Translator Disclaimer
21 November 1995 Speckle reduction of SAR images using neural networks
Author Affiliations +
Given a high-quality, distortion free SAR image generated by a SAR processor, we would like to remove the speckle to reveal the underlying cross-section. Traditional filter-based techniques all suffer from the fact that the large filter size required for effective smoothing will cover a highly inhomogeneous region. A CPU intensive algorithmic noise removal process called simulated annealing has been successfully applied to both SAR intensity images and SAR texture images generated using a mean-normalized log texture estimator. To improve the execution time of the smoothing process, we have adopted a neural network based solution which emulates simulated annealing. A factorized neural network was chosen, consisting of a vector-quantizer first stage which is used to select a specific multi-layer perceptron from the second stage. This technique reduces both the training and run times for large neural networks. A further reduction in training times is achieved by the use of self-adjusting training algorithms. Statistical analysis of test data has shown that the network produces a good approximation to the estimated cross-section. Simulated annealing has the advantage of a much larger adaptive input window than the neural network, and a better comparison can be made by restricting simulated annealing to operate on a window with dimensions comparable with that of the neural network. A comparison with alternative techniques based on multi- dimensional lookup tables is also presented, comparing both the quality of the result with the execution time of the algorithm.
© (1995) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
David Blacknell, Christopher John Oliver, and Martin Warner "Speckle reduction of SAR images using neural networks", Proc. SPIE 2584, Synthetic Aperture Radar and Passive Microwave Sensing, (21 November 1995);

Back to Top