CN103020605A - Bridge identification method based on decision-making layer fusion - Google Patents

Abstract

The invention discloses a bridge identification method based on decision-making layer fusion, which comprises the following steps: reading in a multispectral image, a panchromatic image and an SAR image, and performing HSV space conversion on the multispectral image; respectively carrying out expansion and corrosion pretreatment on the H, S, V components; h, S, V are respectively subjected to threshold segmentation; removing a small-area interference area by using 8-neighborhood search to obtain a land and water segmentation result; using morphological expansion and morphological corrosion operators to regulate the water body area to obtain a potential bridge area; extracting texture features of potential bridge regions from the full-color image, determining a threshold value, and identifying a bridge; extracting region mean ratio characteristics of potential bridge regions from the SAR image, determining a threshold value, and identifying a bridge; and performing decision-making layer fusion on the bridge identification results in the full-color image and the SAR image. The method comprehensively utilizes the complementary characteristics of the multispectral, panchromatic and SAR images to identify the bridge, and effectively improves the correct identification rate of the bridge target.

Description

Bridge recognition based on decision-making level's fusion

Technical field

The present invention relates to a kind of information processing technology, specifically, relate to a kind of bridge recognition that merges based on decision-making level.

Background technology

In recent years, along with developing on an unprecedented scale of remote sensing technology, the means of obtaining remotely-sensed data are more and more, and remotely-sensed data is also more and more abundanter.Meanwhile, the application of remote sensing is also constantly enlarging and is going deep into.Along with the explosive growth of remote sensing images, the ability of information extraction and efficient have become the bottleneck problem of restriction remote sensing application development.

Target identification technology is the important content of field of information processing, how exactly from image the identification target have huge using value in every field, such as autonomous vehicle in the identification of part in the identification of focus in the identification of sensitive target in the military affairs, the medical image, the commercial production, the communications and transportation to the identification of barrier etc.

Bridge object is as important transport hub, its automatically identification and accurately location, though military or civilian on all be significant; Bridge is important component parts in the spatial geographic information storehouse as man-made features, the identification and extraction of Bridge object directly has influence on the automatization level of atural object mapping, along with the fast development of urban construction, the atural objects such as bridge also are the easiest part that changes and need to upgrade in time in geographical data bank in addition; What is more important, when disaster occured, the situation of understanding accurately, timely bridge helped relief goods are sent to rapidly the disaster area very much.

Main method to bridge recognition can be summarized as following a few class:

1, thinks that the notable feature of bridge is the pair of parallel line, on the edge by the image after cutting apart in land and water, directly seeks parallel lines and judge bridge that method commonly used is the Hough conversion.

2, according to geometric properties, moment characteristics, the transform characteristics of target, carry out the method for template matches, moment characteristics coupling and transform characteristics coupling.The selection of the parameter of the method has a great impact for the result who processes.

3, by setting up the model of river, road etc. and bridge relation, in image, seek the position of coupling with the method for tree search and differentiate bridge.The method is applicable to high-resolution topography.

4, first to Image Segmentation Using, then seek potential Bridge object point according to priori, perhaps isolate bridge by the cluster impact point, perhaps by checking the minimum rectangle at impact point place, target is verified.

In the prior art with the immediate more typical bridge recognition of the present invention:

1, the people such as Xu Shengrong is in the bridge target identification method research [J] based on knowledge. pattern-recognition and artificial intelligence, 1993.5 (2): propose among the 23-128 that the identifying that remote side is clapped bridge is divided into basic, normal, high three and process levels.When low layer was processed, the histogram by former figure and gradient image carried out binaryzation, obtains waters agglomerate primitive and corresponding boundary curve; The middle level is processed with the HOUGH conversion and is carried out the analysis that lines detect and lines concern, the target that obtains supposing; The high-rise processing carried out " the female reason of Model Matching ", adopts the method for degree of confidence to verify.

