Abstract In 2003 AlBinHassan and Marfurt explained that the Hough transform can be used to detect fault lines in 2D seismic data. Their conclusion was that “The Hough Transforms method is a good candidate for enhancing lineaments associated with fractures and faults.” But, they also observe that “the problem with the running window code is that the new code will be more sensitive to noise such that the image may not be as clear as we wish.” Another problematic issue with the windowed Hough transform, though not discussed in the AlBinHassan’s paper, is that the run-time is not very favorable for long operators (i.e. , long line segments). In general, if an image contain m*n pixels, and we search for line segments of length k , the runtime complexity will be in the order of m*n*k*k computational units. In practice, this means that if we double the scan line length, the run-time will be four times as long. Clearly, if we look for very long fault lines for every 2D section in a 3D seismic attribute volume, run-time will be an issue. The Hough transform is generally known to be very similar to the Radon transform (it can be convincingly argued they are equivalent), which in turn is closely related to the Fourier transform, so any issues observed with the Hough transform will have the same issues in a Radon- or Fourier-based approach, although Fourier can sometimes be more computationally efficient through the use of the FFT. We have over the last year tried to overcome these two problems for the windowed Hough transform. Specifically, by improving the transform slightly and by calculating varying properties along the scanned line segments, we manage to make the approach more robust in the presence of random noise. We have also found a twist to the algorithm which makes the run-time linearly rather than fourth-power dependent on the line segment length, giving rise to an algorithm that can be applied to large seismic 3D volumes. In this paper, we present some of the theoretical details of how these two issues are solved, and also demonstrate the results on a 3D seismic dataset.