SRA image weak target detection optimization method based on local filtering

Abstract

The invention discloses an SRA image weak target detection optimization method based on local filtering, and belongs to the field of SRA image weak target processing. The problem that a weak target in an existing SAR image is not easy to detect is solved. The method comprises the following steps: performing local mean filtering on a to-be-detected SRA image to obtain local mean images under different scales; expanding the local mean image under each scale by three times, obtaining eight window filters at different positions under each scale, respectively filtering the to-be-detected SRA image, and respectively subtracting pixel values of corresponding pixel points in the local mean image under the corresponding scale from eight filtering result images under each scale to obtain eight pixel difference matrixes; carrying out dot multiplication on the two pixel difference value matrixes with the mark number difference 3, obtaining four correlation values for each pixel point, and taking the minimum value in the four correlation values as the pixel value of the pixel point; and taking the maximum values of the pixel values under different scales to form a two-dimensional image matrix Im, and obtaining an optimized image by using the two-dimensional image matrix Im. The method is suitable for SRA image weak target detection.

Description

Technical Field

[0001] The invention belongs to the field of SRA image weak target processing. Background Art

[0002] With the rapid development of science and technology, synthetic aperture radar (SAR) technology is widely used in security, environment, remote sensing, climate and other fields. SAR images have unique advantages and can obtain target information at night or in bad weather conditions, so they have important application value in weak target detection.

[0003] SAR image texture contours have less feature information, and the background contains a lot of noise and clutter. When detecting distant targets, the weak target obtained in the SAR image is a dim dot on the imaging picture. The target is easily submerged in the background noise and clutter, and the existing algorithms cannot directly distinguish between small hole-like objects and weak target objects. They are extremely sensitive to noise and high-frequency background textures, and are also affected by cloud clutter and other surrounding maneuvers. When the target is very close to the edge of the cloud, high-grayscale cloud pixels will occupy a small neighborhood range of the weak target. The problem becomes more serious when the contrast between the target and the background is low. Weak target detection remains a challenging problem. Summary of the invention

[0004] The present invention aims to solve the problem that weak targets in existing SAR images are difficult to detect, and provides an SRA image weak target detection optimization method based on local filtering.

[0005] The SRA image weak target detection optimization method based on local filtering of the present invention comprises:

[0006] Step 1: using mask matrices of different sizes, performing local mean filtering on the SRA image to be detected at different scales, and obtaining local mean images at different scales; the mask matrix is ​​at least a 3×3 matrix;

[0007] Step 2: enlarge the local mean image at each scale by 3 times, slide the mask matrix at different scales described in step 1 on the corresponding enlarged image, and obtain 8 window filters at different positions at each scale; the 8 window filters at different positions are 8 windows arranged around the central window position; and the 8 window filters at different positions are numbered in sequence;

[0008] Step 3: Filter the SRA images to be detected by using the 8 window filters at different positions, obtain 8 filter result images at each scale, subtract the pixel values ​​of the corresponding pixels in the local mean image at the corresponding scale from the 8 filter result images at each scale, obtain 8 pixel difference matrices, and label the difference matrices according to the labels of the 8 window filters at different positions; obtain 8 labeled pixel difference matrices;

[0009] Step 4: multiply the two pixel difference matrices whose labels differ by 3, obtain 4 correlation values ​​for each pixel, and take the minimum value of the 4 correlation values ​​as the pixel value of the pixel;

[0010] Step 5: Select the maximum pixel value of the same pixel at different scales as the pixel value of the pixel in the SRA image to be detected, and obtain the maximum correlation two-dimensional image matrix Im;

[0011] Step six: calculate the local feature response of the two-dimensional image matrix Im; and obtain the optimized image out02.

[0012] Furthermore, in the present invention, in step 2, the 8 different positions of the 8 different position window filters are 8 non-overlapping window positions excluding the center window position.

[0013] Furthermore, in the present invention, in step 6, the method for obtaining the optimized image out02 is:

[0014] Step 61: Use the average filter h to perform local mean filtering on the image matrix Im and the transposed matrix G of the image matrix Im to obtain the local mean mu and the local mean nu;

[0015] Step 62: Multiply the image matrix Im with the corresponding pixels in the transposed image matrix G to obtain a product matrix, and perform a convolution operation on the product matrix and the filter h to obtain a difference result;

[0016] Step 63: Calculate the pixel difference between the local mean of the square of the transposed image matrix G and the square of the local mean to obtain the local variance varSigma of the transposed image matrix G, use the local variance varSigma to calculate and obtain the normalization factor normalizeFactor, and then use the normalization factor normalizeFactor to calculate and obtain the first weighting coefficient A and the second weighting coefficient B;

[0017] Step 64: After adjusting the image matrix G by the first weighting coefficient A and the second weighting coefficient B, divide it by the normalization factor normalizeFactor to obtain the optimized image out02.

