CN108109156B - SAR Image Road Detection Method Based on Ratio Feature - Google Patents

Abstract

The invention discloses a ratio feature-based SAR image road detection method, which mainly solves the problems of inaccurate road edge positioning and high false alarm rate in the prior art. Its implementation includes: 1) perform speckle reduction on the SAR image and extract 9 texture features, 2) select the 3 texture features that contribute the most to the classification from the 9 texture features according to the Bhattacharyian distance; 3) calculate the speckle reduction point by point Contrasting ratio feature R1 and similar ratio feature R2 in the image; 4) Construct road dictionary D ₁ and background dictionary D ₂ with samples of 2) and 3) results; 5) Solve for each pixel point and road dictionary D ₁ respectively The average difference E ₁ and the average difference E ₂ with the background dictionary D ₂ are used to classify the pixel points to obtain the preliminary detection result; 6) optimize the preliminary detection result to obtain the final detection result. The invention can relatively complete and clearly detect roads in SAR images, and is suitable for detecting roads in different directions and widths in SAR images.

Description

SAR Image Road Detection Method Based on Ratio Feature

technical field

The invention belongs to the technical field of digital image processing, in particular to a road detection method for SAR images, which can be used for map updating, transportation logistics and urban planning.

Background technique

In recent years, with the continuous development of key technologies of synthetic aperture radar, SAR imaging resolution has been continuously improved, signal processing capability has been continuously enhanced, data transmission rate has been continuously increased, equipment volume has been continuously reduced, and quality has been continuously reduced. SAR images can be widely used in intelligence. Collection, battlefield surveillance, attack guidance, strike effect evaluation, etc. SAR image technology has made considerable progress and development since its birth to the present for more than 50 years. In SAR image analysis, line features are very important, because some objects in the image have linear structures, such as roads, bridges, rivers, coastlines, etc. It has been people's wish for many years to use computers to automatically extract linear feature information such as roads from SAR images. Line feature detection is used in multi-sensor image registration, cartography applications, image segmentation and object recognition, etc. For objects with a certain width range, accurate contour lines are helpful for the segmentation of different objects and the recognition of targets.

As an important man-made feature, road is the main part of the modern traffic system, and has important geographical, political, economic and other significance. Because the synthetic aperture radar (SAR) system has the advantages of all-day and all-weather, road extraction from SAR images has been paid more and more attention and applied. In a large number of acquired high-resolution SAR images, road extraction is often not only the intermediate process of SAR interpretation, but also a result of interpretation. The extraction of roads from SAR images is currently a hot spot in the academic circles. At the same time, SAR image road extraction has important application value in the fields of military intelligence interpretation and civilian urban planning.

The general SAR image road detection algorithm is to first determine the road seed points according to the primitive detection, and then use the line detection algorithm to connect the road seed points into road segments, such as the methods proposed by Tupin et al., Kartartzis et al. and Jeon et al. These conventional primitive detection methods all have problems such as inaccurate positioning of the road edge and high false alarm rate.

Contents of the invention

The object of the present invention is to address the deficiencies of the above-mentioned prior art, and propose a SAR image road detection method based on ratio features, so as to reduce the false alarm rate and improve the accuracy of edge positioning.

In order to achieve the above object, the implementation scheme of the present invention includes as follows:

(1) Read in the SAR image to be detected, and perform speckle reduction preprocessing on the SAR image;

(2) Construct a gray level co-occurrence matrix point by point for the image after spot reduction, and use the gray level co-occurrence matrix to calculate energy, entropy, contrast, mean, variance, correlation, dissimilarity, inverse gap and uniformity 9 texture features ;

(3) Optimize and screen the nine texture features extracted above according to the Bhattacharyya distance index, and only select the three features that can effectively contribute to the road detection and classification in SAR images, namely: mean, variance and contrast;

(4) Extract ratio features:

(4a) Extract two kinds of ratio features point by point from the image after speckle reduction, that is, the contrast ratio feature R1 between the road and the areas on both sides and the similarity ratio feature R2 between the areas on both sides of the road;

(4b) normalize the 3 texture features extracted in step (3) and the two ratio features extracted in step (4a), to obtain a 5-dimensional feature vector of each pixel;

(5) Randomly select some pixel points with labels as samples, including road points and non-road points, construct road dictionary D ₁ and background dictionary D ₂ with this sample;

