Measuring classification complexity of image databases: a novel approach

摘要

A novel approach for estimating classification complexities of various image databases holding similar types of data is proposed. A new measure indicating the relative complexity of classification for an image database has been defined in terms of the statistical similarities existing between various samples belonging to various classes in the database. The proposed measure has been applied to five widely available databases containing samples of handwritten and machine printed characters and has been found to be very accurate in predicting the relative classification complexity of these databases. The proposed complexity measure is completely generalised and is applicable to the assessment of the relative classification complexity of any image database relative to any other databases containing similar types of images.