Edge-Preserving Solution

Abstract

For an efficient visualization of hyperspectral image, we would like to combine as many features from the spectral bands as possible, typically characterized by the edges, boundaries, or textures. Various features in the data are clearly visible and observable only over a small subset of bands in the hyperspectral image. Typical pixel-level fusion techniques calculate fusion weights based on some relative importance of the pixels, also known as the saliency. In this process, the features that are available across only a few bands may not receive an adequate representation in the final image, if the weights are not chosen carefully. Thus, the weak features may get lost during the process of fusion. One would like to assign comparatively a higher weight to the pixels belonging to weak features in order to obtain a fused image where these features are appropriately represented. Most of the schemes of calculation of fusion weights or pixel saliency, therefore, employ some kind of high pass filtering to extract the necessary information. The process of high pass filtering is equivalent to the subtraction of a suitably low pass filtered image from the original one. A conventional low pass filtering such as Gaussian filtering, however, distorts the edges in the image. If this low pass filtered image is used for the calculation of fusion weights, the edges and other sharp features do not get an appropriate representation, and hence the fused image contains visible artifacts. This chapter explores an application of an edge-preserving filter known as bilateral filter for the fusion of hyperspectral image bands. We explain how the edge-preserving property of a bilateral filter enables us to extract the minor features in the data, and obtain a visually sharp yet artifact free fused image.