D. Zhang and A.K. Jain (Eds.): ICB 2006, LNCS 3832, pp. 713 – 720, 2005. © Springer-Verlag Berlin Heidelberg 2005 A Novel Strategy for Designing Efficient Multiple Classifier Rohit Singh 1 , Sandeep Samal 2 , and Tapobrata Lahiri 3,∗ 1 Wipro Technologies, K-312, 5th Block, Koramangala, Bangalore – 560095, India rohit.singh@wipro.com 2 Tata Consultancy Services, Bangalore sandeep.sam@tcs.com 3 Indian Institute of Information Technology, Allahabad – 211012, India tlahiri@iiita.ac.in Abstract. In this paper we have shown that systematic incorporation of decision from various classifiers following a simple decision decomposition rule, gives better decision in comparison to the existing multiple classifier systems. In our method each classifier were graded according to their effectiveness of providing more accurate results. This approach first utilizes the best classifier. If this classifier classifies the test sample into more than one class or fails to classify the test data then the feature next to the best is summoned to finish up the remaining part of the classification. The continuation of this process, along with the judicious selection of classifiers, yields better efficiency in identifying a single class for the test data. The results obtained after the experiments on a set of fingerprint images shows the effectiveness of our proposed classifier. 1 Introduction Personal Identification systems based on fingerprints or facial images, diagnosis of diseases by analyzing the histopathological images, etc. are some applications where accuracy cannot be compromised with, as it may be a case of identifying an authorized person for access to critical or highly restrictive places, or it might be the case of saving the life of a patient through proper diagnosis. More often it is seen that a single classifier struggles to give a high accuracy and reliability level that some critical applications demands. As a result of this, a multiple classifier can be a viable solution for the accuracy and reliability constraints. Work has been going in this field from last decade. From the point of view of analysis, the classification scenarios can be of two types. In the first scenario, all the classifiers use the same representation of the input pattern. Here each classifier produces an estimate of the same aposteriori class probability. In the second scenario each classifier uses its own representation of the input pattern. They can be either sequential or pipelined [1], [7], or hierarchical [8], [9]. Other studies done in the gradual reduction of the set of possible classes are shown in [3], [4], [6]. The combination of ensembles of neural networks (based on ∗ Corresponding author.