(6) Preliminary detection:

(6a) Respectively solve the difference E 1 of each pixel point and the road dictionary D ₁ and the difference E ₂ of each pixel point and the background dictionary D _{2 ;}

(6b) Classify the pixel points according to the two differences E ₁ and E ₂ : if E ₁ -E ₂ <0, it means that the pixel point of the test sample has a stronger correlation with the atom in the road area, and it is initially judged as a pixel The point belongs to the road area; if E ₁ -E ₂ ≥ 0, it means that the pixel point of the test sample has a stronger correlation with the atoms in the background area, and the pixel point is initially judged to belong to the background area;

(7) For the preliminary detection results obtained in step (6), according to the characteristics of the slender shape of the road, the optimization algorithm of the area-to-perimeter ratio is used to exclude false alarm areas, and the final result of road detection is obtained.

Compared with the prior art, the present invention has the following advantages:

1. Since the present invention defines a new calculation method of ratio features, the algorithm fully considers the large difference between the road and the areas on both sides, and the strong similarity between the two sides of the road, and overcomes the edge positioning of the conventional primitive detection operator Inaccurate shortcomings, the extracted two ratio features can accurately locate the edge of the road;

2. In view of the redundancy between texture features, the present invention proposes a rule for screening features with Bhattacharyya distance. Because Bhattacharyya distance fully considers the correlation between the mean value, variance and standard map of features, it not only reduces the complexity of features , also improves the operating efficiency of the algorithm;

3. The present invention aims at the characteristics of high real-time requirements for road detection, and proposes the design of using difference classification by constructing a dictionary, which fully considers the characteristics of strong similarity between similar classes and large differences between different classes. Compared with some A classifier with high complexity and low practicability, the difference classifier of the present invention simplifies calculation and can meet the requirements of real-time performance;

4. The optimization algorithm using the area-to-perimeter ratio proposed by the present invention can effectively eliminate false alarm areas. Due to the existence of false alarms in the preliminary detection results, known roads have the characteristics of a slender shape, that is, a closed curve with a constant length. According to the integration principle, the closer the shape is to the slender shape and the smaller the area, the more consistent characteristics of the road.

Description of drawings

Fig. 1 is the general flowchart of road detection of the present invention;

Fig. 2 is the subflow chart of optimization algorithm among the present invention;

Fig. 3 is a schematic diagram of road areas of three different widths in the present invention;

Fig. 4 is the test image that the present invention uses;

Fig. 5 is the image after denoising among the present invention;

Fig. 6 is the preliminary result figure of road detection among the present invention;

Fig. 7 is the result figure after the present invention uses optimization algorithm to remove false alarm;

Fig. 8 is a standard diagram of a test image.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in detail:

With reference to Fig. 1, the realization steps of the present invention are as follows:

Step 1: Enter data and perform speckle reduction.

The suppression of SAR coherence speckle is the key to SAR image processing, and also the premise of subsequent image recognition and other work.

The present invention uses the PPB algorithm to reduce the speckle of the input SAR image to obtain a clearer image of the road, so as to weaken the influence of noise, increase the contrast between the road and the background, and lay a foundation for accurate road detection features.

Step 2: Extract texture features.

SAR images contain rich texture features, if the texture information is used reasonably, it will be beneficial to road detection. According to the denoised image, the gray level co-occurrence matrix is constructed point by point, and then 9 kinds of texture features are extracted by the gray level co-occurrence matrix, namely: energy, entropy, contrast, dissimilarity, inverse gap, uniformity, mean, variance and correlation sex. The texture feature reflects the detailed information of the road, which is obviously different from the background. The road and the background can be distinguished through the texture feature.

