Offline text-independent Persian handwriting forgery detection using texture analysis

Nowadays handwriting is known and accepted as a behavioral biometric. Handwriting-based authentication has become a hot topic of research as a result of increasing handwriting forgery. In this research an offline text-independent method for detection of forgeries in Persian handwritten documents is presented. The proposed approach is based on extracting two dense textures of basic alphabet and some text details. A one class classification (OCC) with only positive (original) samples is used in classification together with Binarization and Skeletonization in the preprocessing phase. Also, a combination of Gabor filter and LPQ descriptor is used to complete the feature vector. Both descriptors have been used for writer identification and verification in English and Persian but this is the first time that the combination of these two is used on a Persian handwriting. We also introduce a new dataset from 62 individuals with 2 pages per each. In the first instance the writers were instructed to copy a preset text and in the second one they were asked to write at least 6 lines about a picture. These instances were to be used as training and testing samples respectively. In order to generate counterfeit samples, we asked 5 qualified calligraphers to duplicate the second sample of each writer. Three forgeries per writer were generated in this way. Tests results are satisfactory and present more than 84% accuracy, demonstrating the quality of the proposed method.