Feature extraction by grayscale morphological operations-a comparison to DOG filters

A comparison is made of the conventional difference-of-Gaussians (DOG) filter and a morphological filter based on opening with a spherical structuring element in realistic images from X-ray inspection. The task is the detection of bright blobs in low-signal/high-noise images containing other objects, such as step edges and corners. The study shows that the morphological filter has a superior selectivity, as it is insensitive to edges, but the resulting images are not free from extra signals. Postprocessing is required to remove undesired responses at sharp, bright corners and similar narrow structures. The morphological filter is more susceptible to noise than its linear competitor, since no averaging is performed. Presmoothing can alleviate this problem, but the choice of the smoothing filter is nontrivial if the advantages of the morphological filter are to be preserved