(2a) Normalize the speckle-reduced image first, generate a gray-level co-occurrence matrix GLCM through the probability density function of the gray-level co-occurrence matrix GLCM, and use the generated gray-level co-occurrence matrix GLCM to describe the normalized SAR image,

The probability density function of the gray level co-occurrence matrix GLCM is as follows:

g(i,j)={(x ₁ ,y ₁ ),(x ₂ ,y ₂ )∈M×Nf(x ₁ ,y ₁ )=i,f(x ₂ ,y ₂ )=j},

Among them, g(i,j) represents the frequency of elements in the SAR image, also known as the GLCM probability density function, M is the horizontal dimension of the SAR image f(x,y), N is the vertical dimension of the SAR image f(x,y) Dimensions, (x ₁ ,y ₁ ) and (x ₂ ,y ₂ ) are two pixel points of the image SAR image f(x,y), f(x ₁ ,y ₁ ) and f(x ₂ ,y ₂ ) are the (x ₁ ,y ₁ ) and (x ₂ ,y ₂ ) grayscale values of the SAR image, i and j are the grayscale values of pixels f(x ₁ ,y ₁ ) and f(x ₂ ,y ₂ ) value;

(2b) According to the probability density function g(i, j) of the gray level co-occurrence matrix and the number of times that two gray pixels appear at the same time in the normalized SAR image when the distance is (Δx, Δy), generate the gray level co-occurrence matrix p (i, j), if the gray level is divided into w levels, a w×w gray level co-occurrence matrix can be obtained;

(2c) According to the generated gray level co-occurrence matrix p(i, j), calculate 9 common texture features:

(2c1) Calculate the energy feature Ene:

Energy is a parameter to measure the stable state of the gray value change in the image. It can describe the degree of image texture precision and the uniform state of the pixel gray value. Its calculation formula is as follows:

Among them, the p(i,j) gray level co-occurrence matrix, in the homogeneous area, because the gray level co-occurrence matrix values are mostly 1, and a few are 0, so the energy value is large, indicating that the current texture is a texture with regular changes and relatively stable texture; In the non-homogeneous area, the gray level co-occurrence matrix values are mostly 0 and a few are 1, so the energy value is small, indicating that the current texture is a texture with unstable changes.

(2c2) Calculate the entropy feature Ent:

Entropy is a measure to describe the amount of random information in an image, and its calculation formula is as follows:

When the matrix element values in the gray level co-occurrence matrix p(i,j) are relatively close or the distribution of element values in the matrix has the greatest uncertainty, the larger the entropy value, the more chaotic the image and the more complicated the texture;

(2c3) Calculate the contrast feature Con:

Contrast is a parameter to measure the number of local changes in the image, and its calculation formula is as follows:

The more local changes appear in the image, the greater the contrast value is, and the building area in the SAR image is mixed with dark spots between bright spots caused by specular reflection, so its contrast value is larger;

(2c4) Calculate the dissimilarity feature Dis:

Similarity is a parameter to measure the number of local changes in the image, and its calculation formula is as follows:

If the local contrast is higher, the dissimilarity is also higher;

(2c5) Calculate the inverse moment characteristic Idm:

Inverse moment: It is a parameter to investigate how much the matrix element values in the gray-level co-occurrence matrix are gathered on the main diagonal. The calculation formula is as follows:

If the inverse moment value is larger, it means that the image is more uniform;

(2c6) Calculate the homogeneity feature Hom:

Uniformity is a parameter that examines how much the matrix element values in the gray-level co-occurrence matrix are gathered on the main diagonal. The calculation formula is as follows:

The larger the uniformity value, the more uniform the image;

(2c7) Calculate the mean feature Mea:

Mean value: It is a parameter describing the matrix element value in the gray level co-occurrence matrix, and its calculation formula is as follows:

The mean value reflects the central tendency of the element values in the gray level co-occurrence matrix, and can usually be used as an estimated value of the matrix elements in the gray level co-occurrence matrix.

(2c8) Calculate the variance feature Var:

Variance: It is a parameter that describes the degree to which the matrix element values in the gray level co-occurrence matrix deviate from the mean value. Its calculation formula is as follows:

The variance and the mean together reflect the uniformity of the image;

(2c9) Calculate the correlation feature Cor:

Correlation: It is a parameter that describes the correlation between the gray values of each pixel in the image. The similarity between the elements of the gray co-occurrence matrix is examined from the row direction and the column direction, and the texture direction in the image can be identified. The calculation formula is as follows :

Step 3: Feature Screening

According to the imaging characteristics of the road area in the SAR image, only the features suitable for this type of SAR image are selected as representative features. On the one hand, it can reduce the interference of redundant useless information on useful feature information, and on the other hand, it can improve the detection efficiency. In the present invention, feature optimization and screening are carried out according to the Bhattacharyya distance. This method of selecting features is simple, fast and effective. The Bhattacharyya distance is defined as follows:

Among them, μ ₁ and σ ₁ represent the mean and variance of the pixel values of the first type of ground object on the same texture feature image respectively; μ ₂ and σ ₂ respectively represent the mean and variance of the pixel values of the second type of ground feature on the same texture feature image, The larger the BD value, the stronger the feature's ability to distinguish these two types of ground objects.

