The Application of Average Gradient Matrices for Fingerprint Classification using Neural Networks

Abstract

Currently the task of classifying fingerprints consumes much time and labour by trained fingerprint technicians. Manual fingerprinting storage and retrieval systems require a vast amount of human skill and expertise. Technicians who classify fingerprints have first to undergo several years training. Many law enforcement agencies use a fingerprint classification system known as the Henry Classification system in order to partition fingerprint databases. Using this system fingerprints are sub-divided into one of eight possible fingerprint pattern types. Numerous attempts have been made to automate such classification methods using conventional image processing techniques but very few have been embraced by law enforcement agencies due to their limited successes in solving the problem. The re-emergence of interest in neural networks (NNs) in recent years has caught the attention of those involved in the fingerprint field as they begin to recognise the potential advantages of a NN approach.In this paper a method of classifying fingerprints using Back Propagation (BP) NNs and average gradient matrices is presented. For this research we assembled a database of fingerprint images made up of 512 by 512 pixels and having 256 grey levels. A dynamic grey level threshold transformation was used in order to obtain a 512 by 512 bi-level image. As such an image contains a high degree of redundant pictorial information which is not required in any input data set to a neural network, the image was reduced to a matrix of varying sizes. This matrix contained the average direction of the ridge lines within a certain pixel subregion. The average directions were estimated by counting the occurrences of five types of “micropattern” which may occur in any group of four neighbouring pixels.Rather than having one large BP network for the task of classifying the fingerprints presented the network was decomposed. This means that rather than train a network to learn the classifications of perhaps seven classes of fingerprint, separate individual networks were trained to classify a particular class of fingerprint as opposed to another class or all other classes. The advantages of such decomposition are that the convergence-training time is greatly reduced and the results obtained are often superior to those for a fully composed network. Input to each BP network consisted of varying array sizes depending on the average direction matrix size chosen. Thus each neuron represented a fingerprint subregion average gradient. Each BP network contained one hidden layer which in each case had twenty neurons. The output of a particular network consisted of a single neuron. This neuron represented the degree to which an input fingerprint matched the ‘learned’ classification.Results of the our research have been encouraging. Using one part of our fingerprint database for training exemplars and the remainder for test data, various experiments have been conducted using various average gradient matrix sizes. The results being both noteworthy and surprising. It was found that no one matrix size yielded the best classification accuracy across all the classes. In the case of training a BP neural network to classify a whorl type fingerprint as opposed to any other sort of fingerprint it was found that the use of an average gradient matrix of size 32 × 32 produced the highest accuracy. This was also true for arches, however for loop type fingerprints the best matrix size was found to be one of 16 × 16. The results of this research confirm that using the average gradient technique is a suitable method to use with neural networks for fingerprint classification. The percentage of correctly classified test fingerprints has been found to be in excess of 90%.