2, the people such as HuoBioa is at Segmentation and Recognition of Bridges in High Resolution SAR Images[C] .IEEE, propose to adopt wavelet transformation to carry out denoising and keep the edge, river among the 2001:479-482, then adopt the OTSU method to carry out image segmentation, the axis that Refinement operation obtains the river is carried out in the waters that extracts, namely obtain some line segments, find bridge by seeking at a distance of closer line segment end points again.

3, the people such as WuFna is at Recognition of Bridges by Integrating Satellite SAR and Optical Imagery[C] .IEEE, propose among the 2005:3939-3941 to adopt the way of SAR remote sensing images and optical imagery combination to seek bridge, first in the SAR image, find area-of-interest, the then existence of checking bridge in optical imagery.

Prior art adopts the image of single-sensor to carry out Bridge object identification, and information extraction is comprehensive not, can not effectively utilize multi-source image message complementary sense feature, causes object recognition rate lower, is embodied in:

1, only carries out bridge recognition with multispectral image, because multispectral image resolution is lower, extracts bridge and often can occur failing to report or false-alarm;

2, only carry out bridge recognition with the SAR image, often there is phenomenon of rupture in SAR image bridge edge, and there is a large amount of noises in the SAR image simultaneously, and there is very large impact in identification for target, causes easily false-alarm and fails to report;

3, only carry out bridge recognition with full-colour image, the gray-level in river is abundant, and river region is difficult to extract, if do not utilize the position in river to extract bridge, calculated amount is too large.

Summary of the invention

The invention provides a kind of bridge recognition that merges based on decision-making level, adopt multi-sensor fusion technology, fully utilize complementary characteristic multispectral, panchromatic, the SAR image and carry out bridge recognition, solved the low problem of single image source object recognition rate, Effective Raise the correct recognition rata of Bridge object.

Concrete technical scheme is as follows:

A kind of bridge recognition that merges based on decision-making level comprises:

Step 1 is read in multispectral image, full-colour image and SAR image, and multispectral image is transformed into the HSV space by rgb space, obtains H, S, three components of V;

Step 2 expands, corrodes pre-service to H, S, three components of V respectively;

Step 3 is carried out Threshold segmentation to H, S, V respectively;

Step 4 is removed the small size interference region with 8 neighborhood search, obtains the land and water segmentation result;

Step 5, to the waters expand, corrosion treatment, obtain potential bridge zone;

Step 6 is extracted the textural characteristics in potential bridge zone in full-colour image, definite threshold, identification bridge; 1: definite threshold: repeatedly solve gained according to experimental result.2: behind the definite threshold, the zone that the gained textural characteristics is greater than or equal to threshold value is the bridge zone, obtains the result one of bridge recognition, and textural characteristics is non-bridge zone less than the zone of threshold value.

Step 7, the regional average value that extracts potential bridge zone in the SAR image are than feature, and definite threshold is identified bridge; 1: definite threshold: repeatedly solve gained according to experimental result.2: behind the definite threshold, the zone that the gained regional average value is greater than or equal to threshold value than feature is the bridge zone, obtains the result two of bridge recognition, and regional average value is non-bridge zone than feature less than the zone of threshold value.

Step 8 is carried out decision-making level with the result of step 6 and step 7 bridge recognition and is merged, and obtains the bridge recognition result.

Further: in the described step 1, the formula that is transformed into the HSV space by rgb space is:

S＝(MAX-MIN)/MAX ；

$H=\left\{\begin{array}{cc}60*(G-B)/(\mathrm{MAX}-\mathrm{MIN})& R=\mathrm{MAX}\\ 120+60*(B-R)/(\mathrm{MAX}-\mathrm{MIN})& G=\mathrm{MAX}\\ 240+60*(R-G)/(\mathrm{MAX}-\mathrm{MIN})& B=\mathrm{MAX}\end{array}\right.;$

V＝MAX；

MAX＝max(R,G,B)

MIN＝min(R,G,B)；

Wherein MAX and MIN are respectively the minimum and maximum value of R, G in the RGB color model, three components of B, and R, G, B are respectively the gray-scale value of three components, and H, S, V are the gray-scale value of three components of hsv color model of obtaining.

Further: H, S, three components of V in the step 1 are carried out respectively iteration threshold process.

