Learning human perceptual concepts in a fuzzy CBIR system

In this study, we propose a fuzzy logic framework for content-based image retrieval (CBIR) to achieve the goal of personalization and better retrieval results. Under the proposed framework, typical problems in CBIR such as the semantic gap and the perception subjectivity can be alleviated. Three major components, including: (1) image representation; (2) query expression; and (3) feature matching, are discussed and solutions are introduced. For the image representation, we formulate a mapping from low-level numerical features to high-level linguistic terms by the use of fuzzy membership functions. For the query expression, we define a query description language that provides a flexible query expression for users to specify their information need at various semantic levels. For the feature matching, our CBIR system can construct a unique personalized similarity function that measures similarity between the query and an image according to the user's query and his/her preference. Experimental results are given to show the effectiveness of our CBIR system.