Recently, two piecewise smooth models L0smoothing and relative total variation (RTV) have been proposed for feature/structure-preserving filtering. One is very efficient for tackling image with little texture patterns and the other has appearance performance on image with abundant uniform textural details. In this work, we present a general relative total variation (GRTV) method, which generalizes the advantages of both approaches. The efficiency of RTV depends on the defined windowed total variation (WTV) and windowed inherent variation (WIV), which focus on edge enhancing and texture suppressing respectively. The key innovations of the presented GRTV method are to extend the norm of WTV in RTV from 1 to [0, 1] and set the norm of WIV inversely proportional to the norm of WTV. These modifications substantially improve the structure extraction ability of RTV. The presented GRTV also improves the edge-boundary enhancing ability of L0smoothing and further enables it to deal with images containing complex textural details and noises. Furthermore, the L2-norm data fidelity term replaced by L1-norm is discussed. Experimental results demonstrate that the proposed method presents better performance as the state-of-the-art methods do.
- Image smoothing
- Iterative Reweighed Least Square
- non-convex regularization
- Structure preserving