Quad Tree Segmentation of Specular Imagery Via Besov Space Merge Criterion.

摘要

Certain specular types of sensor images (e.g., laser) contain vital information which is difficult to glean from non-specular sources. The present and increasing deluge of these types of images has created a critical need for image processing algorithms which reduce the workload for the image analyst by performing some of his/her functions automatically. Many of these algorithms are based on image segmentation a procedure having 1.) high separability between target object and background and 2.) low computational intensity implementation as two key goals. A large number of algorithms for automatic segmentation of images have been tendered, most occuring within the Mumford-Shah paradigm which uses approximation error, boundary length, and variance as weighted terms in an energy functional. The new idea of the present paper is to generalize the Mumford-Shah variance energy so that it directly measures the relative smoothness memberships of target and object background. This is especially important in the application to segmentation of specular types of images which tend to require the separation of subtle grades of smoothness and the unraveling of delicate smoothness space interpolations. The fast wavelet transform answers the purposes of efficient determination of smoothness membership at global as well as local levels and works well in a quad tree architecture.