Further: H, S, three components of V in the step 1 are carried out dimension-reduction treatment.

Further: in the described step 3, the process of Threshold segmentation is:

Obtain maximum gradation value and the minimum gradation value of image, be designated as respectively Z _MaxAnd Z _Min, make initial threshold be:

T ₀＝(Z _max+Z _min)/2；

According to threshold value T _kBe prospect and background with image segmentation, obtain respectively both average gray value Z _OAnd Z _BObtain new threshold value:

T _k+1＝(Z _O+Z _B)/2 ；

If T _k=T _K+1, T then _kBe threshold value; Otherwise turn second step, iterative computation is until reach final threshold value.

Further: in the described step 4, the process of removing the small size interference region with 8 neighborhood search is:

If the object pixel of edge image is 0, background pixel is 255, and scan image runs into object pixel P from top to bottom, left to right, then is labeled as new mark value L;

Take P as Seed Points, the object pixel in its 8 neighborhood is labeled as L, until this connected component labeling is complete;

Continue scan image, until in the image all object pixels all mark is complete;

The regional reference area complete to mark arranges threshold value, removes area less than the zone of threshold value.

Further: the detailed process of described step 5 is:

Use form expansion and morphological erosion operator to carry out regular to the water body zone;

The structural element of definition dilation and erosion;

Pixel in the image is first dilation operation n time, and then is erosion operation n time, the river region that obtains being communicated with; Wherein n is natural number;

Binary image is cut apart in river region after being communicated with and land and water subtracted each other, obtain comprising a connected region in potential bridge zone;

Centered by each pixel of connected region, to the length of k pixel of 8 directions extensions, the width of k 〉=bridge; Wherein k is natural number;

If variable is a, if the value of rightabout a pair of pixel is identical, a value gets 1, if opposite, a value gets 0;

A value addition with 4 pairs of pixel gained;

Setting threshold obtains potential bridge zone.

Further: the process of described dimension-reduction treatment is:

Obtain the poor of the maximal value of component gray scale and minimum value, poor dimension divided by needs obtains the numerical value that each gray scale increases progressively.

The present invention adopts multi-sensor fusion technology, utilizing the color characteristic of multispectral image to carry out land and water cuts apart, obtain potential bridge zone, extract respectively textural characteristics and regional average value panchromatic, the potential bridge of SAR image zone and compare feature, in conjunction with three result, obtain the bridge recognition result, Effective Raise the correct recognition rata of Bridge object, brought following technique effect:

1, improved the accuracy rate of bridge recognition;

2, improved bridge recognition speed;

3, reduced the loss of bridge recognition;

4, reduced the false alarm rate of bridge recognition.

Description of drawings

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

Fig. 2 is full-colour image bridge texture feature extraction figure as a result among the present invention;

Fig. 3 is SAR image bridge feature extraction figure as a result among the present invention;

Fig. 4 is on-water bridge 1 source images that comprises a bridge among the present invention;

Fig. 5 is on-water bridge 2 source images that comprise two bridges among the present invention;

Fig. 6 is on-water bridge 3 source images that comprise two bridges among the present invention;

Fig. 7 is on-water bridge 4 source images that comprise five bridges among the present invention;

Fig. 8 is four groups of multispectral fusion method bridge recognitions of image figure as a result among the present invention;

Fig. 9 be among the present invention four groups of images based on the bridge recognition of radiation statistical nature figure as a result;

Figure 10 is collimation method bridge recognition figure as a result in four groups of image rivers among the present invention;

Figure 11 is the on-water bridge 1 single image recognition result figure that comprises a bridge among the present invention;

Figure 12 is the on-water bridge 2 single image recognition result figure that comprise two bridges among the present invention;

Figure 13 is the on-water bridge 3 single image recognition result figure that comprise two bridges among the present invention;

Figure 14 is the on-water bridge 4 single image recognition result figure that comprise five bridges among the present invention;

Figure 15 is the on-water bridge 1 recognition result figure that comprises a bridge among the present invention;

Figure 16 is the on-water bridge 2 recognition result figure that comprise a bridge among the present invention;

Figure 17 is the on-water bridge 3 recognition result figure that comprise a bridge among the present invention;

Figure 18 is the on-water bridge 4 recognition result figure that comprise a bridge among the present invention.

