This paper presents an innovative instance selection method, called Instance Selection Algorithm based on a Ranking Procedure (ISAR), which is based on a ranking criterion. The ranking procedure aims to order the instances in the data set; better the instance higher the score associate to it. With the purpose of eliminating irrelevant instances, ISAR also uses a coverage strategy. Each instance delimits a hypersphere centered in it. The radius of each hypersphere is used as a normalization factor in the classification rule; bigger the radius smaller the distance. After a comparative study using real-world databases, the ISAR algorithm reached promising generalization performance and impressive reduction rates when compared with state of the art methods.

After the handwritten segmentation process, it is common to have connected digits. This is due to the great size and shape digit variations. In addition, the acquisition and the binarization processes can add noise to the images. These under segmented images, when given as input to classifiers which are specialists to deal with digits separately, should lead to errors. Aiming to detect the handwritten connected digits, it is herein introduced a hybrid system architecture to be used as a segmentation pos-processing task. The proposed system is based on a prototype selection scheme that combines self-generating prototypes and Gaussian mixtures. Besides, this work presents a set of features for the proposed problem. A real-world database of handwritten digits was used to validate the new approach. The results obtained in the experimental study showed that the hybrid strategy achieved promising accuracy rates.

No abstract

No abstract