Stamos, Epaminondas; (2001) Algorithms for designing filters for optical pattern recognition. Doctoral thesis (Ph.D.), University College London (United Kingdom).

Text Algorithms_for_designing_filte.pdf Download (23MB)

Abstract

Matched filters for optical correlators detect the presence of objects immersed in white noise, but are unable to discriminate between similar, noisy input patterns. Also, the dynamic range of optical systems often limits the size of the images that can be recognised. We develop two algorithms for designing filters for optical pattern recognition. The first algorithm suppresses the similarities between the training images and creates a set of filters, which are mutually orthogonal to them. Our filters tolerate 7 dB more additive input white noise than matched filters and the required dynamic range is reduced by 25 dB. In addition, the filters obtained after only two iterations tolerate 2 dB more additive input white noise than linear combination filters (LCF), which results in an improvement in the probability of discrimination of about 30% for the same amount of noise. The correlation outer products for the 2 iteration Similarity Suppression (SS) algorithm are substantially lower than those for the LCFs. The second. Feature Enhancement and Similarity Suppression (FESS), algorithm designs filters for multi-class pattern recognition. Each of these filters can recognise all the members of a group and distinguish them from other groups. The probability of recognition for a training set of faces is 100% without noise, compared to 90% using matched filters and the required dynamic range is again reduced by 25 dB. We prove the mathematical equivalence between these algorithms, the back-propagation algorithm for training neural networks and the method for designing general synthetic discriminant functions (SDF). Our algorithms also design filters for two or more cascaded banks of correlators and can train multilayer neural networks. Conversely, matrix inversion methods, which are generally used for designing SDFs, can train neural networks and give the same results as obtained with the back-error propagation algorithm.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item