Embodiment

Below with reference to accompanying drawing 1 ~ 3, technical solution of the present invention is described in detail.

Step 1: read in multispectral image, full-colour image and SAR image, multispectral image is transformed into the HSV space by rgb space, obtain H, S, three components of V.

RGB color model space to the space conversion formula of hsv color model is:

S＝(MAX-MIN)/MAX

$H=\left\{\begin{array}{cc}60*(G-B)/(\mathrm{MAX}-\mathrm{MIN})& R=\mathrm{MAX}\\ 120+60*(B-R)/(\mathrm{MAX}-\mathrm{MIN})& G=\mathrm{MAX}\\ 240+60*(R-G)/(\mathrm{MAX}-\mathrm{MIN})& B=\mathrm{MAX}\end{array}\right.$

V＝MAX

MAX＝max(R,G,B)

MIN＝min(R,G,B)

Wherein MAX and MIN are respectively the minimum and maximum value of R, G in the RGB color model, three components of B, and R, G, B are respectively the gray-scale value of three components, and H, S, V are the gray-scale value of three components of hsv color model of obtaining;

H, S, three components of V are carried out respectively iteration threshold to be processed;

Because data volume is excessive, in order to improve running software speed, therefore need to carry out dimension-reduction treatment to three components: obtain the poor of the maximal value of component gray scale and minimum value, poor dimension divided by needs obtains the numerical value that each gray scale increases progressively.Wherein the H component need to multiply by 360 and amplify its numerical value because value is too small.

Step 2: respectively to H, S, three components of V expand, the mathematical morphology pre-service such as corrosion.

Step 3: respectively H, S, V are carried out Threshold segmentation.

Obtain maximum gradation value and the minimum gradation value of image, be designated as respectively Z _MaxAnd Z _Min, make initial threshold be:

T ₀＝(Z _max+Z _min)/2

According to threshold value T _kBe prospect and background with image segmentation, obtain respectively both average gray value Z _OAnd Z _BObtain new threshold value:

T _k+1＝(Z _O+Z _B)/2

If T _k=T _K+1, T then _kBe threshold value; Otherwise turn second step, iterative computation is until reach final threshold value.

Step 4: remove the small size interference region with 8 neighborhood search, obtain the land and water segmentation result.

If the object pixel of edge image is 0, background pixel is 255, and scan image runs into object pixel P from top to bottom, left to right, then is labeled as new mark value L; Take P as Seed Points, the object pixel in its 8 neighborhood is labeled as L, until this connected component labeling is complete.Continue scan image, until in the image all object pixels all mark is complete.The regional reference area complete to mark arranges threshold value, removes area less than the zone of threshold value.The result of three component gained gets respectively the maximal value of its same position.

Because waters and the land color of each multispectral image are different, the characteristic on three component waters and land is also all different, therefore need to judge with the white portion area of acquired results whether acquired results is correct, if area is little, just component is carried out inverse operation, judge again, until area reaches certain threshold value.

Step 5: to the waters expand, corrosion treatment, obtain potential bridge zone.

Use form expansion and morphological erosion operator to carry out regular to the water body zone.Consider the size of noise and Bridge object, the structural element of definition dilation and erosion is:

$B=\left[\begin{array}{ccc}1& 1& 1\\ 1& 1& 1\\ 1& 1& 1\end{array}\right]$

Pixel in the image is done first n dilation operation (n is natural number), and then be erosion operation n time, can remove noise, the river region that obtains being communicated with.Binary image is cut apart in river region after being communicated with and land and water subtracted each other, obtain comprising a connected region in potential bridge zone.

