A new descriptor for person identity verification based on handwritten strokes off-line analysis

Abstract

A new descriptor for the verification of people’s identity through the analysis of handwritten text is presented. The proposed descriptor corresponds to a representation of the pattern of writing pressure computed from the grayscale image of a handwritten stroke. Specifically, the descriptor corresponds to the relative position of the minimum gray value points within the stroke. A repository of images for 50 people was created. Each person wrote 50 samples of 6 different symbols which resulted in a total of 15,000 images to carry out the experiments. For each individual’s identity verification, a supervised classifier for non-linearly separable data of the Support Vector Machine type was used, which resulted in the training of a total of 50 classifiers. 50 groups of balanced data were created through the sub-sampling of the majority class for the proper training of the classifiers. Furthermore, K-Fold Cross Validation was used to assess objectively the descriptor performance. The results of the assessment are positive: a hit rate average higher than 95% was achieved for the six analyzed symbols to verify identity. The overall proposal of the paper is interesting because it presents a method based on the processing of very simple characters (the characters are notoriously simpler than a signature). The proposed descriptor has the advantage of being invariant to rotation, which makes the process robust to involuntary changes in the inclination of the sheet containing the strokes. Besides, the descriptor is invariant to scale, as it considers the obtained sign length resizing. This makes the process robust to characters written with different sizes.