[0018] Furthermore, in the present invention, in step 1, the mask matrices of different sizes are matrices of 3×3, 5×5, 7×7 and 9×9 respectively.

[0019] Furthermore, in the present invention, between step five and step six, it also includes: using a [3,3] average filter to smooth the maximum correlation two-dimensional matrix Im and eliminating the noise in the two-dimensional matrix Im.

[0020] Furthermore, in the present invention, in step 63, the method for calculating and obtaining the first weighting coefficient A and the second weighting coefficient B is:

[0021] The first weight coefficient w is calculated using the local variance varSigma, and the first weight coefficient w is convolved using the filter h to obtain a normalization factor normalizeFactor;

[0022] According to the local variance varSigma, the second weight parameter Beta is calculated, the first weight coefficient w is multiplied by the second weight parameter Beta to obtain a parameter product, and the parameter product is filtered by an average filter h to obtain a first weight coefficient A;

[0023] Calculate the local mean L of the transposed image matrix G, subtract the product of the second weight parameter Beta and the local mean nu of the transposed image matrix G from the mean L, multiply the difference obtained by the first weight coefficient w to obtain a first product result matrix, filter the first product result matrix with a filter h, and obtain a second weight coefficient B.

[0024] Furthermore, in the present invention, the weight parameter Beta is:

[0025]

[0026] Φ=(Im*G)*h-mu*nu

[0027] sign is a sign function that returns positive or negative 1 according to the sign of Φ, kappa is the weight factor of the regularization term, and Epsilon varSigma represents the local variance.

[0028] Furthermore, in the present invention, the weight coefficient w is:

[0029]

[0030] Among them, V is the average value of local variance, and scale is the scaling factor of weight w.

[0031] Furthermore, in the present invention, in step nine, the method for obtaining the optimized image out02 is:

[0032]

[0033] Among them, normalizeFactor represents the normalization factor.

[0034] The present invention processes images at different scales and uses mask matrices of different sizes (3×3, 5×5, 7×7, 9×9) to perform local mean calculations. At different scales, different details of the target can be better captured, from larger structures to smaller details, making the target easier to detect. Correlation calculations are performed on local features at different scales, which can fuse information at different scales together to more comprehensively capture the features of the target; the common features of the target are extracted, the influence of the background is weakened, and the target is more prominent in the image, thereby improving the recognition of the target. The improved local feature response is calculated, the contrast of the image is adjusted, and the target in the image is clearer and more prominent. The local features in the image are highlighted using weighted coefficients, and the target in the image is effectively highlighted, making it easier to accurately detect the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 It is the distribution diagram of the mean window with different positions in 8 locations. DETAILED DESCRIPTION

[0037] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0038] Specific implementation method 1: refer to Figure 1 and Figure 2 Specifically describing this embodiment, the SRA image weak target detection optimization method based on local filtering described in this embodiment includes:

[0039] Step 1: using mask matrices of different sizes, performing local mean filtering on the SRA image to be detected at different scales, and obtaining local mean images at different scales; the mask matrix is ​​at least a 3×3 matrix;

[0040] Step 2: enlarge the local mean image at each scale by 3 times, slide the mask matrix at different scales described in step 1 on the corresponding enlarged image, and obtain 8 window filters at different positions at each scale; the 8 window filters at different positions are 8 windows arranged around the central window position; and the 8 window filters at different positions are numbered in sequence;

[0041] Step 3: Filter the SRA images to be detected by using the 8 window filters at different positions, obtain 8 filter result images at each scale, subtract the pixel values ​​of the corresponding pixels in the local mean image at the corresponding scale from the 8 filter result images at each scale, obtain 8 pixel difference matrices, and label the difference matrices according to the labels of the 8 window filters at different positions; obtain 8 labeled pixel difference matrices;

[0042] The index of the difference matrix is ​​the same as the index of the corresponding position window filter; for example, the index of the difference matrix obtained by subtracting the pixel values ​​of the corresponding pixels in the local mean image at the corresponding scale from the filtered result image obtained by using the filter with the index 1 is also 1;

[0043] Step 4: multiply the two pixel difference matrices whose labels differ by 3, obtain 4 correlation values ​​for each pixel, and take the minimum value of the 4 correlation values ​​as the pixel value of the pixel;

[0044] Step 5: Select the maximum pixel value of the same pixel at different scales as the pixel value of the pixel in the SRA image to be detected, and obtain the maximum correlation two-dimensional image matrix Im;

[0045] Step six: calculate the local feature response of the two-dimensional image matrix Im; and obtain the optimized image out02.

