CN110782471A - Multi-scale SAR image edge detection method - Google Patents

Abstract

The invention discloses a multi-scale SAR image edge detection method, which uses windows of different sizes to calculate edge gradients in different directions in an image, and then retains the maximum value among all gradients of different sizes and directions and its corresponding direction information Finally, an edge image with a final width of one pixel is obtained through non-maximum suppression. This method can use windows of different sizes and directions to detect edges, effectively overcoming the defect of inaccurate edge positions obtained during edge detection of a single window.

Description

A Multi-scale SAR Image Edge Detection Method

technical field

The invention belongs to the technical field of image processing, and in particular relates to a multi-scale SAR image edge detection method.

Background technique

The edge in the image is the most basic and difficult to change feature in the image. It is the place where the information in the image is most concentrated. It is the basis for image feature extraction, target recognition, image segmentation and other processing. Therefore, edge detection is the most important in the field of image processing. one of the basic questions. However, due to the influence of factors such as sensor, imaging principle, imaging position and imaging object in the imaging process, the performance of edge information in the image varies widely, which makes it difficult to accurately detect the edge.

The current image edge detection algorithms can be roughly divided into: 1) edge detection algorithm based on spatial gradient; 2) wavelet edge detection algorithm based on frequency domain; 3) edge detection algorithm based on machine learning; 4) edge detection algorithm based on other theoretical techniques detection algorithm. Among these methods, the edge detection algorithm based on spatial gradient is the most classic and the most influential, and among them, the Canny edge detection algorithm is the most typical. However, the detection windows of these algorithms are all fixed size, and the performance is poor when the image noise level is high.

In the field of Synthetic Aperture Radar (SAR) image processing, due to the strong speckle noise, it is difficult for general edge detection algorithms to achieve ideal results. The commonly used edge detector is the Ratio of Average (ROA) detector. The detector is a spatial gradient-based detector, and its detection window and direction can be adjusted manually as parameters, and it has better noise immunity because the regional mean is used instead of a single pixel value. However, the window size of the detector is still single, and for an image, the appropriate detection window size may be different in different regions. For example, in areas with rich detailed information, it is generally more suitable to use a small window for detection, because when a larger window is used for edge detection, the detected edge positions around some points, lines and other objects will have a large deviation; Areas with weak information and more homogeneous pixels are generally more suitable for detection with large windows, because small windows are more sensitive to noise and detect false edges. Therefore, a detection window of a single size cannot meet the requirements at the same time.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the defect that the conventional ROA edge detector cannot meet the requirements by using a single-sized detection window, and proposes a multi-scale SAR image edge detection method that comprehensively utilizes detection windows of different sizes.

Technical solution: In order to achieve the purpose of the present invention, the technical solution adopted in the present invention is: a multi-scale SAR image edge detection method, comprising the following steps:

Step 1: Set the maximum detection window W _max and the minimum detection window W _min to obtain N _W =(W _max -W _min )/2+1 windows W of different sizes;

Step 2: For N _W windows of different sizes, set the detection template parameters l, w, d, θ respectively, and each window of size W can obtain N _θ = π/θ templates, then N=N _W *N _θ templates of different sizes W and different directions of θ; wherein, l is the length of the template, w is the width of the template, d is the distance between the template regions R ₁ and R ₂ , and θ is the angle corresponding to the template;

Step 3: For each pixel p, use N templates of different sizes W and different directions θ to calculate the gradient D _W (θ) between the template regions R ₁ and R ₂ , and obtain N gradient values and corresponding angles in total. value;

Step 4: Select the maximum value among the N gradient values, record the value and its corresponding angle, and continue to process the next pixel until all pixels are processed to obtain a gradient image and an angle image;

Step 5: Perform non-maximum suppression in the gradient direction to obtain a refined gradient image; the details are as follows:

For each pixel p in the gradient image, along the normal direction of the angle indicated in the corresponding angle image, compare the magnitude of the adjacent pixel values; if the pixel p is smaller than its adjacent pixels, set the gradient value of the pixel p as 0, otherwise, keep its original value, and obtain a refined gradient image after all pixels are processed;

Step 6: Binarize the refined gradient image to obtain a final edge image with a width of one pixel.

Further, in the step 3, for the SAR image, the gradient calculation formula is:

Wherein, x ₁ and x ₂ represent the SAR image amplitude or intensity of template region R ₁ and region R ₂ in the direction θ, respectively;

For polarimetric SAR images, the gradient is calculated as:

D _W (θ)=2ln|X ₁ +X ₂ |-ln|X ₁ |-ln|X ₂ |+2qln2

Wherein, X ₁ and X ₂ represent the q×q polarimetric SAR image coherence matrix or covariance matrix of template region R ₁ and region R ₂ in the direction θ, respectively, and q represents the matrix dimension.