Centered by each pixel of connected region, to the length (k is natural number) of k pixel of 8 directions extensions, k can not be less than the width of bridge.If variable is a, if the value of rightabout a pair of pixel is identical, a value gets 1, if opposite, a value gets 0.With a value addition of 4 pairs of pixel gained, gained and larger, it is larger to illustrate that this central pixel point belongs to the possibility in bridge zone.Setting threshold through judging, obtains potential bridge zone.

Step 6: in full-colour image, extract the textural characteristics in potential bridge zone, definite threshold, identification bridge.

At first produce the gray level co-occurrence matrixes GLCM in potential bridge zone, obtain gray level co-occurrence matrixes by the number of times that calculates two gray-scale values level in the zone adjacent (upper right corner 45 degree directions, vertical 90 degree directions, the upper left corner 135 degree directions), each element among the GLCM represents the number of times of gray scale i adjacent with gray scale j level in the zone (upper right corner 45 degree directions, vertical 90 degree directions, the upper left corner 135 degree directions).

In above-mentioned formula, the molecule on equal sign the right be have certain spatial relationship, gray-scale value is respectively i, the right number of pixel of j, denominator is the right summation number of pixel.P (l ₁, l ₂) expression co-occurrence matrix value.

Because dynamically ask for the interval cost prohibitive of the GLCM in zone, the present invention at first arrives grayvalue transition between the gray area of I.Because image is a gray level image, is transformed into 8 grades.In order to guarantee accuracy rate, return the gray level co-occurrence matrixes of four direction.

In order to carry out Texture classification, at first extract Different categories of samples, add up all kinds of textural characteristics, find out the amount of texture of maximum difference, identify as characteristic quantity.The input full-colour image is carried out texture analysis, and the result of output texture analysis extracts the textural characteristics in potential bridge zone in full-colour image, definite threshold, identification bridge.

Experimental result shows, energy, contrast, these three kinds of textural characteristics of homogeney textural characteristics difference on texture characteristic amount is maximum, therefore calculate the feature of gray level co-occurrence matrixes, extract respectively energy, contrast, the homogeney textural characteristics of full-colour image, and realize the full-colour image bridge recognition.Energy, contrast, homogeney textural characteristics computing formula are as follows:

Energy theorem:

$L_r=\underset{{\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_1}{\mathrm{\&Sigma;}}\underset{l_2}{\mathrm{\&Sigma;}}{P}^2({\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_1,l_2);$

The contrast formula:

$L_d=\underset{{\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_1}{\mathrm{\&Sigma;}}\underset{l_2}{\mathrm{\&Sigma;}}{(l_1-l_2)}^{2}P\left({\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_{1,}{l}_2\right);$

Homogeney textural characteristics formula:

$L_j=\mathrm{\&Sigma;\&Sigma;}\frac{P({\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_1,l_2)}{1+|l_1-{\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="HDA00002669355500031.png"\; state="\{\{state\}\}">l</figure-callout>}}_2|};$

P (l ₁, l ₂) expression co-occurrence matrix value.

Fig. 2 has provided the extraction result of the potential bridge regional texture feature of full-colour image, and wherein 2a is that full-colour image bridge 1 texture extracts the result; 2b is that full-colour image bridge 2 textures extract the result; 2c is that full-colour image bridge 3 textures extract the result.

Step 7: the regional average value that extracts potential bridge zone in the SAR image is than feature, and definite threshold is identified bridge.

Step 8: the result of step 6 and step 7 bridge recognition is carried out decision-making level's fusion, obtain the bridge recognition result.

Utilizing panchromatic, SAR image bridge recognition result is genuine recognition result for genuine or relation (OR rule) provide last bridge recognition result, utilize panchromatic and SAR image recognition result for vacation to provide last bridge recognition result with relation (AND rule) be false recognition result.Adopt decision-making level to merge, required data volume to be processed is little, can improve recognition speed.

