Normalisation of Confidence Voting Methods Applied to a Fast Handwritten OCR Classification

Abstract

Summary. In this work, a normalisation of the weights utilized for combining classiers decisions based on similarity Euclidean distance is presented. This normalisation is used by the condence voting methods to decrease the nal error rate in an OCR task. Dierent features from the characters are extracted. Each set of features is processed by a single classier and then the decisions of the individual classiers are combined using weighted votes, using dierent techniques. The error rates obtained are as good or slightly better than those obtained using a Freeman chain codes as contour representation and the string edit distance as similarity measure, but the complexity and classication time decrease dramatically.

Improving classification using a Confidence Matrix based on weak classifiers applied to OCR

“…The class that obtains the maximum average from among these values is selected. This method was chosen as baseline because of the good results obtained previously for this type of tasks [22]. In this previous work, the classification error rate was lower than 1-NN technique applied to each group of individual features (image, background, contour) and than a 1-NN technique gathering as input the three groups of features.…”

“…The same feature extraction as that detailed in [22] was therefore used. Nevertheless, a brief explanation is provided in the following paragraphs.…”

“…Each one of these groups has to be used with a different weak classifier, so that the final meta-classifier does not have to discover these underlying points of view by itself. Hence, our work establishes a case of study in which the CM representation is applied to the OCR task since this kind of data is known to allow several ways of extracting features [22].…”

Confidence voting method ensemble applied to off-line signature verification

“…Nevertheless, in most cases, the zero probability is not a realistic situation. A revised formula was proposed by Rico-Juan and Iñesta [12], and involved finding the nearest neighbor to every class, computing their inverse distances, and normalizing them to estimate their posterior probabilities. This way, every class has a non zero probability.…”