A two-stage SAR image edge detection method based on saliency

Abstract

The application discloses a two-stage SAR image edge detection method based on saliency, and belongs to the technical field of image processing. Anisotropic filtering is used to smooth the input image, so as to filter out noise and retain edge information as much as possible. A Gauss difference operator is used to highlight edge regions and suppress flat regions; a Gabor filter bank and maximum processing are used to generate an edge saliency map, on the saliency map, edge regions have higher saliency values, and other regions are suppressed more thoroughly. Non-maximum suppression and double-threshold processing are used to perform edge rough extraction from the SME, so as to obtain primary edges containing false alarms. A neighborhood discrimination strategy is used to eliminate false alarms from the primary edges, so as to obtain a final edge detection result. Therefore, the application has good scalability, and the detection result has excellent accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to a two-stage SAR image edge detection method based on saliency. BACKGROUND

[0002] An edge usually exists between two regions with different characteristics, and the gray scale of the corresponding pixels thereof presents obvious discontinuity in spatial distribution. The edge contains a large amount of key information and is one of the most basic features of an image, and can play an important role in various applications such as image segmentation, linear target detection, feature description and noise suppression. Due to the importance of the edge, the edge detection has always been a key issue in SAR image processing. At present, the algorithms for SAR edge extraction can be mainly divided into three categories: a mean ratio algorithm based on a window, a method based on an active contour model and a method based on wavelet transform.

[0003] The mean ratio algorithm based on the window measures the edge intensity response of the current pixel by solving the ratio of the mean values of two sub-windows. Such an algorithm has a constant false alarm rate (CFAR) feature, and can resist the influence of noise to a certain extent. However, as a CFAR detection, the mean ratio method also faces the problem of clutter modeling. If the model and the actual clutter are mismatched, the edge detection based on the mean ratio will suffer from CFAR loss. The method based on the active contour model (ACM) can obtain a continuous, smooth and closed boundary; however, it has a high computational complexity, and is sensitive to the position of the initial contour, and often needs to use other means to place the initial contour near the target region. These deficiencies limit the application of the ACM method in actual tasks to a certain extent. The method based on wavelet transform is difficult to directly obtain the edge. It is more similar to filtering, and must be combined with other subsequent methods to realize edge extraction.

[0004] In summary, the current SAR edge detection methods all have some defects, and it is necessary to try to start from a new angle to better solve the SAR edge extraction problem. SUMMARY

[0005] Based on the current situation that the traditional SAR edge detection algorithm has many defects, the present application proposes a new two-stage SAR image edge detection method based on saliency by referring to the visual saliency mechanism in the field of biological vision.

[0006] SAR images contain a large amount of texture noise, and conventional filtering methods can blur object boundaries. Therefore, this invention first employs anisotropic filtering to smooth the input image, preserving edge information as much as possible while filtering out noise. Secondly, it uses the difference of Gaussian operator to highlight edge regions and suppress flat regions; then, it uses a Gabor filter bank and extremum optimization to generate a saliency map (SME). On the saliency map, edge regions have high saliency values, while other regions are more thoroughly suppressed. Next, it further uses non-maximum suppression (NMS) and dual-threshold processing to coarsely extract edges from the SME, obtaining primary edges containing false alarms. Finally, this invention employs a neighborhood discrimination strategy to remove false alarms from the primary edges, obtaining the final edge detection result.

[0007] Therefore, the technical solution of the present invention is: a two-segment SAR image edge detection method based on saliency, characterized in that it specifically includes the following steps:

[0008] Step 1: Input the original SAR image I, and denote the size of I as... Anisotropic diffusion is used to filter I, resulting in a smoothed image I. AD ;

[0009] Step 2: Using the difference of Gaussians operator and I AD Convolution yields I DoG ;

[0010] Step 3: Use Gabor filter banks to filter I DoG Filtering is performed to obtain the edge saliency map SME and the edge orientation map OME. On the edge saliency map SME, the higher the saliency value, the greater the probability that the corresponding position is an edge.

