A high dynamic SAR image corner anti-discriminating method

Abstract

The application discloses a high-dynamic SAR image corner anti-discrimination method, which utilizes a multi-core DSP processor to perform parallel operation, converts high-dynamic M-bit image data to be processed into M / 2-bit image data through power transformation, performs a maximum value searching operation on the M / 2-bit image data after the power transformation, performs target pixel position merging on the pixel maximum value, performs ROI slice segmentation on the merged pixel maximum value, performs ROI processing on the segmented ROI slice, generates an ROI binary graph, calculates a connected region area value in the ROI binary graph, and distinguishes targets and corner anti-discriminations. The application can quickly screen out high-value targets and corner anti-discriminations in a high-dynamic scene, has strong robustness, wide applicability, high automation degree and rapid processing.

Description

Technical Field

[0001] This invention relates to the field of radar target detection and recognition, and in particular to a high dynamic SAR image angle inversion discrimination method based on DSP parallel processing. Background Technology

[0002] SAR terminal guidance technology is highly favored due to its excellent characteristics of all-weather, all-time, and high resolution, and will play an indispensable role in the field of terminal guidance. Most existing SAR image recognition technologies are based on algorithms such as deep learning, neural networks, and sparse representation. These algorithms have high complexity, require a large number of training samples, and are not suitable for working platforms with high real-time requirements. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying angular inversion in high-dynamic SAR images, which can quickly screen out high-value targets and angular inversions in high-dynamic scenes. It is robust, widely applicable, highly automated, and fast in processing.

[0004] To achieve the above objectives, the present invention provides a high-dynamic SAR image corner inversion identification method, comprising the following steps:

[0005] Step S1: Utilize all cores in the multi-core DSP processor to perform synchronous parallel operations, converting the high dynamic range M-bit image data to be processed into M / 2-bit image data through power transformation;

[0006] Step S2: Use all cores of the multi-core DSP processor to perform synchronous parallel operations to find the maximum value of the M / 2 bit image data after power transformation;

[0007] Step S3: Use the main control core in the multi-core DSP processor to merge the target pixel positions based on the maximum pixel value;

[0008] Step S4: Utilize all cores of the multi-core DSP processor to perform synchronous parallel operations and divide the merged pixel maximum value into ROI slices;

[0009] Step S5: Use all cores of the multi-core DSP processor to perform synchronous parallel operation, perform ROI processing on the segmented ROI slices, and generate ROI binary map.

[0010] Step S6: Use all cores of the multi-core DSP processor to perform synchronous parallel operations to calculate the area value of the connected regions in the binary ROI graph, in order to distinguish between the target and the corner reflection.

[0011] Step S1 includes the following steps:

[0012] Step S1.1: Divide the original N×N image into WN×WN sub-images of size W×W pixels, and number the sub-images sequentially as R1, R2, ..., Rt, where t = WN×WN;

[0013] Step S1.2: Distribute the WN×WN subgraphs evenly to all cores in the multi-core DSP processor;

[0014] Step S1.3: Let f represent the input SAR image, with an image size of N×N. After power transformation, the image expression is as follows:

[0015]

[0016] Where c is a constant greater than 0, and γ is an adjustable parameter, taking the value 0 < γ < 1. Power transformation is a non-linear transformation; when processing M-bit images, taking γ = 0.5 ensures that the maximum pixel value in the M-bit image is 2. M / 2 -1.

[0017] Step S2 includes the following steps:

[0018] Step S2.1: All cores in the multi-core DSP processor perform parallel computation on the N×N original image to calculate the value of the maximum pixel in each sub-image Max(i) and its position in the original image (PosX(i), PosY(i)), i = 1, 2, ..., WN×WN;

[0019] Step S2.2: Sort the WN×WN maximum values ​​and take the top DN maximum pixels and their positions in the original image. These positions are the target potential pixel maximum value points.

[0020] In step S3, the main control core of the multi-core DSP processor merges the largest pixels in the first DN sub-images according to distance. The merging distance threshold is the target length / image resolution. The distance between the largest pixels i and j (i = 1, 2, 3, ..., DN-1, j = i+1) is:

[0021]

[0022] Determine the distance d[i,j] between the largest pixels i and j. If it is within the distance threshold, merge these two largest pixels, assign the center coordinate value between the two largest pixels to the largest pixel i, and clear the coordinates of the largest pixel j to 0. The calculation formula is as follows: The number of iterations in the merging calculation process is

[0023] The merged pixel positions are used as the center points of the Region of Interest (ROI) to be generated.

[0024] In step S4, the main control core in the multi-core DSP processor distributes the selected n (n≤DN) maximum value pixels equally to all cores in the multi-core DSP processor, and each core takes an R×R ROI slice centered on the position of the maximum value pixel.

