Cambridge Neurodynamics Ltd, Cambridge, UK, has brought its experience in image and signal processing and pattern recognition technology to create the world’s first neural network automatic fingerprint recognition system for the South Yorkshire Police (CI No 2,676). Data General Ltd provided the Windows-based workstations for civilian fingerprint experts and police officers to access a 20Gb AViiON Unix server containing names and addresses and 100,000 compressed fingerprint images, a thirtieth of their original size. The JPEG compression is to UK Home Office specifications but Richard Gaunt, technical director of Cambridge Neurodynamics, said the company could achieve much more compression with its proprietary methods. The database is also linked to other computerised police systems such as the crime management system for records, general administration, command and control and the custody system. Each workstation has printed circuit boards for compression, a document scanner and feeder which enable the fingerprint expert to scan A4 fingerprint forms and photographs of, or actual, scene-of-crime prints. The workstations can also capture images from live video of the scene of the crime. The expert sees two windows on-screen: one interrogates the server’s database and the other enables the expert to zoom, rotate and generally manipulate (but not tamper with or alter) the unidentified scanned print. The on-screen functions assist the expert in assessing the match of 16 points of similarity against a selection chosen by the system, in order of merit. Crucially, according to Cambridge Neurodynamics, the machine does not make the final selection, the expert does, which means the evidence is acceptable as scientific by the court.
Distorted, smudged, unclear
There is no maximum to the network configuration. However, the size of the image database affect throughput. The latter depends on how many sets of compressed 10-print records it can search. Two types of units are plugged into the South Yorshire network server: a single encoder and a single matcher. Gaunt said that there was no limit to how many such units could be linked; the more there are, the more the throughput. The encoder extracts the detailed information from the image, the matcher stores it and searches for comparison under the control of the database. There is no comparison based on images, as in an optical jukebox system, or older file systems, for example, said Gaunt. The unknown image looks nothing like the original print taken in the police station. Those from the scene of the crime particularly can be distorted, smudged, unclear or superimposed. The matcher searches through the database for minutiae – 16 point of comparison – that it has been ‘trained’ to seek. It also looks for additional information in a probabilistic way to cut down on database search time. The search is unseen by the user, who can continue working. Gaunt said the basic speed of search is 15 minutes for a typical ‘scene of crime’ but pointed out that by comparison with other systems, it is not speed but accuracy that is the issue. A manual system is well organised so that the database can be split and subdivided. But if the force wants to do a murder enquiry where the whole database is searched for an unknown print, it would take an entire detective bureau a week to do it manually. Most searches are not blind, the officers have some idea of who they are looking for, Gaunt said. Physically, the encoder and matcher each contain a processor board. One processor handles one print, so 50 processors for example, will handle 50 prints, therefore there will be 50 times the throughput. The boards hold five Texas Instruments Inc TMS320C50 processors which, Gaunt said, are very suited to image processing since the application itself is very similar to a signal processing application. The software containing the algorithms directs the sequence of processes and is ‘trained’ in simple terms with thousands of prints passing through and a human expert guiding it. For each of the
functions in its operation, it reverses the process and tells the expert where a particular feature is. It does not ‘learn’ by experience in the way claimed for other neural network systems. Cambridge Neurodynamics is small but profitable, says Gaunt. It specialises in developing systems for banking, social security and immigration. Anything Gaunt says, where complicated character reading is required. The system for recognition of images is generic so we can apply it to any number of specific markets. In particular, he added, the container ships market where handwritten identification marks are written on the container sides has huge potential.
