A supervised Lloyd algorithm and segmentation of handwritten Japanese characters

A generalization of a supervised Lloyd algorithm to multilevel image thresholding is provided, whose algorithm is an iterative and fast one based on a minimum weighted-squared distortion criterion. After deriving an important relation between weight coefficients in the criterion and convergent thresholds, we consider a procedure on a reasonable selection of the weight coefficients by a supervisor through segmentation experiments. Using these results, the segmentation of the ETL8 handwritten Japanese character data base is studied.