PROJECT TITLE :

Depth Super-Resolution via Joint Color-Guided Internal and External Regularizations

ABSTRACT:

Many real-world applications make heavy use of depth information. In practise, however, depth maps tend to have a lower resolution than colour images because of the limitations of depth sensor technologies. To achieve depth super-resolution, we propose to combine the internal smoothness prior and the exterior gradient consistency requirement in graph domain. Firstly, a novel graph Laplacian regularizer is presented to preserve the depth's inherent piecewise smoothness, which has useful filtering qualities. The weight matrix of the respect graph has been defined to take advantage of both depth and the related guidance image. In contrast, we introduce a graph gradient consistency constraint to guarantee that the graph gradient of depth is near to the thresholded gradient of guidance, inspired by the finding that the gradient of depth is modest except at edge separation regions. The gradient thresholding model is reinterpreted as variational optimization with a sparsity constraint. Here, we address the issue of structure inconsistency between the depth and guidance. Finally, a unified optimization framework incorporating both internal and external regularizations may be addressed effectively by ADMM. Experiments show that our strategy outperforms the current state of the art in terms of both objective and subjective quality evaluations.