Forensic Image Analysis

Some Results...


Image Denoising

Here are some results from a very poor surveillance video of a holdup. Technically I am very pleased with the results, but they are not good enough to obtain a conviction.

The denoising technique is a new one that I have developed. It is an adaptation of standard wavelet denoising techniques but it uses non-orthogonal wavelets, and unlike existing techniques, ensures that phase information is preserved in the image. Phase information is of crucial importance to human visual perception.

Please note that the image reproduction obtained through your web browser may not be optimal - the original images are probably a bit better than what you see here. It helps if you stand back a bit - don't look up close; the dithering and quantisation in the display tend to mask what you want to see.

Image 1

The face of one of the holdup men, his partner is behind him with a coat pulled up over his head.

Original Image - a bit dark...

Original image with grey scale adjustment to enhance the dark regions. Note that in fact this is two images in one. The television transmission standard results in images being displayed in an interleaved manner with one field of the image displayed in the even numbered rows of the image and the other field in the odd numbered rows. These two fields are taken roughly .16 seconds apart (on a video recoreder in 24 hour mode) so if there is significant movement we will have a confused result.

This is an image constructed just from the odd numbered rows of the image. The extra rows were made up by averaging the row above and below each one we needed to fill in

This is the image constructed just from the even numbered interleaved rows. Notice how his head is in a slightly different position and his partner behind is in a different position relative to the image on the left

A denoised version of the image above with grey scale adjustment.

The denoised version of the image above.


Image 2

One of these guys has a distinctive emblem on the back of his coat...

Original image with grey scale adjustment to enhance the dark regions. Again, note that this is two images in one. In this particular image the interleaving of the two fields produces a very confused result.

This is an image constructed just from the odd numbered rows of the image. The extra rows were made up by averaging the row above and below each one we needed to fill in. Note how this has greatly clarified the image

This is the image constructed just from the even numbered interleaved rows.

A denoised version of the image above with grey scale adjustment.

The denoised version of the image above.

Averaged Image This image was obtained by averaging 130 frames of the image sequence. This very effectively removes the noise - but it only allows you to see the static background regions of the scene. Temporal averaging is not useful for improving details on the holdup men because they are constantly moving.


Image Deblurring

There are a number of image deconvolution techniques. We have had most success with the Richardson-Lucy algorithm. It is fast and relatively insensitive to noise in the image. The algorithm does require you to provide an estimate of the point spread function. However, we have found that even if the estimate of the point spread function is inaccurate useful results can still be obtained. These images were prepared by David Leow as part of his honours project on image deblurring.

Original Image.

Motion blurred image.

Image reconstructed from the blurred image .


Fingerprint Enhancement

Fingerprint image

Enhanced fingerprint image.

Rather than try to enhance the ridge pattern I have found it easier to enhance the groove pattern. The grooves in the fingerprint tend to be more consistent and less broken up. The algorithm used here involves searching for regions of phase symmetry. This works on the assumption that the grooves form lines of local bilateral symmetry. Measuring local symmetry using phase information provides a dimensionless measure of local groove symmetry that is largely invariant to image contrast and brightness.

See: Peter Kovesi, "Symmetry and Asymmetry From Local Phase" AI'97, Tenth Australian Joint Conference on Artificial Intelligence. 2 - 4 December 1997. http://www.cs.uwa.edu.au/pub/robvis/papers/pk/ai97.ps.gz.


Return to: