Real-time large-window binary filter design Academic Article uri icon

abstract

  • Optimal translation-invariant binary windowed filters are determined by probabilities of the form P(Y = 1|x), where x is a vector (template) of observed values in the observation window and Y is the value in the image to be estimated by the filter. The optimal window filter is defined by y(x) = 1 if P(Y = 1|x) > 0.5 and y(x) = 0 if P(Y = 1|x) 0.5, which is the binary conditional expectation. The fundamental problem of filter design is to estimate P(Y = 1|x) from data (image realizations), where x ranges over all possible observation vectors in the window. A challenging aspect of optimal translation-invariant binary windowed filters is the implementation for large windows. In the context of Bayesian multiresolution filter design recently published by the authors, the training requirements for an accurate prior are more stringent. As such the practical feasibility of the filter design becomes an issue. This paper discusses the real time issues for large window filter designs and how the bottlenecks were overcome to design practical large window multiresolution filters. The most crucial bottlenecks are the real memory required for training and the time required for training to obtain a satisfactory estimation of P(Y = 1|x) or its prior (in the context of Bayesian filter design) for large windows. Among other improvements a method for data representation is developed that greatly reduces storage space for the large number of templates that occur for larger windows during the training of the filter. Parallel algorithms are designed that reduces hardware related time loss during training. In addition we take advantage of Bayesian filter methodology to train for large windows. While the algorithm works for larger windows, we demonstrate the feasibility of Bayesian multiresolution filter design for window sizes of up to 31 31.

published proceedings

  • Proceedings of SPIE

author list (cited authors)

  • Kamat, V. G., & Dougherty, E. R.

citation count

  • 0

complete list of authors

  • Kamat, Vishnu G||Dougherty, Edward R

editor list (cited editors)

  • Kehtarnavaz, N.

publication date

  • April 2001