A measure of recognition difficulty for a character image database

A character image database plays an important role not only in the development stage but also for the evaluation of a handwritten character recognition system. Such a database is obtained from outside or customly made. At this point there is no measure which tells the level of recognition difficulty of a given database. If such a measure is available, we can use it in many useful ways. Especially, it will be valuable when we compare and evaluate the performance results of various systems since different databases whose recognition difficulties are unknown are usually used. In this paper we propose such a measure. We first define the entropy of a point of an image in the database. Then we obtain the measure by applying some normalizing factors to the entropy mentioned above. Note that such a measure should be used to compare the recognition difficulties of databases only of the same character set. We show that the proposed measure can be used for databases not only of different numbers of images per class but also of different image sizes. Finally we confirm that the proposed measure really reflects the relative recognition difficulties of databases by using real databases