Multiscale detection of linear features in speckled imagery

We are concerned with multiscale extraction of linear features in speckled images. The methodology combines the a trous algorithm, Hough transform and the line detectors D1 and D2. The noisy image is decomposed into several scales by applying a redundant discrete wavelet algorithm (a trous) and the detectors D1 and D2 extract in each scale the straight lines. An associative symmetrical sum is used to merge information from both D1 and D2. In order to supress responses due to small isolated structures, a local Hough transform is applied to each scale and the best scale results are associated by using the symmetrical sum again. Real SAR images were used to test the algorithm.