[0025] Step S5 includes the following steps:

[0026] Step S5.1: Perform mean filtering on the ROI slices;

[0027] Select the filter window size W based on the size of the scanned image slice. f ×W f Calculate the half-window length based on the filter window size. From the ROI slice (1+W) half ) line, (1+W) half Starting from the data points in column (mW), the data points are used to calculate the range from the data points in column (mW) to the data points in column (mW). half ) line, (nW) half ) column of data points As the endpoint, ensure that all data points within the selected filtering window are within the ROI slice, and calculate the mean M of the sum of all data points within the filtering window centered on the current data point. i,j :

[0028]

[0029] Calculate the filtered mean M (m+1),(n+1) Assign the value to the center point within the current window, and so on, traversing all points in the ROI that meet the conditions. Data points within the ROI slice that are not traversed use the original image pixel values.

[0030] Step S5.2: Adaptive threshold calculation;

[0031] First, find the maximum pixel value within the ROI slice after mean filtering, and then calculate the segmentation threshold T0 for the corresponding ROI slice based on the maximum pixel value:

[0032]

[0033] In the formula, P max TH represents the maximum amplitude of pixels within the ROI slice after mean filtering, and TH is the main lobe and side lobe energy difference adaptively adjusted according to the detection power.

[0034] Step S5.3: Perform adaptive threshold segmentation on the filtered ROI slices;

[0035] The grayscale image is segmented according to the selected segmentation threshold T0, and pixels with grayscale values ​​greater than the threshold T0 are assigned a value of 2. M / 2-1 is used if the value is otherwise 0, resulting in a binary image, which is then assigned a value of 2. M / 2 The pixels with a value of -1 represent the pixel region corresponding to the potential target, which serves as the basis for subsequent target area feature extraction.

[0036] In step S6, the area value of the connected region near the center point of the ROI is selected, and an area feature threshold value is set. Areas greater than the threshold value are considered as suspected targets, and areas less than the threshold value are considered as angular reflections.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. Compared with existing target recognition methods based on deep learning, neural networks, etc., this invention is more applicable to DSP engineering application verification platforms. It can solve the problem of quickly detecting and distinguishing high-value targets and angular reflections in high-dynamic scenes on high-speed platforms without the need for a large number of training samples, and has good engineering application efficiency.

[0039] 2. Compared with the conventional KSW entropy threshold segmentation to obtain the ROI binary image and then use it to obtain the target feature vector to filter out inverted corners, this invention can improve the accuracy of inverted corner removal. Attached Figure Description

[0040] Figure 1 This is a flowchart of a high-dynamic SAR image angle inversion identification method provided by the present invention.

[0041] Figure 2 The simulation experiment of this invention processes images. Detailed Implementation

[0042] The following is based on Figure 1 and Figure 2 The preferred embodiments of the present invention will be described in detail below.

[0043] like Figure 1 As shown, this invention provides a high-dynamic SAR image corner inversion identification method based on DSP parallel processing. The DSP parallel processing implemented in this algorithm is to ensure the timeliness and accuracy of the algorithm's processing. It requires DSP core 0 as the main control core, sending control processing data instructions to DSP cores 0 through 7, and receiving and summarizing the data information processed by DSP cores 0 through 7. The specific operation of this method includes the following steps:

[0044] Step S1: To improve recognition accuracy, the high dynamic M-bit image data to be processed needs to be converted into M / 2-bit image data through power transformation. Due to the large amount of data and the long time required for the DSP to perform power function operation on the high dynamic M-bit image data, the DSP 0 core divides the M-bit image to be processed into 8 equal parts and distributes them to the DSP 0 core to 7 cores to complete the power transformation processing of the SAR original image synchronously by the DSP 8 cores within the time limit.

[0045] In step S1, the original SAR image with an input size of N×N and a pixel bit depth of M bits is processed by a power transform of the DSP 8 cores to convert the M bits into M / 2 bits, which can reduce the burden of subsequent processing.

[0046] The specific implementation is as follows:

[0047] Step S1.1: Divide the original N×N image into WN×WN sub-images of size W×W pixels, and number the sub-images sequentially as R1, R2, ..., Rt, where t = WN×WN.

[0048] Step S1.2: Assign WN×WN subgraphs to the 8 cores of the DSP (subgraphs R1, R2, R3, R4, R5, R6, R7, R8, R9, R1, R1, R2 ...3, R4, R1, R2, R3, R4, R3, R4, R3, R4, R3, R4, R3, R4, R3, R4, R3, R4, R3, R4, R3, R4, R3, 8+1 R 2*8+1 ...Assigned to kernel 0; subgraphs R2, R 8+2 R 2*8+2 ...assigned to core 1, and so on).

[0049] Step S1.3: Let f represent the input SAR image, with an image size of N×N. After power transformation, the image expression is as follows:

[0050]

[0051] Where c is a constant greater than 0, and γ is an adjustable parameter, taking the value 0 < γ < 1. Power transformation is a non-linear transformation; when processing M-bit images, γ = 0.5 is generally chosen to ensure that the maximum pixel value in the M-bit image is 2. M / 2 -1.

