A Multiphase Level Set Method Based on Total Variation Density Estimation for Image Classification

Remotely sensed imagery with high spatial resolution often shows serious intra-class spectral variations and details disturbances. This leads to disadvantages on automatic image classification. To increase accuracy of classification, this paper presents a novel multiphase level set method by an optimization of probability density function(pdf) estimation using Total Variation(TV). Specifically, density estimation method using Total Variation originally from image denoising is introduced to well improve "roughness" of pdf caused by spectral variations and details disturbances. Then, the optimized pdf is used to improve Mansouri's model so as to alleviate local minimum solutions and to further increase classification accuracy. Evidential experiments on IKONOS, QuickBird-2 satellite imagery have demonstrated that our proposed density estimation method is very effective and robust even if in complex scene. Consequently, the improved multiphase level set model has yielded a great increase in classification accuracy. The classification result is more approaching to that of human vision interpretation.