CN103106655A - Construction site unsupervised extraction method based on remote-sensing image - Google Patents

Abstract

The invention provides a construction site unsupervised extraction method based on a remote-sensing image according to the characteristics of a construction site. According to the construction site unsupervised extraction method, seed points are generated again by using the texture features of the construction site in the remote-sensing image, an under-construction area is formed through increase of the seed points, and unsupervised extraction of the construction site is realized. The construction site unsupervised extraction method mainly comprises four steps: step (1), green plants which have texture features similar to the texture features of the construction site in the image are removed; step (2), the image is converted into a YCbCr space; step (3), similarity between each pixel and pixels in a 3*3 neighborhood of the pixel is computed to obtain backup seed points, the Euler distance between each backup seed point and a neighbor pixel of the backup seed point is computed to obtain the seed points according to judgment; and step (4), the seed points in a dense area are increased to form the under-construction area according to certain rules. The construction site unsupervised extraction method based on the remote-sensing image fills a blank in an extraction algorithm in allusion to under-construction buildings at present, provides powerful evidence for investment and construction progress monitoring, casualty loss detection and government decision making, and plays an important role for identification of targets with specific texture features.

Description

The non-supervisory extracting method in a kind of building site based on remote sensing image

Technical field:

The present invention is a kind of practicality, efficiently building site recognition methods, is applicable to high resolution remote sensing satellite data and the unmanned plane data of taking photo by plane, and is specifically related to pattern-recognition and digital image processing techniques.Can extract in the building ground, the aspects such as urban development monitoring, the monitoring of engineering construction progress, disaster detection, assessment have a wide range of applications.

Background technology:

It is one of basic problem of carrying out monitoring of land use that buildings extracts, for urban development monitoring analysis, investment construction process monitoring and government decision provide reliable foundation.

The algorithm that utilizes at present remote sensing image that buildings is identified is all to extract for the buildings that is using, and has ignored the buildings of building.These methods are mainly based on information such as spectral information and texture, shape, edges, and shape, height and the color characteristic basic according to buildings are by setting empirical value and the pattern of setting up is extracted the buildings that has built up.Can only be for the buildings with given shape or spectral information and height, and higher to their spectral information and feature and the distribution limitation of texture information, to texture and the more complicated identification of building buildings of spectral information blank out especially.

The present invention has utilized region growing thought, proposes a kind of building site extracting method, can be to identifying at the buildings of building.The special spectral characteristic in application and construction building site is set the formation condition of Seed Points, will be in the violent regional Seed Points of spectral signature variation and extract, and by formulating new growth rule, the point in its neighborhood scope is classified as in the building site.

Summary of the invention:

At present, buildings identification is mainly for the buildings of having built well, does not comprise the building ground of building.And the method spectrum distinctive according to most of buildings and the information designs such as textural characteristics and geometric configuration that generally use, has very large randomness, be only applicable to some specific environment, only specific buildings had good extraction effect, universality is not high.

The present invention is a kind of non-supervisory building ground extraction method, for building ground spectrum and the trifling feature of texture information, selected spectrum and texture are concentrated the Seed Points of region of variation, give region growing new growth rule, utilize the method for region growing that the building ground is extracted accurately and efficiently.

Concrete method step is as follows:

The first step: remove unmanned plane aviation image Green vegetation

Green vegetation occupies very large ratio in remote sensing images, and they and the buildings built have equally and change violent texture information, can extract the building ground and cause interference.Because their spectral information is more stable, consistent, can utilize analysis of spectrum threshold at first vegetation to be removed.

Second step: the color of image space is transformed into YCbCr from RGB

Three passage R of RGB color space have very high correlativity between G and B, be suitable for showing image.But the distance between pixel in rgb space can not be expressed the mankind to the difference of different atural object perception on consistent yardstick, therefore be not suitable for utilizing rgb space that spectrum and the texture of atural object are analyzed.(Y refers to luminance component to YCbCr, Cb refers to the chroma blue component, and Cr refers to the red color component) be applicable to image segmentation and extract, and the mankind can be with Euler apart from representing (referring to document: Shih to the perception of color distortion, F.Y.and S.Cheng, Automatic seeded region growing for color imagesegmentation.Image and Vision Computing, 2005.23 (10): p.877-886).Therefore we utilize the YCbCr color space that aviation image atural object is analyzed.