[0011] Step 4: Perform Non-Maximum Suppression (NMS) processing on SME to obtain SME NMS ;

[0012] Step 5: Based on SME NMS Perform dual-threshold processing to obtain the primary edge:

[0013] Step 6: Perform neighborhood discrimination processing on the primary edges to obtain the final edge detection results.

[0014] Furthermore, the specific method for step 3 is as follows:

[0015] Step 3-1: Direct the Gabor filter The values ​​are set sequentially to 0, π / 8, ..., 7π / 8, resulting in 8 filter templates G0, G1, ..., G7.

[0016] Step 3-2: Use the template G i and I DoG Convolution, respectively, get the filter results in each direction I G_i , as shown below;

[0017] ;

[0018] where i = 0, 1, 2,···,7, * indicates convolution;

[0019] Step 3-3: On the edge saliency map SME, the saliency value of the point at position (m, n) is calculated as follows:

[0020] ;

[0021] where, represents the filter result of the point at position range (m, n), represents the maximum value of ;

[0022] Step 3-4: On the edge direction map OME, the value OME(m, n) of the point at position (m, n) is calculated as follows:

[0023] ;

[0024] That is, the Gabor filter bank acts on I DoG , the direction corresponding to the filter with the maximum modulus of the filter result is OME(m, n); if there is an edge at (m, n) of the image I, OME(m, n) represents the direction of the edge here, otherwise, OME(m, n) is considered as an invalid value.

[0025] Further, the specific method of the step 5 is:

[0026] Step 5-1: First, calculate the high threshold T H and the low threshold T L ;

[0027] ;

[0028] ;

[0029] where STD( ) is the standard deviation of all elements in the matrix;

[0030] mean( ) is the mean of all elements in the matrix;

[0031] Step 5-2: SME NMSregions with the upper saliency value exceeding are directly determined as edges;

[0032] Step 5-3: SME NMS regions with the upper saliency value below are directly determined as non-edges;

[0033] Step 5-4: SME NMS regions with the upper saliency value in the interval are also determined as edges if they are connected with the edge regions in Step 5-2; the edges obtained at this time are called primary edges, which contain false alarms, indicating false edge recognition.

[0034] Further, the specific method of Step 6 is as follows:

[0035] Let the primary edge pixel set E init = {(r1, c1), (r2, c2), …, (r p , c p )}, (r p , c p ) represents the pixel position coordinates;

[0036] Step 6-1: Calculate the discrimination threshold , ;

[0037] Step 6-2: Take the i-th element in the set E init , and construct a Gabor neighborhood reference window W i in the direction of the primary edge at (r i , c i ), 1≤i≤p; the specific steps are as follows:

[0038] Step 6-2-a: Construct a Gabor template with as the Gabor filter direction;

[0039] Step 6-2-b: , set the elements not less than 0.5 to 1, the elements not more than -0.5 to -1, and the elements in the interval to 0 to generate the neighborhood reference window W i ; the region composed of the elements 1 in W i is the positive domain ‘A+’, and the region composed of the elements -1 is the negative domain ‘A-’;

[0040] Step 6-3: Take a 11x11 size region from I with as the center, and mark it as ;

[0041] Step 6-4: Obtain The mean of all elements corresponding to the 'A+' region is denoted as m. + ; obtain The mean of all elements corresponding to the 'A-' region is denoted as m. - ;

[0042] Step 6-5: If or Then E init If the i-th element is a real edge, it is retained; otherwise, it is a false alarm and is filtered out.

[0043] Step 6-6: Return to step 6-2, extract the (i+1)th element, and continue until E is traversed. init By analyzing all elements in the dataset, we can identify all primary edge pixels and obtain the final edge detection result.

[0044] In summary, this invention can be divided into two stages. The first stage generates an edge saliency map to highlight edge components in the SAR image and suppress other regions. The second stage uses the edge saliency map to achieve refined edge extraction, obtaining the final detection result. The beneficial effects of this invention are as follows:

[0045] 1) It generates an edge saliency map that highlights only the edge regions while suppressing other locations. It can serve as the basis for other edge detection methods and has good scalability.

