چكيده به لاتين
In recent years, the proposed approaches based on sparse representation as a function of the spectral dictionary and sparse coefficients have reached acceptable results in hyperspectral images super-resolution. Among these approaches, the ones used the concept of similar image patches grouping, have been able to use two important prior knowledge in hyperspectral images i.e. non-local resemblance and local sparsity in sparse representation simultaneously. However, the use of resemblance criteria based on differentiation of corresponding pixels on the one hand and the use of non-accurate dictionary learning algorithms on the other hand have prevented the full exploitation of “precision capacity” of the sparse representation of these images. In this thesis, the similar image patch-blocks have been grouped by using a new approach towards the grouping concept and introducing a new modified criterion based on Euclidean norm. As well as at the first time in hyperspectral image super-resolution, spatial-spectral dictionary (SSD) and sparse coefficients of each group are learned by using OSDL algorithm, then the whole image is restored by extracting each block from its corresponding group and putting them together. To maintain spectral structure, “spectral total-variations” was used as a regulating constraint in the final objective function. The simulation results represent a much better performance of the proposed method compared to Bicubic method.
Keywords: Hyperspectral image, Image patch-block-group, OSDL algorithm, Sparse
Representation, Super-resolution.