Due to the redundancy among texture features, in order to simplify the calculation, this example optimizes and screens the features according to the Bhattacharyya distance index. In the specific operation, select 3 of the 9 types of texture features that can distinguish the road and the background best. These features are mean u, variance var, and contrast cor.

Step 4: Extract ratio features

Texture features cannot fully express the characteristics of the road area in the SAR image. In order to extract features that accurately locate the road edge, the present invention extracts two kinds of ratio features point by point from the image after speckle reduction, that is, the distance between the road and the areas on both sides. Compare the ratio feature R1 with the similar ratio feature R2 between the regions on both sides of the road. The extraction steps are as follows:

4a) Extract a 15×15 image block Q point by point from the speckle-reduced image;

4b) Extract the central image block P of 7×7 in the center of the image block Q and calculate the contrast ratio feature R1 of the central pixel point:

R1＝min(max(u ₃ /u ₁ ,u ₁ /u ₃ ),max(u ₂ /u ₁ ,u ₁ /u ₂ ))

Among them, u ₁ represents the average value of pixels in area 1, u ₂ represents the average value of pixels in area 2, and u ₃ represents the average value of pixels in area 3.

Referring to Figure 3, calculate the situation of 3 different road widths and keep the maximum value; according to the resolution of the SAR image and the width of the actual road, the road in the SAR image is usually 1-3 pixels wide, where Figure 3 (a ) represents the case where the road width is 1 pixel, Fig. 3(b) represents the case where the road width is 2 pixels, Fig. 3(c) represents the case where the road width is 3 pixels, area 1 is the hypothetical road , area 3 is the background area on the left side of the road, and area 2 is the background area on the right side of the road. In the present invention, calculate the average ratio between area 1 and area 3, area 1 and area 2 respectively, that is, first get the maximum and then get the minimum. In this example, the mutual ratio of the area 1 mean and area 3 mean is the largest, and the area 1 mean and area 2 Take the maximum of the mutual ratio of the average value; then take the minimum from the above two maximum, that is, obtain the contrast ratio feature of the road area and the background area; calculate the value of the contrast ratio feature R1 under the three road width conditions and take the maximum value, Temporarily used as the contrast ratio feature of the central pixel of the central image block P;

4c) For the image block Q rotated 22.5 degrees counterclockwise around the center pixel point, perform step 4b) to obtain the contrast ratio feature R1 corresponding to the image block center pixel point at this angle, in order to find the direction closest to the road, A total of 7 rotations were performed, and a total of 8 direction contrast ratio features were obtained;

4d) According to the fact that the contrast between the road and the areas on both sides is relatively large, the greater the value of the contrast ratio feature R1, the greater the possibility that the pixel point is a road point. In this example, a total of 8 directions are considered, and 8 contrast ratios are compared. Feature R1 and keep the maximum value as the contrast ratio feature R1 of the central pixel point of the image block Q. Since the central image block P is extracted from the image block Q, R1 is also the contrast ratio feature of the central pixel point of the central image block P . During the calculation process, the image block Q when the contrast ratio feature R1 is the largest is retained, and the direction of the image block Q is regarded as the direction closest to the road.

4e) extracting a central image block P of 7×7 from the image block Q retained in step 4d);

4f) Calculate the similarity ratio feature R2 of the central pixel point of the image block P:

R2＝max(u ₃ /u ₂ ,u ₂ /u ₃ )

Among them, u ₂ represents the average value of pixels in area 2, and u ₃ represents the average value of pixels in area 3.

The similarity ratio feature R2 represents the size of the regional similarity on both sides of the road, with reference to Fig. 3, Fig. 3 represents the situation that the road width is 1-3 pixel points, calculates respectively the average value ratio of region 2 and region 3 in the present invention and gets big; Calculate the similarity ratio feature R2 of the three road widths respectively, and keep the minimum value as the similarity ratio feature of the center point of the image block. The closer the value of R2 is to 1, the more likely the center pixel point of the image block P is a road point. big.

Step 5: Construct a dictionary.

