Verification of Frequently Used Korean Handwritten Characters Through Artificial Intelligence

Abstract

AbstractHandwriting verification is a behavioral biometric that matches handwritten characters to determine whether it is written by the same person. Because each person has a different handwriting, it is used by investigative agencies for the purpose of presenting court evidence. However, it cannot be defined as a rule because the standards for visual reading of experts are ambiguous. In other words, different experts can make different decisions for the same pair. Therefore, we propose a handwriting verification method based on artificial intelligence that excludes human subjectivity. For 4 frequently used Korean characters, genuine or imposter pairs of the same character were trained with a Siamese-based ResNet network. The verification accuracy for the trained model was about 80%. Through this experiment, the objectivity of handwriting biometric through deep learning was confirmed, and a basis for comparison with verification performance through human eyes was prepared.KeywordsHandwriting verificationBehavioral biometricSiamese networkResNetArtificial intelligence