Digital image processing: Offline handwritten signature identification using local binary pattern and rotational invariance local binary pattern with learning vector quantization

Abstract

The Handwritten signature is one of the authentication tools that humans have to determine the authenticity of a file or document. The difference between offline and online handwritten signatures is that offline signatures requires a scanner while online signatures use a stylus. The feature extraction methods used are Local Binary Pattern (LBP), uniform LBP, LBP 8 rotation, and uniform LBP 8 rotation. Learning algorithm used is Learning Vector Quantization (LVQ). This study used 200 images data, derived from 20 participants handwritten signature with contribution of 10 signatures each participants. Training data and test data will be divided based on the k-fold cross validation classification then tested based on 4 groups of data sets with certain parameters. Data set 1 consisted of 5 participants, data set 2 consisted of 10 participants, data set 3 consisted of 15 participants, and data set 4 consisted of 20 participants. The best system performance results with the highest accuracy are in training and testing of data set 1 with maximum of 100 iteration for each method. The results of system accuracy is very affected by total number of participants used, which is the more the number of participants with signatures processed, the accuracy will be reduced. The detailed results of system performances accuracy using data set 1 for the LBP extraction method obtained an accuracy of 84%, uniform LBP obtained an accuracy of 90%, LBP 8 rotation obtained an accuracy of 86%, and LBP 8 uniform rotation obtained an accuracy of 80%.