[0046] 2) An identification strategy was designed to filter out false alarms from the detection results, so that the final detection results have excellent accuracy. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the structure of the present invention.

[0048] Figure 2 The original image generated for the experiment during simulation.

[0049] Figure 3 for Figure 2 The corresponding edge label.

[0050] Figure 4 This invention is aimed at Figure 2 The generated edge saliency map.

[0051] Figure 5 This invention is aimed at Figure 2 The final edge detection result obtained. Detailed Implementation

[0052] The following points are worth noting during the implementation process:

[0053] 1) The DoG operator is used in step 2, and its formula is shown below.

[0054]

[0055] Where DoG(x, y) represents the coefficient at (x, y) on the DoG template; and Set them to 1 and 1.2 respectively; set the template size to 23x23, i.e. ;Will and Substitute the value into the DoG formula and iterate through it. Take the value, generate a 23x23 DoG template, and use the template and I AD Convolution yields I DoG ;

[0056] 2) A Gabor filter was used in step 3, and its formula is shown below.

[0057]

[0058] in, Indicates Gabor template Filter coefficients at the location; Let the phase offset be set as follows: ; The spatial aspect ratio of the Gabor filter shape was determined and set to 0.8; Let the Gaussian kernel standard deviation be 4; This reflects the wavelength of the sine wave, set to 15. Indicates the filter direction; the filter template size is set to 31x31, i.e.: ; Traversal By taking values ​​and substituting the above parameters into the Gabor formula, we can obtain a 31x31 matrix with the following dimensions and direction: Gabor template; Let's take the values ​​as 0 and π / 8 respectively. 7π / 8 gives us filter templates G0, G1, and G1 in 8 directions. G7;

[0059] 3) In step 6, a Gabor filter is used to generate a neighborhood reference window. At this time, the parameter settings in formula (6) are as follows: Set as ; Set to 1; Set to 2; Set to 15; set the filter template size to 11x11, that is: ; According to the OME value, for ; traverse (x, y), generate a Gabor template with a size of 11x11 ;

[0060] 4) Step 2, use the DoG template and I AD When convolving, to ensure that the result after convolution has the same size as the input image, still , I AD needs to be expanded; here, the mirror padding method is used to expand I AD to size; in this way, I DoG will remain size;

[0061] 5) Step 3, filter I DoG using the Gabor filter set; to ensure that I DoG remains size after Gabor filtering, I DoG needs to be expanded; before filtering, the mirror padding method is used to expand I DoG to size;

[0062] 6) When performing neighborhood discrimination in step 6, if the primary edge pixel to be judged is located near the image boundary, the neighborhood reference window will fall outside the image; therefore, when performing neighborhood discrimination, the input image I also needs to be expanded; the mirror padding method is used to expand I to size.

[0063] Based on the above considerations, referring to Figure 1 , the present application is implemented through the following steps:

[0064] Step 1: input the original SAR image I, and let the size of I be , filter I using anisotropic diffusion to obtain the smoothed image I AD ;

[0065] Step 2: convolve I AD using the Gaussian difference (DoG) operator to obtain I DoG ;

[0066] Step 3: filter I DoG using the Gabor filter set to obtain the edge saliency map SME and the edge direction map OME; on SME, the higher the saliency value, the more likely the corresponding position is an edge, and the specific process is as follows:

[0067] Step 3-1: set the direction are set to 0, π / 8, …, 7π / 8 in turn, and correspondingly, 8 filter templates G0, G1, …, G7 are generated;

[0068] Step 3-2: the template G i is used in turn to convolve with I DoG , and filter results in each direction I G_i are obtained respectively, as shown in the following formula,

[0069]

[0070] wherein, * represents convolution;

[0071] Step 3-3: the point with position (m, n) on the edge saliency map SME, the saliency value of which can be calculated according to the following formula, the higher the saliency value, the higher the possibility that the point is an edge;

[0072]

[0073] Step 3-4: the point with position (m, n) on the edge direction map OME, the value OME(m, n) of which is calculated according to the following formula,

