|
book reviews
books on image forensicsreviewed by T. Nelson |
|
|
book reviews
books on image forensicsreviewed by T. Nelson |
Reviewed by T. Nelson
In this age of defensive science we all have to take time out from curing diseases to become experts on image forensics. This short book is a great introduction. After a brief discussion of how to identify which model of camera was used (evidently a big deal in legal cases), the authors focus on copy-move forgery, which they say is the most common type of image manipulation.
According to the authors, the most important tools in image forensics are principal component analysis (PCA), discrete cosine transform (DCT), and dyadic wavelet transform. They describe the algorithms clearly using math that's familiar to anyone who does image analysis, though no source code is provided. They claim the discrete wavelet transform algorithm is the best and that it has a false-positive rate below 2% and an accuracy over 99%.
More research is still needed to find algorithms that aren't fooled by rotation, scaling, or filtering. A weakness in this book is that their test images are mainly color pictures of everyday objects like ducks and ‘Caution Wet Floor’ signs rather than scientific images. There is no coverage of convolutional neural networks.
A good, understandable description of PCA and (just as important) what to
do with the gigantic pile of numbers it dumps on your screen can be found
in An Introduction to Applied Multivariate Analysis with R by Brian
Everitt and Torsten Hothorn. Everitt and Hothorn use the princomp
and prcomp commands interchangeably; some other R books use
principal, which is in the psych package.
Another book is Generalized Principal Component Analysis by Vidal, Ma, and Sastry. This one is oriented toward image analysis and has chapters on robust PCA, spectral methods, and segmentation, though its 566 mathematical pages can be tough sledding.
sep 17 2022