Alan Forghani
of criteria. These include edge continuity, minimal width (sharpness), accurate location, and completeness in terms of
discriminating all relevant edges. The literature describes many edge detectors. Typical examples are Sobel, Roberts,
Canny, and Deriche edge detectors. There is, of course, no shortage of literature dealing with the edge detection
problem. Numerous authors have surveyed and applied edge operators, among them Fischler et al (1981); Schanzer et al
(1982); Canny (1986); Deriche (1990); Gonzalez and Wintz (1987); Pratt (1987); Argialas and Harlow (1990); Monga
et al (1991); Mason and Wong (1992); Faugeras [13]; Petrou (1993); Domenikiotis (1994); Aghajan and Kailath (1994);
Forghani et al (1997), Forghani (1995, 1997a and 1997b.
2.2 Thresholding
After applying an edge detector, the result produced is noisy. Thresholding is one of the most frequently employed
processes for classifying a pixel as an edge or non-edge (background) pixel from filtered images. Fine tuning of the
threshold is crucial in order to optimise the display of edge features. However, it is a tedious task and difficult to define
a best threshold. Faugeras (1993) says "there is no good answer to this question, and the [threshold value] must be
guided by application and the lighting conditions of the scene". The threshold (T) value depends mostly on image
content, image quality, image geometry, and the amount of noise present in the data (Rosenfeld and Kak, 1982). If the
threshold is set too low, there will be many dominant edges, points, and lines; while if the threshold is set too high,
there will be some dominant features from which much useful information is removed.
Hysteresis thresholding of Canny's edge detector provides one solution to this problem. The optimal threshold for
Canny's operator has been reported to be 70 to 80 percent (Canny, 1986, Guestari, 1995) using hysteresis thresholding.
Experience has shown that the output of edge and line algorithms decreases significantly if the line or edge detectors are
applied to imagery of urban scenes in comparison to rural areas (Forghani et al 1997; Geman and Jednynak, 1996;
Forghani, 1997b).
2.3 Mathematical Morphology
A considerable effort has been made to examine mathematical morphology as a tool for delineating linear structures
from remotely sensed data (Forghani, 1997b; Yokoya,1981; Destival, 1986; Forghani, 1997a). The traditional approach
of mask convolution was first combined with mathematical morphology by Destival (1986). The results of this
approach can be fragmentary and error-prone (Wang and Liu, 1994; Forghani, 1997a), with difficulties arising because
of the number of human interpretations required for rejoining and reconstructing line segments (Destival, 1986). As
mentioned earlier, the linear features in an urban area are very complex, and the use of mathematical morphology can
be affected by the complexity of the scene and because of noise in the data. To overcome these ambiguities and improve
the accuracy of the output, domain knowledge and global information about roads and surrounding areas can be used.
Noise responses may be largely removed by discarding all responses that do not rely on either near or edge borders.
Simultaneously, gaps in the lines can be filled by joining line pixels into continuous lines. Edge linking is a well known
technique. The method is to find local edge pixels using some low-level process, and to join them into contours on the
basis of proximity and orientation.
Possible tools for morphological operations are: bottom hatting, bridging, cleaning, closing, diagonal filling, dilation,
erosion, gating, filling, horizontal breaking, opening, removing, shrinking, skeletonization, spurring, thickening,
thinning, and top hatting (Mathworks, 1995). Further details on the principles of mathematical morphology are
presented by Serra (1988).
3 LINEAR FEATURE DETECTION AND ANALYSIS
31 Program Functions
The ILFDP was implemented in MATLAB. The input for this program can be either remotely sensed imagery or other
types of digital photographic data. In this study, aerial images were processed for detection of linear features. Five
different types of spatial tools were implemented in ILFDP:
I. Standard histogram manipulation techniques.
2. Noise removal filters, namely Wiener and Median. The user is able to apply different filter sizes for median
filtering (eg 3 by 3, 5 by 5, 9 by 9). If the presence of noise in images is ignored, then edge detection is
primarily based on intensity gradients and subsequent thresholding of their magnitude. Non-linear edge
enhancement algorithms are based on gradient operators.
290 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
as a ma 1m AR o XO. .
bl A A RÀ S
es