The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
(a) 
(b) 
(c) 
y 
(d) 
(e) 
(f) 
S 
¡- 'ï,'-' \ 
|\c 
(g) 
(h) 
(i) 
Figure 2. Sketch map of different edge types. (a)(b)(c) 
Intensity, gradient and detected shape of strong 
edge. (d)(e)(f) Intensity, gradient and detected 
shape of weak edges. (g)(h)(i) Intensity, gradient 
and detected shape of false edges. 
The aim of the algorithm is to keep useful edges (include strong 
and weak edges) and get rid of useless edges (include false 
edges). The strong edges have distinct characteristic, they are 
hard to confuse with the other two. So, the crux is how to 
distinguish weak edges from false edges. There are some 
similar characteristic between weak edges and false edges: they 
are both detected as broken segments and the difference in 
intensity is not so obvious. But the distribution of weak edges is 
so regular that it is sensitive to our vision. However, the 
previous edge detection algorithms ignored the regular 
distributing, they mainly focused on local intensity change and 
the neighbouring relation between certified and candidate edge 
points. It led to the losing of weak edges with slim difference in 
intensity. The main idea of the proposed algorithm is that 
simulate the sensitive perception of human vision to regular 
distribution, and investigate the distribution characteristic of 
edges in a global view, then pick, link and expand the weak 
edges. For example, an edge is short and with low gradient 
magnitude, if it can form into a line with some other edges, the 
edge will be marked as candidate edge to gain further 
processing, otherwise it will be deleted. 
3. ALGORITHM DESCRIPTION 
The flow of the nroposed algorithm is shown as Figure 2. In 
preliminary edge detection procedure, credible edges can be 
gained by previous edge detection algorithm with tight 
thresholds and candidate edges can be gained by the same 
algorithm with loose thresholds. All the detected edges are 
fitted into line segments, and the segments’ parameters in polar 
coordinates system are calculated in the line fitting procedure. 
Then, the candidate edges are grouped, picked and linked under 
the idea of linear perception, in this procedure, most false edges 
are deleted. And then, expand the edges from their endpoints to 
gain more complete edges. At last, the result of edge detection 
with more weak edges and less false edges are output. 
Input Image 
Figure 3. The flow chart of the proposed algorithm 
3.1 Preliminary Edge Detection 
Edison algorithm(Meer and Georgescu 2001) is improved based 
on Canny edge detection algorithm, and it applies embedded 
confidence to all procedures of edge detection and satisfies the 
three performance criteria for detecting edges: good detection, 
good localization and only response to a single edge(Canny 
1986). Embedded confidence is the information which is 
independent from gradient estimation and used for evaluating 
the comparability between data model and ideal edge model, 
thus, can detect edges of all kinds of figures by reducing the 
uncertainty in edge detection(Jiang 2004). So, we adopt Edison 
algorithm to perform preliminary edge detection. 
In this procedure, credible edges can be gained by using tight 
thresholds and candidate edges can be gained by using loose 
ones. Credible edges are strong edges, most of them are useful 
edges; while candidate edges consist of weak and false edges, 
both useful and useless edges are included. Therefore, candidate 
edges are main part in the next procedures. 
3.2 Line Fitting 
Line fitting is aimed at finding out the linear feature shown by 
the edges and calculating the parameters of 0 and p in polar 
coordinates system. It could be a good preparation for the next 
procedures. Hough transform has been widely used as a typical 
algorithm. But Hough transform is a global calculating process, 
it is hard to detect short segments, because the voting mode and 
accumulation has problems that the peak value caused by the 
isolated points forming into line by accident is higher than the 
peak value caused by short segments, and the segments at a 
distance but with similar parameters effect each other(Furukawa 
and Shinagawa 2003). The proposed algorithm solves these 
problems with localized Hough transform. We clustered edge 
points into edge strips according to their neighbouring 
relationship. Then, calculate the parameters of each edge strip 
respectively by Hough transform, the process could avoid the 
negative effect brought by global calculating, and gain a clear 
peak value even when processing on short segments. To reduce 
the computational complexity and accelerate the computational 
speed of Hough transform, the numerical interval of parameter 
space can be extend properly, after voting, get the points voting 
for the peak value to fit an accurate line by least square. 
3.3 Grouping and Linking Edges 
Traditional methods of perception grouping only considered 
498