Abstract: Presented in this paper is an estimating method of multi\|feature and multi\|classifier combination based on the posterior probability estimators. Also presented is a method to extract effective discriminant features of handwritten digits based on a set of uncorrelated optimal discriminant features and KL transform. Experiments have been performed with Concordia University CENPARMI’s handwritten digit database based on the nearest\|neighbor distance classifier and the nearest\|neighbor correlation classifier, and 12 features of handwritten digits. Experimental results show that the estimating method is better than the polling method or the counting method respectively and the recognition rate of the estimating method is as high as 97%. A new combination classifier is finally brought forward, which is based on the strict structure classifier and the estimating method. Better experimental results have been obtained by means of this new combination classifier: the recognition rate, the reject rate, and the reliability are as high as 97.15%, 2.05%, and 99.18% respectively, which are the best results up to now on the handwritten digit database.