| Hyperspectral images(HSI)are classical high-dimensional data,which contain abundant image and spectrum information.HSI classification is an important content of feature extraction and object interpretation.Researchers can comprehend changes in agriculture,environment,and region according to the classification results and adopt corresponding measures.Potts model is a semi-supervised variation classification model,which presents a universal theory frame for data classification.And the Potts model can expand to solve high dimensional data classification problem.Meanwhile,this model can combine with the nonlocal method and constrained optimization method for achieving high precision and high performance of data classification,and improved the accuracy and computational efficiency for HSI classification.For constructing a new hyperspectral images classification method based on Potts model,the main research work is as follows:1.Based on the classical Potts model frame,a semi-supervised discrete nonlocal model is designed.Compared to the common Potts model,the proposed model has not data term which reduces the model complexity and improves operational performance.The designed model adopts a nonlocal operator to combine feature information of regions to structure weight matrix,which increases the classification accuracy of HSI.For HSI with a huge scale of data,the designed model can maintain steady operational efficiency.2.Based on the designed semi-supervised discrete nonlocal model,a discrete nonlocal model with a continuous max-flow method is designed.The model transforms into the continuous max-flow problem by importing dual variables.Therefore,the designed model has an efficiency of graph cutting and improves the accuracy and operation performance of discrete nonlocal Potts model.Due to the designed semisupervised don’t have data term,simplify flow conservation constraint to reduce the running time of designed model and improve the operational performance.3.For the non-convex optimization problem of the structured models,a corresponding alternating multiplier algorithm is designed.The process of solving models is decomposed into several suboptimization problems by importing auxiliary variables and adopted formulas to solve variables like the soft threshold formula.After each iteration computation,the label function is projected by the projection formula,so that the models can obtain the classification results with high accuracy and smoothness efficiently.In order to prove accuracy and performance,three classical semi-supervised classification methods are compared in the numerical experimentation and adopting five real hyperspectral data set to experiment.The experiments proved that the designed model can efficiently obtain accurate and smooth results. |
