Road Network Detection by Mathematical Morphology
2.2 Granulometry
The criterion used in trivial opening can be regarded as a threshold which should be determined from the image analysis with respect to granulometry or pattern spectrum. The idea is to apply opening operation to decompose an image through a series of structure elements with a specific shape. The opened image are compared with the original image to generate measures with respect to different size of structure element with same shape.
Let S(n) = 1,2, ………N be series of structure elements, and X be original image, A(n), n = 1,2, ……….N is a sequence of images being opened as follows
A(0) = AOS0, ……., A(n) = AOSn
Since opening is anti-extensive, A(n) < A(m) for n>=m. let C[A(N)] be the measure of the cardinality of A(n), the following relation are obtained.
C[A] > = C[A(0)]>= C[A(1)]>= C[A(2)]………>=C[A(N)]
Then the size distribution is defined as
C[A(n)]/C[A] is percentage of the filtered objects. SD increases from 0 to 1 and can be regarded as a probability distribution function. C[A(n)]/C[A], SD therefore provide shape-size description of objects in image. Trivial opening can also be used to perform granulometery. However, the notion of structure element size must be reconciled with that of an increasing criterion T. this can be achieved by ordering a criterion T in a set of criteria T(i), I= 1,2,…..N under the following constraints: if a connected component CC does not satisfy Ti, it does not satisfy Ti+1.
3. Experiment on road detection
An experiment was designed and conducted to detect road network from a simulated 1m high resolution image in Toronto, Canada as shown in fig.1. Image is orthorectified and distributed by Earth Watch Inc. the roads in the image make continuous region which form a network, Houses are very dense in the image, some of which roofs almost similar spectral characteristics as roads, and trees are dark in the image along the road standing with the location close to houses.
Fig .1 Test Images and Historgram
The first step is to separate road from background. As the image resolution is high, road network appears to be area with certain width rather than thin lines, this creates an opportunity for classification based method (Benjamin et al. 1990, Gong and Wang 1996).
Image segmentation in this study is achieved through ISODATA (Iterative Self Organizing Data Analysis Technique).
Morphological trivial opening is then applied with the long axis of minimum ellipse that encloses an object to eliminate houses. Granulometry analysis gives size distribution of objects in image as shown in fig.2 with C[A(n)]/C[A] against T. it can be seen with T>=110, the remaining objects are not changed until T>=440, all objects are filtered out. The trivial opening with this criterion T>=110 preserve the road area and filters out almost all the houses and small clusters of noises as well. Morphological closing is applied to fill small holes on the road caused by pixel spectral difference. An initially extracted road network image is then obtained as shown in fig. 3. Some houses are still remained connecting with road network via paths after trivial opening, further processing is needed to remove small paths and these houses.
Fig.2 Size Distribution of Test Image
Fig.3 Initially Extracted Road Network
Fig.4 Final Result of Road network Detection with Thinned Center Line Superimposed on Original Image