[0046] Furthermore, in the present invention, in step 2, the 8 different positions of the 8 different position window filters are 8 non-overlapping window positions excluding the center window position.

[0047] Furthermore, in the present invention, in step 6, the method for obtaining the optimized image out02 is:

[0048] Step 61: Use the average filter h to perform local mean filtering on the image matrix Im and the transposed matrix G of the image matrix Im to obtain the local mean mu and the local mean nu;

[0049] Step 62: Multiply the image matrix Im with the corresponding pixels in the transposed image matrix G to obtain a product matrix, and perform a convolution operation on the product matrix and the filter h to obtain a difference result;

[0050] Step 63: Calculate the pixel difference between the local mean of the square of the transposed image matrix G and the square of the local mean to obtain the local variance varSigma of the transposed image matrix G, use the local variance varSigma to calculate and obtain the normalization factor normalizeFactor, and then use the normalization factor normalizeFactor to calculate and obtain the first weighting coefficient A and the second weighting coefficient B;

[0051] Step 64: After adjusting the image matrix G by the first weighting coefficient A and the second weighting coefficient B, divide it by the normalization factor normalizeFactor to obtain the optimized image out02.

[0052] Furthermore, in the present invention, in step 1, the mask matrices of different sizes are matrices of 3×3, 5×5, 7×7 and 9×9 respectively.

[0053] Furthermore, in the present invention, between step five and step six, it also includes: using a [3,3] average filter to smooth the maximum correlation two-dimensional matrix Im and eliminating the noise in the two-dimensional matrix Im.

[0054] Furthermore, in the present invention, in step 63, the method for calculating and obtaining the first weighting coefficient A and the second weighting coefficient B is:

[0055] The first weight coefficient w is calculated using the local variance varSigma, and the first weight coefficient w is convolved using the filter h to obtain a normalization factor normalizeFactor;

[0056] According to the local variance varSigma, the second weight parameter Beta is calculated, the first weight coefficient w is multiplied by the second weight parameter Beta to obtain a parameter product, and the parameter product is filtered by an average filter h to obtain a first weight coefficient A;

[0057] Calculate the local mean L of the transposed image matrix G, subtract the product of the second weight parameter Beta and the local mean nu of the transposed image matrix G from the mean L, multiply the difference obtained by the first weight coefficient w to obtain a first product result matrix, filter the first product result matrix with a filter h, and obtain a second weight coefficient B.

[0058] Furthermore, in the present invention, the weight parameter Beta is:

[0059]

[0060] Φ=(Im*G)*h-mu*nu

[0061] sign is a sign function that returns positive or negative 1 according to the sign of Φ, kappa is the weight factor of the regularization term, and Epsilon varSigma represents the local variance.

[0062] Furthermore, in the present invention, the weight coefficient w is:

[0063]

[0064] Among them, V is the average value of local variance, and scale is the scaling factor of weight w.

[0065] Furthermore, in the present invention, in step nine, the method for obtaining the optimized image out02 is:

[0066]

[0067] Among them, normalizeFactor represents the normalization factor. Specific embodiment:

[0069] Local mean filtering: Use mask matrices mask3, mask5, mask7, and mask9 of sizes 3×3, 5×5, 7×7, and 9×9 respectively as the windows used for local mean calculation. The element values ​​in these matrices are the same, and the sum is 1. Sliding calculations are performed on the image to achieve mean calculation of the local area. The local mean window passed in each time is different, and the local mean images l3, l5, l7, and l9 calculated using local mean windows of different sizes are obtained.

[0070] Calculate the correlation between local features at different scales: The following is the processing process using a 3×3 mask matrix as an example:

[0071] First, get 8 9×9 matrices m31-m38 with all 0s, respectively. Figure 2 A 3×3 matrix of all 1s is placed at positions F31-F38 in the matrix. These window positions are at different positions in the matrix, and local mean windows with different shapes are formed by setting 1 at the window positions. Then, the image is convolved with the local mean windows at these eight different positions to obtain a LCM3 three-dimensional array, where each layer represents the local mean image of the image at different positions.

