Identifying the Characteristic Scale of Satellite Images with Scale Space Representation and Normalized Derivatives

摘要

This paper presents an approach that identifies the characteristic scale of satellite images. It consists of three steps: (1) calculating the scale space representation of satellite images using convolution with the Gaussian kernel; (2) determining the local characteristic scale of every pixel by a local extremum over scale of normalized derivatives; (3) deriving the local characteristic scale histogram and determining the characteristic scale of satellite image based on maximum of this histogram. The local characteristic scale histogram reflects the local texture and geometric features while the characteristic scale represents the global salient structures of satellite images. Experiments on real data show that our proposed characteristic scale accurately identifies global structure of satellite images. We use the normalized Laplacian operator to determine the local characteristic in our experiments. Our proposed characteristic scale shed light on the structure understanding of satellite images. Moreover, the proposed local characteristic scale histogram can be extended to represent image structure for further scene classification of remote sensing images.