A Fuzzy Approach to Complex Linguistic Query Based Image Retrieval

The current trend in the rapid growth of on-line image databases has brought forth several innovative approaches to content-based image retrieval. Most current techniques retrieve images based on an example image or object shapes/features extracted from images. Retrieval based on linguistic queries has not recieved much attention. In this paper, we present a fuzzy connective approach to handle complex linguistic queries consisting of multiple attributes. We represent each (fuzzy) attribute in a complex query by a (multi-dimensional) membership function. The degree to which an image satisfies the attribute is obtained by finding the membership value of the feature vector corresponding to the image in the membership function for the attribute. We propose the use of fuzzy connectives to combine the degrees of satisfaction of multiple attributes in a complex query to arrive at an overall degree of satisfaction. This overall degree is used to rank images for retrieval. The membership functions for the attributes can be generated in various ways. In this paper, we consider attributes such as “Bark-like” and “Water-like” which correspond to fuzzy categories in the images. In order to determine the membership functions corresponding to the fuzzy attributes we use a maximum-likelihood-based robust mixture decomposition technique to organize the given database according to categories that are inherent in the database. The robust mixture decomposition algorithm provides us with membership functions to represent the categories and these membership functions can be used to handle complex queries. We report experimental results on a database of 2384 texture images created from the VisTex texture database available at the MIT media lab. The database can be arranged into 16 categories such as Bark, Metal, and Water. Our results indicate that this approach can be used to facilitate meaningful image retrieval based on complex linguistic queries.