Automated forensic ink determination in handwritten documents by clustering

Even in today`s highly digitalized world, the use of handwriting is still widely in use for legal documents such as testaments, contracts, bank cheques or professional certificates. Thus, forgery analysis of handwriting still poses challenges for criminalistics forensic document examiners and one of the investigation questions is, if a questioned document has been written with more than one ink. If so, this may indicate a forgery by manipulations of a possible counterfeiter after production of the genuine original. By means of chemical analysis, it is possible today to identify an ink with almost 100% certainty. However, this process is manual, tedious and needs an initial suspicion by a human expert, that more than one ink was applied on a specific area on the document. Further, most chemical approaches are destructive and limited to very small areas. To improve and to automate the initial investigation on the use of multiple inks, this work proposes a pattern recognition approach based on signal processing, feature extraction and classification by data clustering, which is based on spectral imaging, acquired in almost non-destructive and contact-less manner. The goal is to support forensic examiners by an automated digital detection of regions, which have been written using different ink, which they then can further examine. For experimental evaluation, a benchmark is introduced to evaluate the accuracy of detection results on a specifically created test set, which is also presented. Test results indicate that the best clustering in our investigation has been achieved by the expectation maximisation (EM) approach, with a correct ink cover rate of above 80% for the first and 73% for the second ink in average. Even more relevant for forensic experts is the observation, that false detections occurred in less than 1% of the cases in average. Future work will include extension of data sets and automatic analysis and parameter adjustments in the clustering process.