Multi-scale Data Term
To benefit from the many advantages of a multi-scale analysis (less computational load, robustness, multi-resolution description of the image), a wavelet transform is applied to get a series of images at different scales. In each scale, based on the observed histograms, we model the image in the road and the background using two-component Gaussian mixture distributions for pixel intensity, and Gamma distributions for the variance in a window. The probabilistic models define the data energy ED.
This data energy plus the prior energy in [1] is then used to extract the road network from a QuickBird panchromatic image (Fig. 1(a)). At 1/8 resolution, the complete main road network is successfully retrieved (Fig. 1(b)), but the road region is actually not very accurate. To improve on the result, extraction needs to be performed at full resolution.
However, we observe experimentally that if we try to use the same model at finer resolutions, the details of the scene in the image make road extraction more difficult (Fig. 1(c)-(e)). In an attempt to overcome this problem at finer resolutions, we define a multi-scale data energy as a sum of the data energies at several different scales. The use of several scales allows the combination of coarse scale data, in which the details to disrupt recognition can be eliminated, with fine scale data to increase precision. The result is not perfect, but is very promising considering the complexity of the image (Fig. 1(f)).
Fig. 1 Experiments. (a) a QuickBird image
(size: 2560 × 2560); (b)-(e): result obtained using the
single-scale data energy respectively at 1/8, 1/4, 1/2 and
full resolution; (f) result obtained using the multi-scale
data energy at full resolution.
Last update: Dec. 10, 2008