Often there is phenomenon of rupture in SAR image bridge, and bridge edge linear feature is not obvious, and is responsive for noise ratio, therefore only detects bridge with shape facility, position feature and causes easily false-alarm and undetected.The angle scatterers that the SAR image forms man-made target (particularly metal target) etc. are very responsive, and man-made target shows as bright especially point or zone in the SAR image.Regional average value is than having reflected the relative brightness average size of SAR image in the zone, and regional average value then is tending towards SAR image non-stationary zone, i.e. urban architecture, highway bridge etc. than large.Regional average value is conducive to keep target shape than feature, can protect preferably the edge in the image, and therefore, the present invention uses regional average value to screen potential bridge zone than feature, helps to solve the problem of SAR image bridge recognition difficulty.The computing formula of regional average value ratio is as follows:

rrm(k)＝u(k)/u

$u\left(k\right)=\frac{1}{n}\underset{(i,j)\&Element;k}{\mathrm{\&Sigma;}}f(i,j);u=\frac{1}{M\&times;N}\underset{i=0}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N}{\mathrm{\&Sigma;}}}f(i,j)$

N is the number of pixels in zone; F (i, j) is the brightness value that the SAR image is located at point (i, j); M, N are respectively height and the width of image; U (k) expression SAR image is in the brightness average of regional k; U represents the brightness average of SAR image.

Fig. 3 has provided the feature extraction result of SAR image bridge, and wherein 3a is SAR image bridge 1 feature extraction result; 3b is SAR image bridge 2 feature extraction results; 3c is SAR image bridge 3 feature extraction results.

By reference to the accompanying drawings 4 ~ 18, the technology of the present invention effect can further specify by following experiment:

Shown in Fig. 4,5,6,7.4a is multispectral image among Fig. 4, and 4b is full-colour image, and 4c is the SAR image; 5a is multispectral image among Fig. 5, and 5b is full-colour image, and 5c is the SAR image; 6a is multispectral image among Fig. 6, and 6b is full-colour image, and 6c is the SAR image; 7a is multispectral image among Fig. 7, and 7b is full-colour image, and 7c is the SAR image; Can find out that from image the resolution of multispectral image bridge is low, be not easy to extract bridge; The bridge edge of full-colour image is obvious, but the waters gray-level is abundant, is difficult for extracting the waters; The edge linear feature of SAR image bridge is not obvious.A bridge is arranged among Fig. 4, two bridges are arranged among Fig. 5,6, five bridges are arranged among Fig. 7; Two bridges of Fig. 5 and Fig. 6 are parallel basically, and five bridges of Fig. 7 are not parallel; Fig. 4,5,6 river are relatively more crooked.

Classical bridge recognition mainly comprises the image segmentation of multispectral fusion, based on collimation method bridge recognition in the bridge recognition of radiation statistical nature and the river.

1, the image segmentation of multispectral fusion.At first gather redness, green and 3 band spectrum images of near infrared, adopt R+IR, 2G-R-IR, R+G+IR, G+IR, the amalgamation modes such as G+R, IR-R, then utilize maximum variance automatically to get the threshold method definite threshold, and according to this Threshold segmentation image.This method of cutting apart by fused image can not take full advantage of the complementary characteristic of bridge in remote sensing images.Multispectral fusion method bridge recognition result as shown in Figure 8,8a is on-water bridge 1; 8b is on-water bridge 2; 8c is on-water bridge 3; 8d is on-water bridge 4.The topological relation feature of only utilizing waters and bridge is cut apart in multispectral fusion, merges river, rear section fringe region and inside, river and is identified as bridge by mistake.

2, based on the bridge recognition of radiation statistical nature.Utilize the Pun entropy of SAR image as the foundation of identification judgement.Because Pun entropy reflection gradation of image distribution situation and abundant information degree generally do not change with target distortion, therefore has good pervasive effect.But the Pun entropy of bridge is little with edge, the river difference that mistake is identified as bridge, false-alarm and undetected therefore occurs easily.Based on the bridge recognition result of radiation statistical nature as shown in Figure 9,9a is on-water bridge 1; 9b is on-water bridge 2; 9c is on-water bridge 3; 9d is on-water bridge 4.Four groups of image bridge recognition false alarm rates are very high.