The 3rd step: automatically generate Seed Points

1, calculate the similarity of each pixel and its surrounding pixel

Setting window size is 3 * 3, and image variance on a certain wave band in window ranges is:

${\mathrm{\σ}}_x=\sqrt{\frac{1}{9}\underset{i=1}{\overset{9}{\mathrm{\Σ}}}{(x_i-\stackrel{\‾}{x})}^2}---\left(1\right)$

Wherein, x is Y, three passages of Cb and Cr.

$\stackrel{\‾}{x}=\frac{1}{9}\underset{i=1}{\overset{9}{\mathrm{\Σ}}}{x}_i---\left(2\right)$

Population variance can be expressed as:

σ＝σ _Y+σ _Cb+σ _Cr(3)

Be standardized as:

${\mathrm{\σ}}_N=\frac{\mathrm{\σ}}{{\mathrm{\σ}}_{\max }}---\left(4\right)$

Wherein, σ _maxIt is variance maximum in 3 * 3 windows take all pixels of view picture image as core

So, the spectral similarity of 3 * 3 windows take a certain pixel as core may be defined as:

H＝1-σ _N (5)

If the similarity that calculates is greater than some threshold values, this pixel is classified as seed reserve point.This threshold value is to be determined by ' Otsu ' method.' Otsu ' method also is poor method between maximum kind, and its threshold value of determining can be divided into image prospect and background two classes, and makes inter-class variance maximum, and the class internal variance is minimum.Core concept is:

t＝Max[w ₀(t)×(u ₀(t)-u) ²+w ₁(t)×(u ₁(t)-u) ²] (6)

Wherein t is the threshold value of cutting apart, w ₀(t) be the backdrop pels number shared ratio of threshold value when being t, w ₁(t) be the prospect pixel number shared ratio of threshold value when being t, u ₀(t) be the background mean value of threshold value when being t, correspondingly, u ₁(t) be the prospect average.U is the average of entire image, that is:

u＝w ₀×u ₀+w ₁×u ₁ (7)

In the computing formula of threshold value, w ₀(t) * (u ₀(t)-u) ²+ w ₁(t) * (u ₁(t)-u) ²Be the calculating variance between prospect and background two classes.Namely make the threshold value of inter-class variance maximum be ' image segmentation threshold that determines of Otsu ' method.

2, calculate the reserve Seed Points and be adjacent Euler's distance corresponding between pixel in 3 * 3 windows

Wherein, Euler's distance can be defined as:

$d_i=\frac{\sqrt{{(Y-Y_i)}^2+{(C_b-C_{b_i})}^2+{(C_r-C_{r_i})}^2}}{\sqrt{Y^2+C_b^2+C_r^2}}---\left(8\right)$

i＝1，2，…，8

And between pixel and its neighbours' pixel, maximum Euclidean distance is:

$d_{\max }=\underset{i=1}{\overset{8}{\max }}\left(d_i\right)---\left(9\right)$

In the reserve Seed Points, if certain any maximum Euclidean distance is set as Seed Points greater than 0.05.

The 4th step: region growing

Traditional region growing method is to be a plurality of zones with image segmentation, and the zone is merged into the true atural object with certain textural characteristics and level and smooth spectral information.But for the buildings of building, owing to having very strong spectrum crumbliness and texture inconsistency, be difficult to utilize traditional method that the zone is merged.Therefore we according to the formation characteristics of Seed Points, have invented a kind of new region growing method.

1, judge whether Seed Points is in close quarters

Because the Seed Points that generates is in spectrum and texture information all changes larger place, and the spectrum of the buildings in process of construction and texture information change and compare comparatively dense.Therefore, in the zone, building ground, intensive Seed Points is arranged.In order to judge whether Seed Points belongs to high-density region, we add up the number of Seed Points in 5 * 5 window ranges centered by each Seed Points.If surpass 16, think that it is in the high density area.

The size of the window ranges of statistics Seed Points number has determined the susceptibility of Seed Points to closeness.If window ranges is very little, be difficult to reach the standard of closeness, can cause like this Seed Points that much is positioned at the zone, building ground to be lost, make the building ground by the vacant lot that is identified as of mistake.If window ranges is very large, can be easy to reach the closeness standard, can increase the building ground area that detects, cause erroneous judgement, the buildings that vacant lot is taken as in process of construction extracts.Through experiment showed, that 5 * 5 windows are comparatively suitable, also conform with ground truth with the 16 extraction results as the standard of weighing the density height.

