Determining the Origin of Digital Imagery

weakassuredΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 4 χρόνια και 1 μήνα)

74 εμφανίσεις

Determining the Origin of Digital Imagery

Project Proposal

Supervisor:

Dr Darrel Greenhill


Abstract

Digital cameras such as mobile phones, webcams, consumer cameras or CCTV cameras have be
-
come widely available in recent years. Supporting technologies have ensured that acquisition,
storage, distribution and visualisation of enormous volumes of video mate
rial is within the reach of
all. Consequently such imagery is likely to play an increasing role as evidence in police inves
-
tigations
or for courts of law. This proposal aims to develop a method of establishing the origin of digital
imagery


specifically,

whether a particular video sequence was acquired from a particular camera
based on the unique intrinsic pixel
-
defect pattern associated with every manufactured camera. The
key challenge addressed by this proposal is the recovery of the pixel defect patter
n buried beneath
the layers of interpolation, white
-
point and gamma correction, and compression post
-
processing.


Aim and Objectives

The Home Office five
-
year strategic framework aims “to ensure that the police service is equipped to
exploit opportunities
in science and technology to deliver effective policing” through the
identification of priority police service capabilities to which new technologies are likely to make a
significant impact. The goal of this project is to develop the technology to support
a specific
investigative capability: providing an evidential link between a video sequence and the camera that
acquired it. The key issues which require addressing can be summarised as the following scientific
objectives:




To characterise CCD ‘fingerprint
s’ and support expert witnesses by providing statistical
evidence about image origin.



To develop a model of the generic digital acquisition process, and algorithms to both
parameterise this model given access to the target camera and to recover the camera
’s
pixel
-
defect ‘fingerprint’.



To develop a method for recovering a probabilistic pixel
-
defect pattern from a digitally
-
acquired target video sequence given a hypothesised model of the camera acquisition
process.



To develop and validate a probabilistic m
atching scheme (capable of supporting a robust and
statistically defensible forensic test) which will establish the likelihood that any specific video
sequence could have originated from a specific camera.


Methodology

The main issues in forensic image sci
ence relate to image integrity, authenticity and camera
identification. In standard film photography, forensic scientists attempt to prove the origin of
photographs by looking for evidence of camera imperfections. These include scratches on the film
caused

by the camera mechanism when an image is acquired, or changes to the image caused by the
characteristics or defects of the lens. Equivalent approaches for the digital domain could make use of
features of digital cameras such as the CCD sensor, the colour
filters, digital storage and so on.

Individual cameras vary but the structure of the digital acquisition process is relatively standard.
Light enters through the lens and first passes through a set of anti
-
aliasing filters. Passing through
colour filter a
rrays (CFA), different wavelengths fall onto an array of millions of CCD sensors which
measure the intensity of light arriving at each pixel. The charge pattern is then read out through shift
registers. Rather than having a CCD for red, green and blue, to
reduce manufacturing costs, sensors
for each colour channel are interlaced such that each pixel detects only one of the colour channels.
Interpolation techniques are then used to calculate the missing colour values. As well as
interpolation, each particula
r camera model performs a number of other image processing
operations on the image before it becomes available to the user. These include white
-
point
correction, gamma correction and often compression. If the proprietary techniques used by different
manufa
cturers can be characterised from captured images, it may be possible to parameterise the
unknown camera model and hence identify the specific camera.


Relatively little work has been carried out in this important forensic area. A number of possible
chara
cteristics of this process could be useful in generating unique camera signatures. An important
feature of CCD devices is the ‘dark current’. Normally charge flows roughly proportional to the
amount of light hitting the sensors. However, even in complete d
arkness a small amount of current
still flows. This is non
-
homogeneous in nature, as some pixels have more or less dark current than
others. Given access to the camera, the pixel distribution of this dark current can be recovered by
averaging several hundr
ed frames together of images taken in darkness. A second feature of such
manufactured devices is the inevitable presence of pixel defects. The unique spatial distribution of
such defects could be used in camera identification. Some cameras have on
-
board pr
ocessing to
eliminate these pixels. Although normally found on more expensive cameras, such technology will
become more widespread. The most common types of pixel defects include dead pixels, which have
poor response; cluster defects where groups of pixels

are faulty and hot point defects where the
pixels are very bright. As a complication, it has been shown that the distribution of pixel defects
changes according to the temperature of the camera. An alternative approach has been to extract
discriminatory c
olour and image quality features train a classifier. Features include the correlation of
pairs of the RGB colour bands and statistical features calculated using wavelets. Camera models (or
even individual cameras) produce images of dif
-
ferent quality as me
asured by darkness, sharpness
and colour quality. No work has previously been done which measures how defects vary across time,
for example a defect in a green channel for a particular pixel may be intermittent. A number of time
-
based metrics could be esta
blished, and statistical techniques such as discriminant analysis used to
identify the best set of features.


In the US, the recently funded “Non
-
Intrusive Forensic Analysis of Visual Sensors Using Output
Images” project aims to identify technologies empl
oyed inside digital cameras based on their output
images. Their aim to characterise the acquisition process is similar to our but they wish to address
technology infringement. We propose to utilise the spatial pattern of pixel defects to generate a
unique
camera ‘finger
-
print’. The post
-
processing operations which are applied to the acquired im
-
age, combine to intentionally suppress any visual error. The solution, therefore, is the inverse
problem of retrieving the pixel defect pattern given the final image

and the unknown acquisition
pipeline. Initially the project will focus on the problem of forensically linking video sequences
(contiguous sets of images) to a camera rather than the more challenging problem of linking a single
frame.


Our approach will b
e to develop a parameterisable generic model of the digital acquisition proc
-
ess.
This model should capture the range of propriety post
-
processing methods applied to an acquired
image as well as the final compression algorithm. A procedure for both paramet
erising this model
and establishing the pixel defect pattern using the target camera will be developed based on the
acquisition of fresh video under controlled lighting conditions using diagnostic im
-
age targets. (It will
also be determined whether the def
ect pattern could be recovered more di
-
rectly by dismantling the
camera.) Having established an appropriate model for the target cam
-
era, a second characterisation
procedure applied the target video sequence will attempt to recover the pixel defect pattern

using
the camera model. The inevitably probabilistic spatial pattern will need to be matched to the pixel
defect pattern recovered from the camera and a statistical probability of match determined.




Vacancy Type

PhD

Closing Date: August 27th 2011


How
to apply

You are strongly encouraged to discuss this project with
Dr D. Greenhil
l
, the Director of Studies
before applying.


Please return the completed Postgraduate Application Form together with an academic CV, TWO
references, academic certificates and any other documentation to Research Admissions, Faculty of
Science, Engineering and Computing
, Sopwith Building, Kingston Univ
ersity, Penrhyn Road, Kingston
upon Thames Surrey, KT1 2EE


or by e
-
mail to
CISMResearchNB@kingston.ac.uk


Please download the application form from:
http://www.kingston.ac.uk/postgraduate/apply
-
now/

Any queries can be directed to the above e
-
mail or by telephone on +44 (0) 208 417 2697

Please assume that your application has not been successful if you have not heard from us 6 weeks
after t
he closing date.