Instance Reduction Techniques for ِArabic and Hindu Digit Recognition

Khalil M. el Hindi, Eman F. Issa, Areej M. Abu Kar


Instance-based learning is competitive, in terms of classification accuracy, in many applications to more sophisticated learning techniques such as neural networks. However, instance based learning requires storing a large training set which requires large storage space and hence long classification time. Several general instance reduction techniques have been developed to deal with this problem. This work presents several instance reduction techniques that are designed for the digit recognition problem. The techniques construct a set of prototypes, which are retained instead of the complete training set. We also present a distance function that is used to measure the distance between an instance and a prototype. The new methods are compared with some of the best-known reduction techniques on the handwritten digit recognition problem. The results show that the techniques presented provide the best combination of amount of reduction and classification accuracy for both Arabic and Hindu digits.


Instance-Based Learning, Instance Reduction, K Nearest Neighbors Algorithm, Hand-written Digit Recognition

Full Text:



  • There are currently no refbacks.