2. determine seed boundary of a set of points rectangle in close quarters

Just the buildings in process of construction has intensive, discrete Seed Points set, but is not connected region.For with discrete point set compartmentalization, we centered by each seed, in 5 * 5 regional extents, find the boundary rectangle of the set that all Seed Points consist of in highdensity seed point set, and the non-Seed Points of rectangle inside all is labeled as Seed Points.Like this, discrete seed point set increases and becomes continuous, smooth zone, is the extraction result of building ground.

Description of drawings:

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is the key drawing of the new Rule of Region-growth of formulating;

Fig. 3 is concrete emulation case;

Embodiment:

Fig. 2 is explaining of the region growing principle that proposes of the present invention.Its Green point is non-Seed Points, yellow Seed Points for selecting, red core seeds point for processing.In this example, red point has 17 Seed Points in 5 * 5 neighborhoods, the threshold value 16 of the judgement dense degree of setting greater than the present invention, so after region growing, in 5 * 5 scopes have a few and all be marked as Seed Points, the building site of namely monitoring.

The process of Fig. 3 for adopting this method that the building ground of high resolving power unmanned plane aviation image is extracted

Test intercepts the unmanned plane aviation image of Yunnan Airport being taken on May 30th, 2009, and size is: 886x554

Figure (a) is Yunnan Airport part capital construction areal map.In figure, atural object to build the zone as main, wherein has part more scattered (seeing atural object in blue circle), and major part is more concentrated.Also have the subregion to belong to the non-building area, atural object in the show chromosphere.

The seed point diagram of figure (b) for extracting.Can see, the closeness of Seed Points is along with the building ground changes than intensity.In area, non-building ground, Seed Points distributes very sparse, and in the building ground intensive central area, Seed Points distributes also very intensive.

Figure (c) is the extraction figure as a result of construction area.That distribute to concentrate is just well extracted in the building site, and near distribution more sparse small-sized building site also is extracted.Non-construction area is also well neglected.Extract result smoother, complete, also with practically phase is near for geometric configuration, has certain application potential.

Claims (4)

1. based on the building ground extracting method of high-resolution remote sensing image, it is characterized in that utilizing the formation characteristics of Seed Points, building ground to spectral information and texture information fragmentation positions, the recycling region growing method will be discrete the seed point set generate to be communicated with and distinguish, and then be created on and build the zone, building site.Specific embodiments is as follows:

(1) utilize analysis of spectrum threshold that image Green vegetation is rejected

Although green vegetation has more continuous and level and smooth spectral information, its texture information is more trifling, easily the extraction of building ground is caused interference.Therefore at first utilize spectrum threshold that green vegetation is removed.

(2) with image from the RGB color space conversion to the YCbCr space

Rgb space is adopted in the demonstration of remote sensing images more, but because the correlativity between R, G and B is higher, can't expresses the mankind to the difference of different atural object perception with spectrum intervals, thereby can not be used for the processing such as image segmentation, target identification.And YCbCr (Y refers to luminance component, and Cb refers to the chroma blue component, and Cr refers to the red color component) space is more responsive to spectrum and texture information, and is multiplex in image segmentation, extraction feature of interest.Therefore need to be with color space conversion.

(3) automatic decision and generation Seed Points

At first calculate the spectral similarity that each pixel is adjacent pixel, and utilize the otsu method automatically to generate threshold value, as the evaluation criterion of spectral similarity.If greater than this threshold value, as the alternative point of seed, then calculate texture similarity between itself and adjacent element, generate the seed point set.

(4) region growing generates the building ground and extracts result

For the characteristics of texture information and the spectral information of building ground, and the distribution characteristics of Seed Points, take each Seed Points as core, judge the Seed Points number in its 5 * 5 scope, determine whether it belongs to the compact district.If belong to the compact district, with in its 5 * 5 scope have a few all to be labeled as and building the building site identified region.

2. according to right 1 requirement, it is characterized in that, spectrum and the texture information of building ground are more trifling, match with the Seed Points formation condition, can will both combine.

3. according to right 1,2 requirements, it is characterized in that, Seed Points is intensive at building ground punishment cloth, the Seed Points that is positioned at the building ground can be extracted by setting intensive conditions.Therefore, each Seed Points is placed in the core place of 5 * 5 windows, judges that in its window be the number of Seed Points in abutting connection with pixel, if greater than 16, set this Seed Points and be in close quarters, be i.e. the building ground.

4. according to right 3 requirements, it is characterized in that, a large amount of discrete seed point sets distribute on the zone, building ground.Seed Points in each highly dense zone is as core, and the pixel of its all non-Seed Points in 5 * 5 neighborhood rectangles is labeled as Seed Points, i.e. region growing operation obtains smooth continuous building ground recognition result.

Images