Extraction of Signatures from Document Images for Real World Applications

Various automatic methods for signature verification have been reported in the recent past. A common issue with nearly all of these methods is that they are built on the assumption that signatures are available pre-segmented or pre-extracted from document images. Accordingly, such systems are refined and tested on data containing signatures on the foreground and very little, to ideally no other, information in the background. The authors argue that these settings are not realistic and, in reality, experts encounter cases where signatures are written on documents having a lot of other information, than just the signatures, e.g., machine-printed text, ruling lines, logos, etc. We believe, to better assist forensic experts, a system should have the capability to automatically extract/segment signatures from documents like bank checks, forms, bills, wills, etc. This paper identifies that automatic signature extraction/segmentation from document images is a relevant problem faced by forensic document examiners and compares the various approaches currently available. Furthermore, we present an alternative method for extracting signatures from document images. The proposed method is capable of distinguishing machine-printed text from signatures and is based on a well known local feature descriptor, Speeded Up Robust Features (SURF). We evaluate the proposed method on the publicly available Tobacco-800 dataset in order to compare it to the currently available extraction methods. The research found that the proposed method was able to find all of the signatures in the document images tested. The authors see the application of the proposed method as a tool for document examiners to automatically extract signatures from document images.