3, collimation method in the river.Determine area-of-interest with the neighbouring relations of waters and bridge first, recycling is communicated with the axis in river must pass this characteristic removal interference region of bridge, is difficult to accurately determine the axis in river.In the river collimation method bridge recognition result as shown in figure 10,10a is on-water bridge 1; 10b is on-water bridge 2; 10c is on-water bridge 3; 10d is on-water bridge 4.Bridge 1 some bridge does not detect, and some is bridge by flase drop for bridge 2, and bridge 4 has a bridge not identify.Collimation method is better to the more parallel bridge recognition result in two sides, river in the river, but flase drop is appearred in the river bridge detection of bending easily.

Figure 11 has provided single image bridge recognition result to Figure 14.11a is multispectral image among Figure 11, and 11b is full-colour image, and 11c is the SAR image; 12a is multispectral image among Figure 12, and 12b is full-colour image, and 12c is the SAR image; 13a is multispectral image among Figure 13, and 13b is full-colour image, and 13c is the SAR image; 14a is multispectral image among Figure 14, and 14b is full-colour image, and 14c is the SAR image.Can see that multispectral image is identified as bridge, particularly marginal portion, river with a part of river region mistake.Although panchromatic and SAR image bridge recognition result also has false-alarm and undetected, higher than the accuracy of multispectral image bridge recognition.

The background that the present invention utilizes color characteristic to extract multispectral image Bridge place is the river, then be limited to river region and extract potential bridge, according to potential bridge zone in full-colour image textural characteristics and the regional average value in the SAR image identify bridge than feature, take the bridge recognition result of full-colour image as main, the bridge recognition result of SAR image is auxiliary, utilize the OR rule to merge in decision-making level, improving discrimination and reliability, with remedy remote sensing of optical imaging when the sky, weather, cloud cover, have the deficiency of the situations such as camouflage and concealed target.The method the results are shown in Figure 15 to Figure 18.15a is multispectral image among Figure 15, and 15b is full-colour image, and 15c is the SAR image; 16a is multispectral image among Figure 16, and 16b is full-colour image, and 16c is the SAR image; 17a is multispectral image among Figure 17, and 17b is full-colour image, and 17c is the SAR image; 18a is multispectral image among Figure 18, and 18b is full-colour image, and 18c is the SAR image.

Figure 15 can see in conjunction with the bridge recognition result of multi-source Remote Sensing Images complementary characteristic to Figure 18, obviously improves in conjunction with complementary characteristic axle casing recognition correct rate, and false alarm rate is starkly lower than single image identification.Bridge recognition accuracy rate of the present invention and false drop rate statistics are as shown in table 1.

Table 1 single image and multi-source image bridge recognition comparative result

Target identification is because complicated operation is very consuming time, so that some needs the application of fast target identification to be difficult to finish.The present invention extracts first the waters, and the potential bridge extracted region feature that recycling waters segmentation result obtains has reduced computing time, and recognition speed is as shown in table 2.Different images can cause recognition speed different because it varies in size.

Table Bridge 2 beam recognition speed

Can be found out that by above-mentioned experimental result technical solution of the present invention has following beneficial effect:

1, improved the accuracy rate of bridge recognition;

2, improved bridge recognition speed;

3, reduced the loss of bridge recognition;

4, reduced the false alarm rate of bridge recognition.

Effective Raise the correct recognition rata of Bridge object.

Claims (7)

1. a bridge recognition that merges based on decision-making level is characterized in that, comprises the steps:

Step 1 is read in multispectral image, full-colour image and SAR image, and multispectral image is transformed into the HSV space by rgb space, obtains H, S, three components of V;

Step 2 expands, corrodes pre-service to H, S, three components of V respectively;

Step 3 is carried out Threshold segmentation to three components of pretreated H, S, V of completing steps 2 respectively;

Step 4 is removed the small size interference region with 8 neighborhood search for three components of the Threshold segmentation of completing steps 3, obtains the land and water segmentation result;

Step 5, to the waters expand, corrosion treatment, obtain potential bridge zone;

Step 6 is extracted the textural characteristics in potential bridge zone in full-colour image, definite threshold, and the identification bridge obtains the result one of bridge recognition;

Step 7, the regional average value that extracts potential bridge zone in the SAR image are than feature, and definite threshold is identified the result two that bridge obtains bridge recognition;

Step 8, the result one of the bridge recognition in the step 6 and the result two of the bridge recognition in the step 7 are carried out decision-making level's fusion, obtain the bridge recognition result, namely utilizing panchromatic, SAR image bridge recognition result is genuine recognition result for genuine or relation provide last bridge recognition result, utilizes panchromatic and SAR image recognition result providing last bridge recognition result with relation and be the recognition result of vacation for vacation.