Extract 5 features including mean, variance, contrast, contrast ratio feature R1 and similarity ratio feature R2 from the input image, and normalize them to obtain a series of samples with 5-dimensional feature vectors. These samples are marked, One part belongs to the road area in the image, and the other part belongs to the background area in the image. Construct road dictionary D ₁ and background dictionary D ₂ with these samples;

Step 6: Preliminary testing.

Respectively solve the difference E1 between each pixel point and the road dictionary _D1 and the difference _E2 between each pixel point and the background dictionary D2, and _classify the pixels according to these _two differences _E1 and _E2 . Determine whether the pixel belongs to the road area or the background area.

Calculate the difference E ₁ with the road dictionary D ₁ and the difference E ₂ with the background dictionary D ₂ for the input samples respectively by the following formula:

Among them, x is the input test sample, is the kth atom of the road dictionary D ₁ , and p represents the number of atoms in the dictionary D ₁ ; Is the _lth atom of the background dictionary D2, and q represents the number of atoms in the dictionary D2 _;

Classify the pixels of the test sample by comparing the size of the two differences, and get the preliminary results of road detection:

If E ₁ -E ₂ <0, it means that the pixel point of the test sample has a stronger correlation with the atom in the road area, and the pixel point is initially judged to belong to the road area;

If E ₁ -E ₂ ≥ 0, it means that the pixel point of the test sample has a stronger correlation with the atoms in the background area, and it is preliminarily determined that the pixel point belongs to the background area;

The preliminary detection obtained in this step is shown in Figure 6.

Step 7: Result optimization.

The general outline of the road can be obtained through preliminary detection. How to determine the real road and eliminate false alarm interference is a critical step. The present invention is based on the characteristics of the road itself, that is, the road is slender, and its outline is a closed curve. The closer the shape of this curve is to the slender type, the smaller the area, the more in line with the characteristics of the road. According to the integral principle, the present invention proposes The optimization algorithm of the area-to-perimeter ratio is as follows to effectively eliminate false alarm areas.

Referring to Figure 2, the implementation of this step is as follows:

7a) Input a preliminary detection result graph I to be optimized, see FIG. 6, its size is m×n, take the center of the preliminary detection result graph I as a fixed point, and expand the size of the image to (m+2)×(n +2), set the expanded border to 0, which is the background, and the expanded image is Initialize the number S ^* =0 of the pixel points of the road and the perimeter L ^* =0 of the connected domain in the connected domain;

7b) For the expanded image Start scanning from pixel point (2,2), skip all boundary points until encountering the first pixel point (x, y) with a pixel value of 1, let S ^* = 1, pixel point (x, y) The pixel value of is set to 3;

7c) Take the 4 neighborhoods of the pixel point (x, y), that is, in the 3*3 window with (x, y) as the center point, find the number of points with pixel values of 0 in the 4 neighborhoods of up, down, left, and right m, update L ^* with the value of L ^* + m;

7d) Take its 8 neighborhoods for the pixel point (x, y), that is, find the number n of points with a pixel value of 1 in the surrounding 8 neighborhoods in a 3*3 window with (x, y) as the center point, Update S ^* with the value of S ^* +n, change the n pixel values from 1 to 3 one by one, and record the positions of these pixels;

7e) For the n pixels recorded in step 7d), repeat step 7c) and step 7d) one by one until no pixel value 1 can be found in the pixel point (x, y) and its 8 neighborhoods, that is Traversing a connected domain, record the number S ^* of pixels in this connected domain that are roads and the perimeter L ^* of the connected domain. If the side length of the pixel is set to 1, the number of pixels S ^* will be Equivalent to the area of the road in the connected domain;

7f) Set the threshold value T, calculate the road approach rate with the updated S ^* and the updated L ^* Then compare the road approach rate with the threshold value: if the road P ^* >T, then change those points whose pixel value is 3 recorded in step 7d) to 0, otherwise unchanged, reinitialize S ^* =0, L ^* = 0;

The setting of the threshold value is a floating value, which can be flexibly selected according to the detection requirements. If the value of T is set smaller, the effect of removing false alarms will be better, but at the same time the missed detection rate will be higher. If the value of T If the value is set larger, the detected road is relatively more complete, but the effect of removing false alarms is relatively poor. In this example, the threshold value T is set to 0.35;

7g) For the expanded image Start scanning from the pixel point (x+1, y+1), repeat step 7b) to step 7f), until the last pixel point (m, n) of the preliminary detection result image; all the expanded images Change the value of the point with a pixel value of 3 to 1, remove the expanded boundary, and obtain the optimized road detection result of the image to be detected, as shown in Figure 7.

