Handwritten connected digits detection: An approach using instance selection

Segmentation is a fundamental step in the process of handwritten digits recognition. However, it is common to have images with connected digits after the segmentation task and this affects the classifier accuracy. This paper presents an approach for handwritten connected digits classification based on instance selection. The new technique uses information from all data of the training set to build a ranking of the instances. The instances having the highest scores are chosen to represent the data points of the problem. A set of features especially designed for the problem is extracted. The experimental study using a real world database shows that the proposed technique is quite efficient in the detection of handwritten connected digits.