Implementación y análisis comparativo de algoritmos de clasificación supervisados para identificar a propietarios de firmas manuscritas

Abstract

Financial institutions need to control the errors that the operational user may make when verifying the handwritten signature on documents such as withdrawal slips, checks, and loan promissory notes among others, that involve money transactions. In this work, a model based on convolutional neural networks is proposed to quickly and efficiently classify and identify a person's signature with an accuracy greater than 90%. Two datasets of signatures were used. The first one, entitled CEDAR available for public access, and the second collected by the researchers entitled GC-DB, composed of 121 local signers, who provided 45 signature specimens each one. Before classifying the images, image preprocessing was executed to eliminate implicit noise produced by the capturing device. A problem was found when capturing images on paper of different thicknesses, making the elimination of image noise to be a relatively complex operation. The proposed algorithm's efficiency was compared with two other algorithms that were implemented and validated with the same datasets. The results show that an efficient verification of the handwritten signatures can be executed in such a way as to exceed the established goal, and easy algorithm to implement.