Like the SNN filter, the Kuwahara filter is an adaptive, edge-preserving smoothing filter. It works by dividing the support into four sub-regions (as shown here) and selecting the one with the lowest variance (i.e. the smallest spread of values). It then returns the mean of the pixels in that region. This ensures that only relatively homogenous data contribute to the output, and edges are therefore not averaged across.
This filter preserves the position and magnitude of edges, and even enhances them in most cases, but it tends to blur smoother areas. The result is flat areas with sharp edges. Because of this, this filter is often used in biomedical applications as the first step in identifying discrete anomalies in noisy images, in diagnosing brain tumors from magnetic resonance images for example. The well-defined smooth regions are easier to apply thresholding to, isolating irregularities. The workflow is analogous to mapping amplitude anomalies.
The Kuwahara filter does not cope well with a lot of random noise. A crosshatched appearance can result. As shown in Figure 8 of Hall (2007), a single prior pass of a 3 × 3 median filter alleviates this problem.
- Hall, M (2007). Smooth operator: smoothing seismic horizons and attributes. The Leading Edge 26 (1), January 2007, p16-20. doi:10.1190/1.2431821