An Improved Road Network Segmentation Method Based on Multi-Level Texture Analysis with Regularization

The paper presents an improved method for automatic road network extraction based on multilevel texture analysis using regularization. The method is based on a three-level system of statistical texture analysis: the first level uses the Mahalanobis distance to estimate local features, the second (one and a half) level applies the modified Bhattacharyya distance to analyze distributions in the neighborhoods of pixels, and the third level implements an integrated approach through texture cubes and coincidence matrices. To improve the stability of the algorithm, regularization mechanisms, adaptive parameters and accounting for local texture features are introduced. Particular attention is paid to the post-processing stage, where a binary image is converted into a topologically correct vector representation with the elimination of artifacts and noise. The experimental results demonstrate a significant increase in the accuracy of road network extraction compared to existing methods, especially in rural and suburban areas. The practical value of the work is confirmed by the successful application of the method for processing satellite images and aerial images when updating geographic information systems. The developed approach allows to significantly reduce the time spent on data processing while maintaining high accuracy of the results.