Further, in the step 6, the entropy threshold method is used to perform binarization processing, which specifically includes:

Step 6.1: Perform image enhancement processing on the refined gradient image as follows:

First, logarithmically transform the refined gradient image, and the calculation formula is: x=10*log ₁₀ (x), where x represents the pixel value;

Then, stretch the maximum value and the minimum value. The processing method is: sort the image pixel values in ascending order, take the pixel value at the a% position as the minimum value min, and take the pixel value at the b% position. is the maximum value max;

Finally, the maximum value and the minimum value are used for normalization, and the normalization formula is: x=(x-min)/(max-min), where x represents the pixel value;

Step 6.2: Divide the pixels in the gradient image processed in Step 6.1 into n gray levels, and calculate the ratio p _i of the pixels in each gray level i to the total number of pixels;

Step 6.3: Set the gray level s to divide the image into two parts A and B, and the entropy of each part is defined as follows:

Then the sum of the entropy of A and B is: H _sum =H _A +H _B , calculate the H _sum corresponding to each gray level i;

Step 6.4: Find the largest H _sum among the n gray levels and its corresponding gray level s, then s is the segmentation threshold, set the pixels less than or equal to s to 0, and set the pixels greater than s to 1, That is, a binarized image is obtained.

Beneficial effects: compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

The invention proposes to use window templates of different sizes to perform edge detection on the image, and to synthesize the edge information obtained by each window to obtain the final edge detection result, thereby overcoming the inability of conventional single-window edge detection to take into account texture-rich regions and weak edge regions It can not only detect possible edges to the greatest extent, but also ensure the accuracy of edge detection.

Description of drawings

Fig. 1 is the flow chart of the present invention;

FIG. 2 is an edge detection template diagram in an embodiment of the present invention.

Detailed ways

The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

As shown in Figure 1, a method for detecting edges of a multi-scale SAR image according to the present invention includes the following steps:

Step 1: Set the maximum detection window W _max and the minimum detection window W _min to obtain N _W =(W _max -W _min )/2+1 windows W of different sizes;

In this embodiment, W _max =7 and W _min =3 are set, then there are N _W =3 windows of different sizes in total, that is, W takes 3, 5, and 7.

Step 2: For 3 windows of different sizes, set their template parameters l, w, d, θ respectively, as shown in Figure 2; each window of size W can obtain N _θ = π/θ templates, then a total of Obtain N=N _W *N _θ templates of different sizes W and different directions θ; where l is the length of the template, w is the width of the template, d is the distance between the template regions R ₁ and R ₂ , and θ is the angle corresponding to the template ; In general, the template is a square, that is, l=2*w+d;

In this embodiment, when the window size is 3, set l=3, w=1, d=1, θ=45°; when the window size is 5, set l=5, w=2, d=1, θ=45°; when the window size is 7, set l=7, w=3, d=1, θ=45°; then a total of N=3*(180°/45°)=12 different sizes, different Orientation template.

Step 3: For each pixel p, use 12 templates of different sizes and different directions to calculate the gradient D _W (θ) between the template regions R ₁ and R ₂ , and obtain a total of 12 gradient values and corresponding angle values;

In this embodiment, for the SAR image, the calculation formula of the gradient described in step 3 is:

Among them, x ₁ and x ₂ represent the SAR image amplitude or intensity of template region R ₁ and region R ₂ in the direction θ, respectively;

In this embodiment, for polarimetric SAR images, the calculation formula of the gradient described in step 3 is:

D _W (θ)=2ln|X ₁ +X ₂ |-ln|X ₁ |-ln|X ₂ |+2qln2

Wherein, X ₁ and X ₂ represent the q×q polarimetric SAR image coherence matrix or covariance matrix of template region R ₁ and region R ₂ in the direction θ, respectively, and q represents the matrix dimension.

Step 4: Calculate the maximum value of the 12 gradient values, record the value and its corresponding angle, and continue to process the next pixel until all pixels are processed. After all pixels are processed, a gradient image and a gradient image are obtained. angle image.

Step 5: Perform non-maximum suppression in the gradient direction to obtain a refined gradient image; the details are as follows:

For each pixel p in the gradient image, along the normal direction of the angle indicated in the corresponding angle image, compare the magnitudes of the adjacent pixel values, if the pixel p is smaller than its adjacent pixels, set the gradient value of the pixel p as 0, otherwise, keep its original value, and get a refined gradient image after all pixels are processed.

Step 6: Binarize the refined gradient image to obtain a final edge image with a width of one pixel;

In this embodiment, the entropy threshold method is used for binarization, which specifically includes:

Step 6.1: Perform image enhancement processing on the refined gradient image as follows:

First, logarithmically transform the refined gradient image, and the calculation formula is: x=10*log ₁₀ (x), where x represents the pixel value;

Then, stretch the maximum value and the minimum value. The processing method is: sort the image pixel values in ascending order, take the pixel value at the 6.25% position as the minimum value min, and take the pixel value at the 93.75% position. is the maximum value max;

Finally, the maximum value and the minimum value are used for normalization, and the normalization formula is: x=(x-min)/(max-min), where x represents the pixel value;

Step 6.2: divide the pixels in the gradient image processed in step 6.1 into n gray levels, and calculate the ratio p _i of the pixels in each gray level i to the total number of pixels; n=256 in this embodiment;

