Lawrence Berkeley National Laboratory

Moss et al.(2005) describe, in a recent paper, a filter that they use to detect lines. We noticed that the wavelet on which this filter is based is a difference of uniform filters. This filter is an approximation to the second derivative operator, which is commonly implemented as the Laplace of Gaussian (or Marr-Hildreth) operator (Marr & Hildreth, 1980; Jahne, 2002), Figure 1. We have compared Moss' filter with 1) the Laplace of Gaussian operator, 2) an approximation of the Laplace of Gaussian using uniform filters, and 3) a few common noise reduction filters. The Laplace-like operators detect lines by suppressing image features both larger and smaller than the filter size. The noise reduction filters only suppress image features smaller than the filter size. By estimating the signal to noise ratio (SNR) and mean square difference (MSD) of the filtered results, we found that the filter proposed by Moss et al. does not outperform the Laplace of Gaussian operator. We also found that for images with extreme noise content, line detection filters perform better than the noise reduction filters when trying to enhance line structures. In less extreme cases of noise, the standard noise reduction filters perform significantly better than both the Laplace of Gaussian and Moss' filter.