Page 264 - Digital Analysis of Remotely Sensed Imagery
P. 264

226    Cha pte r  S i x

               duplicating the border rows and columns. However, those linear fea-
               tures very close to the image border cannot be detected well using
               this method.
                   The above discussion applies to ideal situations where there is
               no noise in the image, which is rarely true in reality. In order to reduce
               the random variation of pixels within the same feature and hence
               improve the reliability of edge detection, it may be necessary to smooth
               the image using the methods described in Sec. 6.3.3 before it is used in
               the detection. The detection quality can also be improved by imposing
               a threshold to test the validity of the detected edges in a postdetection
               session. For instance, only those differences exceeding a certain thresh-
               old are regarded as representing genuine edges. All other differences
               are treated as noise and removed. Another postdetection process-
               ing technique is to spatially filter the detected results. All isolated non-
               zero pixels that do not appear to be aligned with any linear segments
               in a meaningful direction are eliminated from the output image.

               6.4.2 Edge-Detection Templates
               Several templates have been devised for edge detection (Fig. 6.16).
               All of them have one characteristic in common: the sum of all ele-
               ments in a kernel is zero. These zero-sum kernels smooth out areas of
               low spatial frequency (e.g., absence of any edge), and cause a low
               output in areas of low spatial frequency. In areas of high spatial fre-
               quency (e.g., the interface of homogeneous patches of pixels), a sharp
               contrast results. It is possible for the edge-enhanced images to con-
               tain only edges and zeros. In areas of high spatial frequency, the dis-
               parity among pixel values is magnified as large values become larger
               while low values become even lower owing to the unequal element
               values along the first/last row/column (Leica, 2006). Edges are
               enhanced in the filtered image as it frequently is made up of only
               edges and zeros.
                   Two kinds of special edge-detection operants deserve more discus-
               sion here, Sobel filters (Fig. 6.16a, b, c, top) and Prewitt (1970) filters
               (Fig. 6.16a, b, c, bottom). The nine elements can be arranged horizontally,

                        –1  0  1        –1  –2  –1      –2  –1  0
                        –2  0  2         0  0  0        –1  0  1
                        –1  0  1         1  2  1         0  1  2

                        –1  0  1        –1  –1  –1       0  1  1
                        –1  0  1         0  0  0        –1  0  1
                        –1  0  1         1  1  1        –1  –1  0

                        (a) Vertical   (b) Horizontal   (c) Diagonal
               FIGURE 6.16  Examples of edge detection operants. (a) Operants designed to
               detect vertically oriented edges; (b) operants designed to detect horizontally
               oriented edges; and (c) operants designed to detect diagonally oriented edges.
   259   260   261   262   263   264   265   266   267   268   269