Fingerprint recognition using self-orthogonal associative memories

This thesis investigates an automated system for the analysis and recognition of fingerprint images. The system involves software for image processing, feature extraction and fingerprint comparison. Extensive preprocessing of the digitized data is required before any classification can be undertaken. Therefore, enhancement, edge detection, thresholding and thinning of fingerprint image are used to transform the digitized image to a binary, skeletal image. From the processed print, features called minutiae are extracted via a multilayer perceptron classifier with back-propagation training algorithm. Selected features are represented in a special way such that they are simultaneously invariant under shift, rotation, and scaling The constructed feature vectors are then classified using a newly developed neural network called self-orthogonal associative memories (SOAM). Unlike the multilayer perceptron, SOAM does not has the bottleneck problem of slow learning rate. This network is trained by an orthogonal learning algorithm which can improve the performance of associative memories in both storage capacity and recognition capability under noisy environment. By combining this algorithm with different network architecture, incremental learning and forgetting of the memories can be achieved. Moreover, an efficient pruning algorithm is also proposed to further reduce the complexity of the network. Simulation results on random vectors and character recognition showed that the proposed network has memory capacity of about two to three times higher than that of traditional networks. When applying SOAM to fingerprint recognition, the system performed satisfactory for a small database of images.