The effects of the present invention can be further illustrated by the following simulation results.

1. Simulation conditions

The present invention carries out simulation experiments under the programming environment of Visual Studio 2010 and in combination with MFC technology.

The size of the detected SAR image is 329×329 and the resolution is 4 meters, as shown in Figure 4.

2. Simulation content

In simulation 1, the SAR image shown in FIG. 4 of the present invention is used for PPB speckle reduction processing, and the result is shown in FIG. 5 . As can be clearly seen from Fig. 5, the image after speckle reduction in the present invention is clearer, and the road features are more obvious, which lays a good foundation for the next step of extracting the features that can accurately represent the road;

In simulation 2, the present invention is used to classify each pixel of the speckle-reduced image, as shown in Figure 5, to obtain a preliminary detection result map, as shown in Figure 6. From the detection results in Figure 6, the general outline and direction of the road have been detected preliminarily, but there is still the problem of wrongly detecting non-road points as road points, so it is necessary to post-process the preliminary detection results by adopting an optimization algorithm of;

Simulation 3, the present invention combines the characteristics of the road itself, that is, the roads are all slender linear objects, and the preliminary detection result diagram shown in Figure 6 is optimized with the optimization algorithm of the area-to-perimeter ratio to effectively remove isolated The interference points and blocky false alarm areas, the results are shown in Figure 7. It can be seen from the results in Figure 7 that the detection results after the optimized algorithm exclude some interference points and blocky false alarm areas in the preliminary detection result map;

Comparing FIG. 7 with the standard image corresponding to the test image shown in FIG. 8, it can be seen that the road detection result of the present invention is highly consistent with the standard image, indicating that the present invention has a good road detection effect.

Claims (6)

1. a kind of SAR image Approach for road detection based on ratio feature, comprising:

(1) SAR image to be detected is read in, and drop spot pretreatment is carried out to SAR image；

(2) gray level co-occurrence matrixes are constructed to the image after drop spot point by point, and calculate energy, entropy, comparison with gray level co-occurrence matrixes Degree, mean value, variance, correlation, non-similarity, unfavourable balance away from 9 kinds of textural characteristics of uniformity；

(3) the 9 kinds of textural characteristics extracted above are optimized and is screened according to Bhattacharyya range index, only chosen Can be respectively: mean value, variance and comparison effectively to Road Detection in SAR image and maximum 3 kinds of the feature of classification contribution Degree；

(4) ratio feature is extracted:

(4a) extracts two kinds of ratio features to the image after drop spot point by point, i.e., the comparison ratio feature between road and two side areas Similar ratio feature R2 between R1 and both sides of the road region；

The 3 kinds of textural characteristics extracted in step (3) and step (4a) are extracted two kinds of ratio features and are normalized by (4b), are obtained 5 dimensional feature vectors of each pixel；

(5) pixel of part tape label is randomly choosed as sample, including road waypoint and non-rice habitats point, is gone out with the sample architecture Road dictionary D₁With background dictionary D₂；

(6) Preliminary detection:

(6a) solves each pixel and road dictionary D respectively₁Difference E₁With each pixel and background dictionary D₂Difference E₂；

(6b) is according to the two differences E₁And E₂Classify to pixel: if E₁-E₂< 0, indicate test sample pixel and The Atomic Correlations of road area are stronger, then are tentatively judged to pixel and belong to road area；If E₁-E₂>=0, indicate test sample Pixel and background area Atomic Correlations it is stronger, then be tentatively judged to pixel and belong to background area；

(7) Preliminary detection that obtains for step (6) is as a result, leptosomatic feature further according to road, with area perimeter ratio Optimization algorithm excludes false-alarm region, obtains the final result of Road Detection.

2. according to the method described in claim 1, wherein in step (3) according to Bhattacharyya range index to step (2) The 9 kinds of textural characteristics extracted optimize screening, are to optimize screening to every kind by following range formula:

Wherein, BD indicates Bhattacharyya range index, μ₁、σ₁Respectively indicate first kind atural object on same texture template image The mean value and variance of pixel value；μ₂、σ₂Respectively indicate the mean value of the second class atural object pixel value and side on same texture template image Difference.

