ng 2008 
497 
EDGE EXTRACTION ALGORITHM BASED ON LINEAR PERCEPTION 
ENHANCEMENT 
Fan ZHANG*, Xianfeng HUANG, Xiaoguang CHENG, Deren LI 
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, 
129 Luoyu Road, Wuhan 430079, China, zhangfanl28@163.com 
KEY WORDS: Image understanding, Edge detection, Computer Vision, Cluster Analysis, Urban planning, Aerial photogrammetry 
ABSTRACT: 
Edges are very important clues for building reconstruction in city-modeling. But some weak edges, such as roofs’ ridges, are often 
similar to the background in intensity which causes low gradient magnitude. The previous edge detection algorithms are hard to 
detect weak edges with tight threshold, while loose threshold will lead to lots of false edges caused by noise and interference. An 
linear features perception based edge detection algorithm is proposed in this paper. The human visual mode of linear perception is 
used as reference, and weak edges are picked according to their linear distribution. The experiments on aerial images of different 
areas show that the proposed algorithm has a significantly enhanced ability to detect weak edges and suppress false edges comparing 
with Edison edge detection algorithm. 
1. INTRODUCTION 
Edges are very important clues for building reconstruction, so 
edge detection is a basic processing in city-modeling. Some 
useful weak edges of city-modeling objects, such as roofs’ ridge, 
are often similar to the background in intensity which usually 
causes low gradient magnitude as shown in the Figure 1. 
Figure 1. Weak edges of roof s ridges 
However, most previous edge detection algorithms are gradient- 
based methods(Canny 1986; C.Gonzalez and E.Woods 2002), 
they detect edges on the basis of the discontinuity in intensity 
between object and background. So, these algorithms have 
problems to detect the weak edges: with tight threshold can’t 
detect weak edge, while with loose threshold will lead to lots of 
false edges caused by noise and interference. To solve this 
contradiction, Canny introduced threshold with hysteresis to 
gain better separation between signal and noise(Canny 1986), 
He and Wang do twice edge detection with tight and loose 
thresholds separately, then process the two results to get 
reasonable edge points(He and Tang 2005; Wang and Wang 
2006). But these methods only select and link the candidate 
edge points locating in the neighbouring area of the reliable 
edge points and ignore the global characteristic presented by the 
edge points. Global characteristic is important information for 
human vision to detective weak edges, human can recognize the 
weak edges that can form into a straight line or a regular curve. 
Based on this assumption, the weak edge with regular 
distribution can be easily detected. 
An edge detection algorithm based on linear perception is 
introduced, the algorithm use the human visual mode of linear 
perception as reference, and pick weak edges according to their 
linear distribution. It has a significant enhanced performance to 
detect weak edges and inhibit false edges. 
2. EDGES CLASSIFYING AND MAIN IDEA OF THE 
ALGORITHM 
The most popular edge detectors are gradient-based technique 
which is on the assumption of the change of intensity value. In 
fact, useful edges for city-modeling in aerial images may be 
missed because of low gradient magnitude, while discontinuity 
in intensity caused by noise and interference may be falsely 
recognized as edges. Therefore, the edges detected by gradient- 
based technique can be classified into three types: strong edges, 
weak edges and false edges, their different characteristics in 
intensity, gradient and shape are shown in the Figure 2. Strong 
edges have evident contrast with background in intensity and 
cause higher gradient magnitude, the edges are continuous and 
steady, strong edges usually are useful edges. Weak edges are 
useful edges as mentioned above, weak edges’ intensity and 
gradient are not as ideal as strong edges’, and they are detected 
as a lot of broken line segments with regular distribution. False 
edges come from noise or the intrusive information that are 
useless for city-modeling such as trees, their characteristics of 
intensity and gradient are similar to weak edges’, they are also 
detected as broken segments, but the segments’ distribution are 
not regular but scattered.