atoms, are trained. Finally we combine each sub-band dictionary into a global one to repair damaged image. The pixel loss of natural images, scratches and text removal experiments, demonstrate that our approach´s applicability over a set of degraded images, at the same time, the PSNR and SSIM are improved.