Image understanding is generally defined as the construction of explicit

meaningful descriptions of the structure and the properties of the 3-dimensional world from 2-dimensional images.A conceptual framework for image understanding is based on Marr’s concept of visual perception as computational process.Marr postulated a hierarchical architecture for vision systems with different intermediate representations and processing levels（low

middle and higher level vision）.According to the description of Marr’s machine vision theory and the characteristics of low resolution remote sensing images

this paper proposes an automatic main-road extraction method

which is based on line segment perceptual grouping and dynamic programming.The method exploits the road model in low resolution remote sensing images at low level and locates the road seeds automatically by line segment perceptual grouping at middle level and then tracks the road seeds to extract the road networks by dynamic programming at high level.Firstly we illustrate the road model in low resolution remote sensing images based on analyzing road characteristics

such as photometric

geometric

topological and contextual characteristics and so on.To increase the precision of road object recognition and to reduce the effects of noise

source images are preprocessed ahead

which include contrast stretching

edge information detection with canny operator and redundant line segments elimination.Edge detection is crucial to line segment perceptual grouping

thus canny operator is applied because of its characteristics of high position precision

single pixel width and low error rate.In the process of line segments elimination

the length and curvature of line segments are considered as the decisive factors to the elimination of redundant line segments.But the directions of the beginning and end line segments are recorded to assist the decision.At middle level

edge line segments are grouped by perceptual grouping technology based on contextual line segments to generate latent road edge line segments.After that

several road seeds are located by computing the latent road edge line segment groups.The relationship of gray value between the regions shaped by the latent road edge line segment and their background is exploited to locate the road seeds thereinto.Then a new road tracking approach using dynamic programming is adopted at high level.The approach introduces the concept of minimized cost route and extends the primitive segment which is formed by direct connection of road seeds to the whole road network in light of minimized cost route.Finally false alarms are eliminated by knowledge inference method

in which inference rules are attained based on the geographic characteristics of road networks in low resolution remote sensing images.We conducted experiments on three datasets of low resolution remote sensing images

which include the Landsat7（B80 band） image with 15m-resolution

the SPOT image of San Diego district with 10m-resolution and the SAR（Synthetic Aperture Radar） image with 12.5m-resolution.Correctness and completeness are introduced to make objective evaluation on the effectiveness of the method.In this way

reference data

which means the road network plotted by observer

should be defined ahead.Experimental results show that our proposed method has high correctness

especially in Landsat7 remote sensing image（98.7%）.Meanwhile the completeness criteria gained from all the source datasets is comparatively high.The lowest value（88.1%） appears in SAR image probably due to the speckle noises.Moreover the followings are proved by the experimental results:（1） the proposed method is effective in lowresolution remote sensing images（high resolution remote sensing images can be sampled to generate its low resolution counterpart）.Especially the images contain some sparse rural roads and intricate city road network;（2） the method is completely automatic and shows better computation efficiency than others

especially compared to semi-automatic road detection methods which need human and computer interaction;（3） the method shows robustness and good performance in remote sensing images such as Landsat7

SPOT and SAR.