2. the bridge recognition that merges based on decision-making level as claimed in claim 1 is characterized in that in the described step 1, the formula that is transformed into the HSV space by rgb space is:

S＝(MAX-MIN)/MAX ；

$H=\left\{\begin{array}{cc}60*(G-B)/(\mathrm{MAX}-\mathrm{MIN})& R=\mathrm{MAX}\\ 120+60*(B-R)/(\mathrm{MAX}-\mathrm{MIN})& G=\mathrm{MAX}\\ 240+60*(R-G)/(\mathrm{MAX}-\mathrm{MIN})& B=\mathrm{MAX}\end{array}\right.;$

V＝MAX ；

MAX＝max(R,G,B)

MIN＝min(R,G,B) ；

Wherein MAX and MIN are respectively the minimum and maximum value of R, G in the RGB color model, three components of B, and R, G, B are respectively the gray-scale value of three components, and H, S, V are the gray-scale value of three components of hsv color model of obtaining.

3. the bridge recognition that merges based on decision-making level as claimed in claim 1 is characterized in that described step 1 is carried out dimension-reduction treatment to H, S, three components of V, then respectively H, S, three components of V is expanded, corrodes pre-service.

4. the bridge recognition that merges based on decision-making level as claimed in claim 1 is characterized in that in the described step 3, the process of Threshold segmentation is:

Obtain maximum gradation value and the minimum gradation value of image, be designated as respectively Z _MaxAnd Z _Min, make initial threshold be:

T ₀＝(Z _max+Z _min)/2 ；

According to threshold value T _kBe prospect and background with image segmentation, obtain respectively both average gray value Z _OAnd Z _BObtain new threshold value:

T _k+1＝(Z _O+Z _B)/2 ；

If T _k=T _K+1, T then _kBe threshold value; Otherwise turn second step, iterative computation is until reach final threshold value.

5. the bridge recognition that merges based on decision-making level as claimed in claim 1 is characterized in that in the described step 4, the process of removing the small size interference region with 8 neighborhood search is:

If the object pixel of edge image is 0, background pixel is 255, and scan image runs into object pixel P from top to bottom, left to right, then is labeled as new mark value L;

Take P as Seed Points, the object pixel in its 8 neighborhood is labeled as L, until this connected component labeling is complete;

Continue scan image, until in the image all object pixels all mark is complete;

The regional reference area complete to mark arranges threshold value, removes area less than the zone of threshold value.

6. the bridge recognition that merges based on decision-making level as claimed in claim 1 is characterized in that the detailed process of described step 5 is:

Use form expansion and morphological erosion operator to carry out regular to the water body zone;

The structural element of definition dilation and erosion;

Pixel in the image is first dilation operation n time, and then is erosion operation n time, the river region that obtains being communicated with; Wherein n is natural number;

Binary image is cut apart in river region after being communicated with and land and water subtracted each other, obtain comprising a connected region in potential bridge zone;

Centered by each pixel of connected region, to the length of k pixel of 8 directions extensions, the width of k 〉=bridge; Wherein k is natural number;

If variable is a, if the value of rightabout a pair of pixel is identical, a value gets 1, if opposite, a value gets 0;

A value addition with 4 pairs of pixel gained;

Setting threshold obtains potential bridge zone.

7. the bridge recognition that merges based on decision-making level as claimed in claim 3 is characterized in that the process of the dimension-reduction treatment in the described step 1 is:

Obtain the poor of the maximal value of component gray scale and minimum value, poor dimension divided by needs obtains the numerical value that each gray scale increases progressively.

Images