Abstract:
Hyper-spectral imaging is a new and novel way of capturing images from invisible parts of the electromagnetic spectrum. In this project, I explored and researched ways to use this technology for image segmentation, image and document verification. This is a very research-intensive project, but it has a marketable side to it in the form of a finished product developed for large corporations, banks, and government, who can use it to perform forensics on official documents and reduce losses due to fraud and counterfeit documents.
For my research, I used two different exclusive non-public datasets of hyper-spectral images. There are multiple problems that I tried to solve in the domain of document verification using HSI. One is the problem of segmentation of handwritten signatures from documents. This has proved to be especially difficult for traditional algorithms to solve in cases where the signature overlaps with some other text, which has become a serious hurdle in signature verification, and in other fields which rely on segmentation as a pre-processing step. I have attempted to solve this problem through various image processing algorithms which I strung together in a specific order to form my own algorithm. The basic step that enabled this was my algorithm of spectrally un-mixing the signature and the printed text.
I have also developed an algorithm which uses HSI images of cheques to determine whether or not the cheque has been written with ink from two different pens, or at two different times. This has been achieved by applying different image processing algorithms in a specific sequence in order to make a judgement on the cheque with surprisingly reasonable accuracy. The program is designed to be used by banks on every cheque that is honoured so that it can alert the bank teller whenever any specific cheque is determined to be suspicious. This data is very useful in forensic investigations of cases where, for example, a person accuses the bank of allowing a stolen cheque to be honoured.