[0072] Then calculate the correlation between local features of different scales and positions in the image. Specifically, first subtract each layer of LCM3 from l3 to obtain the difference of local features at a series of positions, which are stored in F31 to F38 respectively. Then use temp3 to store the products of F31 and F35, F32 and F36, F33 and F37, F34 and F38, and calculate the product of these differences to obtain the correlation between local features at different positions in the image. For example, the first layer of temp3 stores the product of F31 and F35, which represents the correlation between the local features of the image after the mean filtering of the first window position and the local features after the mean filtering of the fifth window position. Extract the minimum value along the third dimension (ie, depth) in temp3 to obtain a two-dimensional matrix out3.

[0073] The processing process of the other 5×5, 7×7, and 9×9 matrices is similar to obtain three two-dimensional matrices out5, out7, and out9.

[0074] Set a three-dimensional array temp of size (row, col, 4) to store the minimum correlation between local features at four different scales, where row and col are the number of rows and columns of the image.

[0075] Then the calculated out3, out5, out7 and out9 are stored in different layers of temp, and the corresponding pixel values ​​of the four feature images are compared along the third dimension (depth) of the temp array, and the maximum correlation value of each pixel at four scales is taken out to obtain a new two-dimensional matrix out01. Use an average filter of size [3,3] to smooth out01, eliminate the noise in the image, obtain the smoothed image Im, and extract the features in the image. Calculate the improved local feature response:

[0076] Create the averaging filter h:

[0077] Define an average filter h of size patchSize×patchSize, where the value of each element is helement to ensure that the sum of all its elements is 1, which is used to calculate the local mean.

[0078]

[0079] Assign Im to G. And use the average filter h to calculate the local means mu and nu of images Im and G.

[0080] Calculate the difference value Φ:

[0081] 1) Multiply the corresponding positions of the images Im and G to obtain the product of the pixels at the corresponding positions of the two images.

[0082] 2) Perform a convolution operation on the above product image and the filter h, and perform local averaging on the product image.

[0083] 3) Multiply the corresponding positions of images mu and nu, where mu is the local mean of image Im and nu is the local mean of image G.

[0084] 4) Subtract the two convolution results to obtain Φ, which represents the difference between the convolution result of image Im and G minus the local mean.

[0085] Calculate the local variation factor varSigma:

[0086] 1) Square the pixel values ​​at the corresponding positions of image G.

[0087] 2) Convolve the squared image with the filter h to achieve local averaging.

[0088] 3) Square the pixel value at the corresponding position of image nu.

[0089] 4) Subtract the two convolution results, compare the result with 0, output the result if it is greater than 0, and output 0 if it is less than 0, ensuring that the result is non-negative. varSigma is used to represent the difference between the local mean of the square of image G and the square of the local mean.

[0090] Calculate the weight parameter Beta:

[0091] 1) First, calculate the intermediate variable a according to formula (2):

[0092]

[0093] 2) Epsilon is a small positive number used to avoid division by zero.

[0094] 3) Use formula (3) to calculate:

[0095]

[0096] Here, sign is the sign function, which returns positive or negative 1 according to the sign of Φ.

[0097] Calculate the weight coefficient w:

[0098] 1) Take the average of all elements in the entire varSigma matrix, that is, calculate the average value V of the local variance.

[0099] 2) Use scale as the scaling factor to adjust the range of weights.

[0100] 3) Calculate using formula (4). Where w is a matrix representing the weight of the pixel. Divide the local variance by its mean value and multiply it by the scaling factor to obtain the weight coefficient w.

[0101]

[0102] Then the weight is calculated and the original variance value is weighted by formula (5) so that the larger variance corresponds to a smaller weight and the smaller variance corresponds to a larger weight.

[0103] Calculate the normalization factor normalizeFactor: Use the filter h to perform a convolution operation on the weight coefficient w to obtain the normalization factor to ensure that the weighted image value is correctly normalized.

[0104] Calculate the weighting factor A:

[0105] Beta and w are multiplied at corresponding positions to obtain a new weighting coefficient. The weighting coefficient is filtered using the normalized filter h to obtain A.

[0106] Calculate the weighting factor B:

[0107] 1) Calculate the local mean mu of Im minus the product of Beta and the local mean nu of G to represent the difference between the average value L of image G and the average value of G after weight adjustment.

[0108] 2) Then multiply L by the weight w.

[0109] 3) Use the normalized filter h to filter the weighted difference to obtain B.

[0110] Calculate the output image out02:

[0111]

[0112] After multiplying image G by A and adding B, we get the weighted result of image G after contrast adjustment and the average difference. The weighted result is divided by the normalization factor normalizeFactor to get the final output image out02, which is then filtered by the threshold to obtain a binary image.

