An adaptive fuzzy approach for modeling visual texture properties

The analysis of the perceptual properties of texture plays a fundamental role in tasks like semantic description of images, content-based image retrieval using linguistic queries, or expert systems design based on low level visual features. The presence of these properties in images is very difficult to characterize due to their imprecision, and, moreover, because their perception may change depending on the user or the image context. In this paper, texture properties are modeled by means of an adaptive fuzzy approach that takes into account the subjectivity of the human perception. For this purpose, a methodology in two phases has been proposed. First, non-adaptive fuzzy models, that represent the average human perception about the presence of the texture properties, are obtained. For this modeling, we propose to learn a relationship between representative measures of the properties and the assessments given by human subjects. In a second phase, the obtained fuzzy sets are adapted in order to model the particular perception of the properties that a user may have, as well as the changes in perception influenced by the image context. For this purpose, the membership functions are automatically transformed on the basic of the information given by the user or extracted from the image context, respectively.