Selecting scales for texture models

Texture models are widely in use for image content description. In remote-sensing images textures occur at very different scales, requiring the application of several texture models. This paper presents a scale selection algorithm based on a multi-scale random field model with a dynamic pyramidal structure. The algorithm returns the scales and the regions where texture is present. It is designed as a fast preprocessing step in texture analysis. Since it allows the restriction of texture analysis to the interesting regions and scales, it makes the analysis faster as well as more robust. Additionally it minimizes the number of features to compute. These are critical issues in the design of large remote-sensing databases with content-based retrieval services.