In the past several years, wepsilave been developing a high performance two-layer SOM classifier for handwritten Bangla numeral recognition. We have reported previously the structure and the learning algorithm of the two-layer SOM. In this paper, we fuse the outputs of the second layer SOMs to improve the classification ability, based on confidence coefficient of the SOMs in the second layer. Experimental results on the numeral images obtained from real Bangladesh envelopes have proved the validity of the proposed method.