A Framework for Combining Statistical and Structural Pattern Retrieval Based on Feature Histograms

Patterns are characterized by the distribution of features. In general detailed geometry of feature locations and statistics of features distribution are important for the characterization. We call these aspects structural and statistical information of patterns and aim for developing framework for the unified description of them. Statistical information can be simply and conveniently picked by feature histograms, structural information description is much more complex. In order to deal with it we are introducing the concept of hierarchical decomposition of pattern areas. Areas are described using statistical information by feature histograms, size and number of areas reflects structural information. This formulation unifies statistical and structural information and the problem of minimizing structural information is stated as reducing the number and size of the histograms. We illustrate this on an example of retrieval from face image database using features based on quantized block transform coefficients. We can show that very limited structural information is needed for nearly perfect retrieval performance equal to the best available algorithms.