[0074]

[0075] , that is, the Gabor filter bank is used to I DoG , and the direction corresponding to the filter with the maximum modulus of the filter result is OME(m, n); if there is an edge at (m, n) of the image I, OME(m, n) represents the direction of the edge at this position, otherwise, it can be considered as an invalid value;

[0076] Step 4: the non-maximum suppression NMS processing is performed on SME NMS to obtain SME

[0077] Step 5: based on SME NMS , the double-threshold processing is performed to obtain the primary edge:

[0078] Step 5-1: first, the high threshold T H and the low threshold T L are calculated;

[0079] ;

[0080] ;

[0081] Step 5-2: the region with high saliency value on SME NMS , which exceeds , is directly determined as an edge;

[0082] Step 5-3: the region with saliency value on SME NMSThe region with high saliency value and lower than is directly determined as non-edge;

[0083] Step 5-4: SME NMS The region with saliency value in the interval is also determined as edge if it is connected with the edge region in step 5-2. The edge obtained by the above processing is called "primary edge", which contains certain false alarm;

[0084] Step 6: Neighbor discrimination processing is performed on the primary edge, and the primary edge pixel set is denoted as E init = {(r1, c1),(r2, c2), …, (r p , c p )}; the specific processing process is as follows:

[0085] Step 6-1: Calculate the discrimination threshold , ;

[0086] Step 6-2: Take the i-th element in the set E init , and construct a Gabor neighborhood reference window W i in the direction of the primary edge at (r i , c i ); the specific steps are as follows:

[0087] Step 6-2-a: construct a Gabor template with as the Gabor filter direction;

[0088] Step 6-2-b: , set the elements not less than 0.5 to 1, the elements not more than -0.5 to -1, and the elements in the interval to 0 to generate the neighborhood reference window W i ; the region composed of the elements 1 in W i is the positive domain 'A+', and the region composed of the elements -1 is the negative domain 'A-';

[0089] Step 6-3: Take a region of 11x11 size from I with as the center, denoted as ;

[0090] Step 6-4: obtain the mean value of all elements corresponding to the 'A+' region on , denoted as m + ; obtain the mean value of all elements corresponding to the 'A-' region on , denoted as m - ;

[0091] Step 6-5: If or Then E init If the i-th element is a real edge, it is retained; otherwise, it is a false alarm and is filtered out.

[0092] Step 6-6: Return to step 6-2, extract the (i+1)th element, and continue until E is traversed. init By analyzing all elements in the dataset, we can identify all primary edge pixels and obtain the final edge detection result.

[0093] The edge detection effect of the present invention is further illustrated by the following simulation experiments.

[0094] The simulation scenario is described as follows: The experimental material used is simulated SAR images, as shown in the attached image. Figure 2 As shown, its size is 729. 700. The regular nested graphics in the left half of the figure can be used to test the algorithm's performance in multi-edge scenarios; the irregular graphics in the right half of the figure can be used to study the algorithm's ability to detect edges in various directions. Figure 2 The corresponding tags (GroundTruth, GT) are as follows: Figure 3 As shown in the diagram, black represents non-edge areas, white represents edges, and the gray areas symmetrically distributed on both sides of the edges are edge-tolerance pixels, with a width of 4 pixels on each side. If the extracted edge falls into either the white or gray area, the detection result is considered correct.

[0095] The experimental results are attached. Figure 4 and attached Figure 5 As shown. By Figure 4 It can be seen that the present invention produces SMEs that, while clearly protruding from the edge region, also thoroughly suppress non-edge regions. It provides a good indication of the presence of edges. On the other hand, due to... Figure 5 It can be seen that the final edge detection result obtained by this invention and Figure 3 The consistency is high, and it has achieved good performance in controlling false alarms and missed detections.

[0096] The above experiments demonstrate that this invention exhibits excellent SAR image edge detection performance. Furthermore, the edge saliency map obtained by this invention can effectively suppress other areas while highlighting edges. Therefore, this saliency map can serve as the basis for other edge detection methods, possessing good scalability.