[0113] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. The optimization method of weak target detection in SRA images based on local filtering is characterized by: include: Step 1: using mask matrices of different sizes, performing local mean filtering on the SRA image to be detected at different scales, and obtaining local mean images at different scales; the mask matrix is ​​at least a 3×3 matrix; Step 2: enlarge the local mean image at each scale by 3 times, slide the mask matrix at different scales in step 1 on the corresponding enlarged image, and obtain 8 window filters at different positions at each scale; and number the 8 window filters at different positions in sequence; Step 3: Filter the SRA images to be detected by using the 8 window filters at different positions, obtain 8 filter result images at each scale, subtract the pixel values ​​of the corresponding pixels in the local mean image at the corresponding scale from the 8 filter result images at each scale, obtain 8 pixel difference matrices, and label the difference matrices according to the labels of the 8 window filters at different positions; obtain 8 labeled pixel difference matrices; Step 4: multiply the two pixel difference matrices whose labels differ by 3, obtain 4 correlation values ​​for each pixel, and take the minimum value of the 4 correlation values ​​as the pixel value of the pixel; Step 5: Select the maximum pixel value of the same pixel at different scales as the pixel value of the pixel in the SRA image to be detected, and obtain the maximum correlation two-dimensional image matrix Im; Step six: calculate the local feature response of the two-dimensional image matrix Im; and obtain the optimized image out02.

2. The SRA image weak target detection optimization method based on local filtering according to claim 1 is characterized in that: In step 2, the 8 different positions of the 8 different position window filters are 8 non-overlapping window positions excluding the center window position.

3. The SRA image weak target detection optimization method based on local filtering according to claim 2 is characterized in that: In step 6, the method for obtaining the optimized image out02 is: Step 61: Use the average filter h to perform local mean filtering on the image matrix Im and the transposed matrix G of the image matrix Im to obtain the local mean mu and the local mean nu; Step 62: Multiply the image matrix Im with the corresponding pixels in the transposed image matrix G to obtain a product matrix, and perform a convolution operation on the product matrix and the filter h to obtain a difference result; Step 63: Calculate the pixel difference between the local mean of the square of the transposed image matrix G and the square of the local mean to obtain the local variance varSigma of the transposed image matrix G, use the local variance varSigma to calculate and obtain the normalization factor normalizeFactor, and then use the normalization factor normalizeFactor to calculate and obtain the first weighting coefficient A and the second weighting coefficient B; Step 64: After adjusting the image matrix G by the first weighting coefficient A and the second weighting coefficient B, divide it by the normalization factor normalizeFactor to obtain the optimized image out02.

4. The SRA image weak target detection optimization method based on local filtering according to claim 3 is characterized in that: In step 1, the mask matrices of different sizes are 3×3, 5×5, 7×7 and 9×9 matrices respectively.

5. The SRA image weak target detection optimization method based on local filtering according to claim 1, 2, 3 or 4, characterized in that: Between step five and step six, the method further includes: using a [3,3] average filter to smooth the maximum correlation two-dimensional matrix Im and eliminating the noise in the two-dimensional matrix Im.

6. The SRA image weak target detection optimization method based on local filtering according to claim 3 is characterized in that: In step 63, the method for calculating and obtaining the first weighting coefficient A and the second weighting coefficient B is: The first weight coefficient w is calculated using the local variance varSigma, and the first weight coefficient w is convolved using the filter h to obtain a normalization factor normalizeFactor; According to the local variance varSigma, the second weight parameter Beta is calculated, the first weight coefficient w is multiplied by the second weight parameter Beta to obtain a parameter product, and the parameter product is filtered by an average filter h to obtain a first weight coefficient A; Calculate the local mean L of the transposed image matrix G, subtract the product of the second weight parameter Beta and the local mean nu of the transposed image matrix G from the mean L, multiply the difference obtained by the first weight coefficient w to obtain a first product result matrix, filter the first product result matrix with a filter h, and obtain a second weight coefficient B.

7. The SRA image weak target detection optimization method based on local filtering according to claim 6 is characterized in that: Weight parameter Beta: Φ=(Im*G)*h-mu*nu sign is a sign function that returns positive or negative 1 according to the sign of Φ, kappa is the weight factor of the regularization term, and EpsilonvarSigma represents the local variance.

8. The SRA image weak target detection optimization method based on local filtering according to claim 7 is characterized in that: The weight coefficient w is: Among them, V is the average value of local variance, and scale is the scaling factor of weight w.

9. The SRA image weak target detection optimization method based on local filtering according to claim 8 is characterized in that: In step nine, the method for obtaining the optimized image out02 is: Among them, normalizeFactor represents the normalization factor.

Images