Image Inpainting Methods Based On Image Decomposetion And Sparse Representation

Image inpainting is an active field in image processing. It is widely applied to digital restoration of valuable artworks, damaged cinefilm, and image object editor and so on. This technology offers new ideas for video transmission error correction, super-resolution, image compression and demosaic etc.The aims of this paper are to cope with both the pixels missing in image spatial domain and missing wavelet coefficient of image, and two different models are proposed. To deal with the pixels missed in image spatial domain, the image can decomposed into two parts cartoon and texture according to the property of natural image, and the PDEs-base and texture synthesis-based image inpainting model are considered to inpaint them. In order to cope with the “intensity discontinuity” problem of TV Inpainting model which proposed by Chan et al to recover cartoon, a four-order PDEs-based image inpainting model which takes advantage of the isophotes direction information is proposed. To decrease the complexity of the proposed model, an improved Fast Marching method based model is used to fill the missing pixels, so that the iteration is hugely decreased. The performance of the Criminisi algorithm used to recover texture is degraded by its error propagation and high complexity, in order to solve these problems, a distance weighted similarity measurement is proposed to replace the original measurement based on the Markov random field theory. By using the statistical tools, the relationship between center distance and patch average distance is discussed, so that a local search method is used to replace the exhaustive search method. The experimental results demonstrate that the proposed model can efficiently alleviate the “intensity discontinuity” problem and suppress the error propagation compared with the TV Inpainting model, the Criminisi algorithm and the TV-Decomposed model. For the same miss rate, the signal to noise ratio and peak signal to noise ratio index of the proposed model is improved 3-6 d B and 1-2 d B, respectively.To deal with the missing of wavelet coefficients of image, a Two-step Wavelet model is proposed based on three facts: the lack of Prior model for wavelet coefficients, the rich of Prior model for image spatial domain and the characteristic of the low-frequency subband and the high-frequency subband. In the first step, the sparse prior is regarded as the Prior model of the low-frequency subband of the image, and the sparse representation model based inpainting model is proposed, which can obtain well estimate of damaged image; In the second step, to recover the high-frequency subband, following the idea of Chan et al., we transform the original difficult problem in wavelet domain to the spatial domain problem with wavelet coefficients constraints. Compared with TV Wavelet model and the Nonlocal TV Wavelet model, when the missing rate is same, the proposed model has better performance of suppressing noise, especially when standard deviation of noise is ≥30, the results of(Nonlocal) TV Wavelet model have visible noise; as the missing rate is increased, the inpainting performance of the proposed model decreased much slower, which means it more robust than the other two models; for the same miss rate, the peak signal to noise ratio and structural similarity index of the Two-step Wavelet model is more than 2 d B and 0.1 higher than the other two models, respectively.