Claims

1. A saliency-based two-stage SAR image edge detection method, characterized in that, Specifically comprising the following steps: Step 1: input the original SAR image I, let the size of I be , filter I by anisotropic diffusion to get the smoothed image I AD ; Step 2: Use the difference of Gaussians operator and I AD Convolve, get I DoG ; Step 3: Use Gabor filter banks to filter I DoG Filtering is performed to obtain the edge saliency map SME and the edge orientation map OME. On the edge saliency map SME, the higher the saliency value, the greater the probability that the corresponding position is an edge. Step 4: Non-maximum suppression (NMS) is performed on the SMEs to obtain the SMEs NMS ; Step 5: SME based NMS Double thresholding is performed to get the preliminary edges: Step 6: carrying out neighborhood discrimination processing on the primary edge to obtain a final edge detection result; Let the set of primary edge pixels be E init = {(r1, c1), (r2, c2), …, (r p , c p )}, where (r p , c p ) denotes pixel position coordinates; Step 6-1: Calculate the discrimination threshold , ; where STD( ) is the standard deviation of all elements in the matrix. mean( ) is the mean of all elements in the matrix; Step 6-2: Take out the set E init The i-th element in the middle of the first edge, the direction of the Gabor neighborhood reference window W i , c i ) at (r i , 1≤i≤p; the specific steps are as follows: Step 6-2-a: With As Gabor filter direction, construct Gabor template ; Step 6-2-b: Elements with a value not less than 0.5 are set to 1, and elements with a value not exceeding -0.5 are set to -1. Elements within the interval... Set the elements in the property to 0 to generate a neighborhood reference window W. i W i The region composed of elements 1 is the positive region 'A+', and the region composed of -1 is the negative region 'A-'. Step 6-3: Take an 11x11 region centered at I from I, denoted as ; Step 6-4: Compute the mean of the region ; Step 6-4: find the mean of all elements corresponding to the 'A+' region, denoted by m + ; find the mean of all elements corresponding to the 'A-' region, denoted by m - ; Step 6-5: If or then E init The ith element in the middle is a true edge, and is preserved; Otherwise, it is a false alarm and is filtered out; Step 6-6: Return to Step 6-2: Step, take out the i+1th element, until all elements in E are traversed init All elements in the middle, complete the identification of all primary edge pixels, get the final edge detection result.

2. The saliency-based two-stage SAR image edge detection method of claim 1, wherein, The specific method of the step 3 is: Step 3-1: setting the direction of Gabor filter are set to 0, π / 8, …, 7π / 8 in turn, and correspondingly, 8 filter templates G0, G1, …, G7 are generated; Step 3-2: Use the templates G i and I DoG Convolution, respectively, get the filter results I in each direction G_i As shown in the following formula; ； Wherein, i = 0, 1, 2,..., 7, * represents convolution; Step 3-3: the point with the position (m, n) on the edge saliency map SME, and the saliency value is calculated according to the following formula: ； wherein, a filtered result of a point representing a position range (m, n), represents a maximum value of the maximum value of Step 3-4: the point with the position (m, n) on the edge direction map OME, and the value OME(m, n) is calculated according to the following formula: ； That is, the Gabor filter bank acts on I DoG The direction corresponding to the filter with the largest modulus of the filtering result is OME(m, n); if there is an edge at (m, n) of the image I, OME(m, n) represents the direction of the edge at this position, otherwise, OME(m, n) is considered as an invalid value.

3. The saliency-based two-stage SAR image edge detection method of claim 2, wherein, The specific method of the step 5 is: Step 5-1: First, calculate the high threshold T H and the low threshold T L ; ； ； Step 5-2: SME NMS The upper saliency value exceeds The region is directly determined as the edge. Step 5-3: SME NMS The upper saliency value is lower than The region whose saliency value is lower than a threshold is directly determined as non-edge. Step 5-4: SME NMS The upper significant value is in the region of the interval If the region is connected with the edge region in step 5-2, it is also determined as an edge; the edge obtained at this time is called a primary edge, which contains false alarms, which represent false edge recognition.

Images