Offline signature verification using ordinal structure fuzzy logic and integrated features based on single signature

Abstract

An offline signature verification system (OSVS) is an industry-driven technology with the ability to verify and recognize a signer’s signature, as required for different situations such as performing financial transactions, undertaking security and identifying processes, and detecting fraud. In the OSVS field, substantial investigations have been undertaken mainly using a sizeable number of sample signatures available, from which a profile of an individual signer is constructed. However, very few studies have been undertaken regarding how a limited number of signatures can be used to build a signer’s profile. Furthermore, most of the previous works in the OSVS field have used isolated signatures to verify system performance, and there are very limited studies on signatures from documents, cheques and forms. This research developed a system, which supports the worst-case scenario where only one sample signature is available to build a profile. This system achieved accurate OSVS through which, one single signature is used to build the signer’s profile with different genuine signatures extracted from forms and cheques. Besides, different types of proposed forged signatures were evaluated using different techniques in the different stages of the system. This work was divided into two different stages called the adaptive representation module (ARM) and reliable verifier (RV). ARM starts by proposing a new adaptive binarization module (ABM) to isolate clear binary objects from the signatures embedded in the forms and cheques. ABM consists of a background-based estimation (BBE) stage that generates different greyscale images, zero-crossing thresholding (ZCT) technique which produces binary images, and fuzzy structured ordinal module (FSOM) designed by rules to select the best binary signature image with clear objects out of three nominated binary images. The second ARM module of is descriptors representation, which proposes generating two sets of features that distinguish signatures, including lines-based features and blob-based features. All the collected features are used to build a statistical feature vector to be applied later in RV. Next, the RV fused the distance-based and statistical verifiers to increase the accuracy of both FAR and FRR. The signature dataset for this research consisted genuine signatures embedded in forms, random signatures generated by signing simple names, unseen forgeries through signing known characters, and seen forgery signatures that simulated real signatures collected from the signer. Genuine signatures embedded into low resolution and noisy background forms were also generated to improve the efficiency of the adaptive offline signature verification (AOSV) system. The calculation showed low error rates for both FAR as seen in the forgery samples at 0.139 and FRR at 0.156. The findings have shown that researcher has successfully developed the an accurate offline signature verification system which uses one single signature to build a signer’s profile with different genuine signatures extracted from forms and cheques, as well as evaluate of different types of forged signatures.