Research On Learning-based Super-Resolution Reconstruction Method For Face Image

Face image super-resolution reconstruction is one of key technologies to enhance the resolution and evaluate the quality of face image, so it attracts more attention and is researched widely in computer vision, video surveillance, public security and other fields.The learning-based super-resolution method via sparse representation is mainly researched for single face image in this paper, after studying the knowledge related to face image super-resolution comprehensively. It is integrated mainly by the dictionary training step and the testing step for high-resolution face image reconstruction. The former can train a pair of dictionaries for the latter, which reflect the relationship between high and low resolution face images. Three approaches are proposed in this paper when researching these two steps. To reduce the error from image feature extraction for the trained dictionaries, two approaches based on learned feature dictionary and sparse representation are proposed. In the testing step, to reduce the structural error for reconstructed image, the approach based on total-variation generalized accelerated poximal gradient(GAPG-TV) and sparse representation is proposed.In the super-resolution reconstruction based on sparse representation, feature extraction directly affects the quality of learned feature dictionaries and is important to evaluate the final reconstruction quality for face image. To reduce the influence from upsampling errors, the feature extraction way based on Gaussians and Laplace(GASS-LAP) is proposed by making full use of the neighborhood pixels' information in this paper. To make the feature-extraction image enhance the association with other face training images, this paper proposes another feature extraction way based on Sparse Autoencoding(SAE). Further more, two super-resolution approaches are proposed correspondingly with trained GASS-LAP-based feature dictionaries or trained SAE-based feature dictionaries. The experimental results with ORL face database show that two proposed super-resolution approaches can both reconstruct better face image than other conventional approaches. Besides, the proposed approach based on GASS-LAP than the SAE-based approach.The estimated details for reconstructed face image by the conventional approaches have some global structural errors compared with the real image, due to the sparse approximation error, dictionary's completeness, image patches division or fusion. After training GASS-LAP-based dictionaries, to further improve the reconstructed face image's quality, in the reconstruction testing step, this paper proposes an approach to repair the reconstructing image globally by combining the GAPG-TV and the degradation factor obtained from sparse representation. The experimental results with ORL face database show the proposed approach based on GAPG-TV and sparse representation performs better in recovering the structure of high-resolution face image, than the conventional approaches and another proposed approaches in this paper.