Image super-resolution(SR)reconstruction addresses the problem of generating a high-resolution(HR)image from its low-resolution(LR)versions.Overcoming the physical limitations of low-cost imaging devices and the effect of complex environment,the SR technique based on joint dictionary learning and sparse representation of color image with depth map achieves more robust performance in extensive applications such as augmented reality,three-dimensional reconstruction,and man-machine interaction,and has become one of the most active research topics in the area of image processing and computer vision these years.In this paper,we study the existing image SR methods and innovatively propose a learning method for gray-depth joint dictionary.Then,on the basis of joint dictionary learning,two methods of image SR are proposed.The main works and innovations are introduced as follows:(1)This paper presents a new SR technique(JSR)for depth image based on gray-depth joint dictionary learning.The joint dictionary learning is first.Then the joint encoding increment is used for coding noise reduction,then the self-and multi-scale similarity are utilized to estimate the joint encoding coefficients for the HR joint patches.In the end,we solve the SR reconstruction problem and obtain the HR depth image by linear Bregman alternative algorithm.Note that the joint dictionary learning method first constructs gray-depth joint feature-patches with regard to the self-and multi-scale similarity.Then patches are clustered into several classes using K-means algorithm.Each class is then trained to form a compact sub-dictionary by Principal Component Analysis(PCA).Experimental results show that the proposed method achieves finer edges,and better reconstruction effect in both subjective and objective evaluation criteria.(2)This paper presents a new color-depth joint SR technique(CD-JSR)based on the improved neighbor embedding methods and gradient histogram preservation regularization term.On the basis of JSR method,the image non-local means is introduced as a regularization term into the neighbor embedding methods,improving the precision of the initial estimation.We incorporate the LR image multi-scalesimilarity and gradient histogram feature into the reconstruction model as regularization terms.Lastly,using the iterative algorithm and to simultaneously reconstruct an HR color image and depth image.Experiment results show that our method can effectively suppress the ringing phenomenon and edge dispersion artifact.It also achieves better refinement on texture and the edges,and simultaneously reconstructs HR color image and HR depth image closer to the original image. |