Abstract:
The aim of our work is to provide an offline signature authentication system that help
bankers to verify whether the signature performed by the customer is genuine or
forger. The point that the sign is extensively used as resource for person verification
stresses the necessity for a verification system. There are two major ways to perform
verification that are following:
• Online systems
• Offline systems
Both systems vary in various aspects. If we talk about online systems so one thing is
common in all online systems that all of these process use active aspects of a
signature that is captured during the moment signatures are performed by the
individual. On the other hand, scanned image is the essential element in offline
system that makes it to perform work.
Our group work objective is to achieve an offline sign verification/authentication that
is based on transfer learning. This report consists of different techniques that we have
applied in order to provide best accuracy results. Techniques and steps like data
acquisition, pre-processing, feature extraction, classification and verification is
explained in detail. Furthermore, inception v3 and mobile Net are the major
highlighting features of our project, these enhances accuracy of the result by using
different layers in order to recognize other scanned images of the signatures
performed. The advantage of using these features with transfer learning is this that
inception v3 is playing a major role over many years in order to provide best
accuracy result by working on image dataset features that are extracted.
Moreover, modelling and training is also part of this system. Training is performed
by softmax layer that is assumed to be normalized as N + 2048*N (or J001*N) model parameters. Beside this all, report also consists of final recommendation and
conclusion that is : consisting ofreasons to use all ofthe above techniques.