3. being carried out as follows according to the method described in claim 1, wherein calculating comparison ratio feature R1 in step (4):

4a) extract 15 × 15 image block Q point by point to the image after drop spot；

4b) in the center image block P of image block Q center extraction 7 × 7, the feelings that road width is 1,2,3 pixel are calculated separately Condition obtains 3 comparison ratio features, and takes maximum value therein, temporarily special as the comparison ratio of image block P central pixel point Sign；

22.5 degree 4c) are rotated counterclockwise centered on central pixel point to image block Q, return step 4b), a corotating 7 times obtains To 8 comparison ratio features；

Maximum value 4d) is found out from 8 comparison ratio features, the comparison ratio feature R1 as image block P central pixel point:

R1=min (max (u₃/u₁,u₁/u₃),max(u₂/u₁,u₁/u₂)),

Wherein, u₁Indicate the mean value of pixel in region 1, u₂Indicate the mean value of pixel in region 2, u₃Indicate pixel in region 3 The mean value of point；

The image block Q in corresponding direction when retaining comparison ratio feature R1 maximum simultaneously, as closest to the direction of road.

4. being carried out as follows according to the method described in claim 1, wherein calculating likelihood ratio value tag R2 in step (4):

7 × 7 center image block P 4e) is extracted in the image block Q retained from step (4d)；

4f) calculate separately road width be 1,2,3 pixel the case where, obtain 3 similar ratio features, and take it is therein most Small value, the similar ratio feature R2 as image block P central pixel point:

R2=max (u₃/u₂,u₂/u₃),

Wherein, u₂Indicate the mean value of pixel in region 2, u₃Indicate the mean value of pixel in region 3.

5. according to the method described in claim 1, wherein solving each pixel and road dictionary D₁Difference E₁With each pixel Point and background dictionary D₂Difference E₂, it is carried out respectively with following formula:

Wherein, x is the test sample of input,For road dictionary D₁K-th of atom, p indicate dictionary D₁In atom number；For background dictionary D₂First of atom, q indicate dictionary D₂In atom number.

6. according to the method described in claim 1, wherein excluding false-alarm area with the optimization algorithm of area perimeter ratio in step (7) Domain carries out as follows:

Preliminary detection result images I 7a) is inputted, it is fixed point with the center of Preliminary detection result images I that size, which is m × n, It is (m+2) × (n+2) that the size of image, which is expanded, will expand the boundary come out and is both configured to 0, as background, the image after expansion ForIt initializes in connected domain as the number S of the pixel of road^*=0 and connected domain perimeter L^*=0；

7b) to image after expansionStart to scan from pixel (2,2), skip all boundary points, until encountering first pixel The pixel (x, y) that value is 1, enables S^*=1, the pixel value of pixel (x, y) is set as 3；

Its 4 neighborhoods 7c) are taken to pixel (x, y), are searched the number m for the point that pixel value is 0 in 4 neighborhoods, are used L^*The value of+m updates L^*；

Its 8 neighborhoods 7d) are taken to pixel (x, y), are searched the number n for the point that pixel value is 1 in 8 neighborhoods, are used S^*The value of+n updates S^*, the pixel value of this n point is changed to 3 one by one, and record the position where these pixel values；

Step 7c 7e) is repeated to above-mentioned n point) and step 7d), it can not find until in pixel (x, y) and its 8 neighborhoods The point that pixel value is 1 obtains in this connected domain as the number S of the pixel of road^*And the perimeter L of connected domain^*；

Threshold T 7f) is set, with updated S^*With updated L^*It calculates road and approaches rateRoad is approached again Rate is compared with threshold value: if road P^*> T is then to be changed to 0 inner the recorded pixel value of step 7d) those of 3 o'clock, no It is then constant, reinitialize S^*=0, L^*=0；

7g) to image after expansionStart to scan from pixel (x+1, y+1), repeat step 7b) arrive step 7f), Zhi Daochu The last one pixel (m, n) for walking detection result image I, image after all expansionsThe value weight for the point that middle pixel value is 3 It newly is changed to 1, removes image after expansionBoundary, obtain the final optimization pass result of Road Detection.