Step 6.3: Suppose the gray level s divides the image into two parts A and B, and the entropy of each part is defined as follows:

Then the sum of the entropy of A and B is: H _sum =H _A +H _B , calculate the H _sum corresponding to each gray level i;

Step 6.4: Find the largest H _sum among the n gray levels and its corresponding gray level s, then s is the segmentation threshold, set the pixels less than or equal to s to 0, and set the pixels greater than s to 1, That is, a binarized image is obtained.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention pertains can make various modifications or additions to the specific embodiments described or substitute in similar manners. For example, the template parameter θ can be set to other values; the gradient can be replaced by other formulas; the binarization method can choose other methods to replace the entropy threshold method; and so on. However, this does not go beyond the framework of the proposed algorithm of the present invention, does not deviate from the spirit of the present invention, or goes beyond the scope defined by the appended claims.

Claims (6)

1. A multi-scale SAR image edge detection method is characterized in that: the method comprises the following steps:

step 1: setting a maximum detection window W _maxAnd a minimum detection window W _minObtaining N _W＝(W _max–W _min) 2+1 windows W of different sizes;

step 2: for N _WSetting detection template parameters l, W, d and theta of windows with different sizes, and obtaining N for each window with the size of W _θPi/theta templates, then obtaining N-N _W*N _θTemplates with different sizes W and different directions theta; wherein l is the template length, w is the template width, d is the template region R ₁And R ₂The distance between the two plates, theta is the angle corresponding to the template;

and step 3: for each pixel p, calculating a template region R by using N templates with different sizes W and different directions theta ₁And R ₂Gradient D between _W(θ), obtaining N gradient values and corresponding angle values in total;

and 4, step 4: selecting the maximum value of the N gradient values, recording the value and the corresponding angle thereof, and continuously processing the next pixel until all the pixels are processed to obtain a gradient image and an angle image;

and 5: carrying out non-maximum suppression in the gradient direction to obtain a refined gradient image;

step 6: and carrying out binarization processing on the thinned gradient image to obtain a final edge image with a single-pixel width.

2. The edge detection method of the multi-scale SAR image according to claim 1, characterized in that: in step 3, for the SAR image, a gradient calculation formula is:

wherein x is ₁、x ₂Respectively representing the template areas R in the direction theta ₁And region R ₂The SAR image amplitude or intensity;

for a polarized SAR image, the formula for the gradient calculation is:

D _W(θ)＝2ln|X ₁+X ₂|-ln|X ₁|-ln|X ₂|+2q ln2

wherein, X ₁、X ₂Respectively representing the template areas R in the direction theta ₁And region R ₂Q × q, q representing the matrix dimension.

3. The edge detection method of the multi-scale SAR image according to claim 1, characterized in that: in the step 5, performing non-maximum suppression in the gradient direction to obtain a refined gradient image; the method comprises the following specific steps:

for each pixel p in the gradient image, comparing the magnitude of adjacent pixel values in the direction normal to the angle indicated in the corresponding angle image; if the pixel p is smaller than the adjacent pixels, the gradient value of the pixel p is set to be 0, otherwise, the original value of the pixel p is reserved, and the refined gradient image is obtained after all the pixels are processed.

4. The edge detection method for the multi-scale SAR image according to any one of claims 1-3, characterized in that: in the step 6, an entropy threshold method is adopted for binarization processing, which specifically includes:

step 6.1: carrying out image enhancement processing on the refined gradient image;

step 6.2: dividing the pixels in the gradient image processed in the step 6.1 into n gray levels, and calculating the ratio p of the pixels in each gray level i to the total number of pixels _i；

Step 6.3: the image is divided into two parts A and B by a gray level s, and the entropy of each part is respectively H _AAnd H _B(ii) a The sum of the entropies of a and B is then: h _sum＝H _A+H _BCalculating H corresponding to each gray level i _sum；

Step 6.4: finding the largest H among n gray levels _sumAnd the corresponding gray level s is obtained, s is the segmentation threshold, the pixels less than or equal to s are set as 0, and the pixels more than s are set as 1, so that the binary image is obtained.

5. The edge detection method of the multi-scale SAR image according to claim 4, characterized in that: the step 6.1 is to perform image enhancement processing on the refined gradient image, and the method comprises the following steps:

firstly, carrying out logarithmic transformation on the refined gradient image, wherein the calculation formula is as follows: x is 10 log ₁₀(x) Wherein x represents a pixel value;

then, maximum value and minimum value stretching is carried out, and the processing method comprises the following steps: sorting image pixel values in a descending order, taking the pixel value at the a% position as a minimum value min, and taking the pixel value at the b% position as a maximum value max;

and finally, carrying out normalization processing by using the maximum value and the minimum value, wherein the normalization formula is as follows: x is (x-min)/(max-min), where x represents the pixel value.

6. The edge detection method of the multi-scale SAR image according to claim 4, characterized in that: in step 6.3, the entropy of the two parts a and B corresponding to the gray level s is defined as follows:

Images