The most important information for image identification is edges of objects. But when objects are seated among nature background

those edges of nature background will be a great burden to the identification process. So unnecessary edges must be removed before identification. In tradition

a method called“huerestic threshold”is used to reach that target. But it performs bad in preserving details of objects

because local information is not taken into account. According to fractal geometry

it is found that the fractal density of edges of men-made objects is close to 1

while that of nature background is close to 0 or 2. Thus a novel method is proposed to fulfill the demand. This method integrates the consideration of both gradient strength of edge points and fractal density of curves

and uses cluster analysis for filter. In this paper

images of planes with cloud and forest as background are used to testify the effectiveness of the new method. From the experiment

it can be seen that most background edges are filtered

yet details of planes are preserved. so the method is beyond the function of the traditional method. At the end of this paper

the limit of this method is discussed.