The Directional Neighborhoods Approach to Contextual Classification of Images from Noisy Data
提出一种有向邻域方法，通过选择最优非对称但相对同质的邻域对像素进行上下文分类，蒙特卡洛模拟表明该方法显著优于参考贝叶斯分类。
Abstract The directional neighborhoods approach (DNA) to classifying pixels and reconstructing images from remotely sensed noisy data is a newly proposed computer-intensive procedure that is partly Bayesian and partly data analytic. It uses the observational data to select an optimal, generally asymmetric, but relatively homogeneous neighborhood for contextually classifying pixels. A criterion for “homogeneity of neighborhood” is developed. DNA involves two stages: a zero-neighbor preclassification stage, followed by selection of the most homogeneous neighborhood, and then a final classification. We provide Monte Carlo simulations for a two-population image and compare DNA results with those from a reference Bayesian contextual classification. We show that DNA improves substantially on the reference classification procedure.