[0052] Step S2: After the 8 cores of the DSP synchronously complete the power transformation process, M / 2 bit image data is obtained. Then, the 8 cores of the DSP are used to perform SAR image segmentation and find the maximum value.

[0053] The specific steps are as follows:

[0054] Step S2.1: The DSP 8-core parallel processing of the N×N original image calculates the value of the maximum pixel in each sub-image Max(i) and its position in the original image (PosX(i), PosY(i)), i = 1, 2, ..., WN×WN.

[0055] Step S2.2: Sort the WN×WN maximum values ​​and take the first DN (depending on the scene settings and the size range of the original image, the value of D is generally selected from 8 to 24, and this invention selects 16) maximum pixels and their positions in the original image. These positions are the target potential pixel maximum value points.

[0056] Step S3: The DSP 0 core summarizes the target pixel positions and merges the position information. The merged pixel positions are used as the center point positions of the Region of Interest (ROI) to be generated.

[0057] In step S3, the pixel positions are merged based on the maximum pixel value extracted in step S2 above, removing the influence of the detection target being divided into multiple sub-images, so that the final maximum pixel value comes from different detection targets.

[0058] The largest pixels in the first DN sub-images are merged based on distance, with the merging distance threshold being the target length / image resolution. The specific calculation is shown below. The distance between the largest pixels i and j (i = 1, 2, 3, ..., DN-1, j = i+1) is:

[0059]

[0060] Determine the distance d[i,j] between the largest pixels i and j. If it is within the distance threshold, merge these two largest pixels, assign the center coordinate value between the two largest pixels to the largest pixel i, and clear the coordinates of the largest pixel j to 0. The calculation formula is as follows: The number of iterations in the merging calculation process is

[0061] Step S4: DSP core 0 evenly distributes the ROI pixel positions to DSP cores 0 to 7, and all 8 DSP cores simultaneously generate ROI slices.

[0062] In step S4, the ROI is sliced ​​and segmented using DSP 8 cores based on the coordinates of the maximum pixel value obtained in step S3 above.

[0063] The specific steps are as follows:

[0064] The n (n≤DN) maximum value pixels selected in step S3 are evenly distributed to the 8 cores of DSP 0 for synchronous processing. Each core takes an R×R ROI slice centered on the position of the maximum value pixel (ROI slices that exceed the edge of the original image are not taken, and the maximum loss is the area size of the region ROISIZE / 2 pixels away from the edge).

[0065] Step S5: After the 8 cores of the DSP generate ROI slices, mean filtering is performed. Considering the difference between the main lobe and side lobe energy of the angular reflection and the target, the maximum pixel amplitude of each filtered ROI slice is selected, and an adaptive threshold is calculated based on the maximum amplitude. The ROI binary map is generated based on the threshold value.

[0066] In step S5, the segmented ROI slices obtained from step S4 are processed by an 8-core DSP, including mean filtering and binarization.

[0067] The specific steps are as follows:

[0068] Step S5.1: Perform mean filtering on the ROI slices.

[0069] Select the filter window size W based on the size of the scanned image slice. f ×W f Calculate the half-window length based on the filter window size. From the ROI slice (1+W) half ) line, (1+W) half Starting from the data points in column (mW), the data points are used to calculate the range from the data points in column (mW) to the data points in column (mW). half ) line, (nW) half ) column of data points As the endpoint (ensuring all data points within the selected filter window are within the ROI slice), calculate the mean M of the sum of all data points within the filter window centered on the current data point. i,j The calculation formula is as follows:

[0070]

[0071] Calculate the filtered mean M (m+1),(n+1) The value is assigned to the center point within the current window, and so on, iterating through all points in the ROI that meet the criteria. Data points within the ROI that are not traversed use the original image pixel values.

[0072] Step S5.2, Adaptive threshold calculation.

[0073] First, find the maximum pixel value within the ROI slice after mean filtering. Then, calculate the segmentation threshold T0 for the corresponding ROI slice based on the maximum pixel value. The formula for calculating threshold T0 is as follows:

[0074]

[0075] In the formula, P max TH represents the maximum amplitude of pixels within the ROI slice after mean filtering, and TH is the energy difference between the main lobe and the side lobe that is adaptively adjusted based on the detection power.

[0076] Step S5.3: Perform adaptive threshold segmentation on the filtered ROI slices.

[0077] The grayscale image is segmented based on the selected segmentation threshold T0, and pixels with grayscale values ​​greater than the threshold T0 are assigned a value of 2. M / 2 -1 is used if the value is otherwise 0, resulting in a binary image, which is then assigned a value of 2.M / 2 The pixels with a value of -1 represent the pixel region corresponding to the potential target, which serves as the basis for subsequent target area feature extraction.

[0078] In step S6, after the 8 cores of the DSP generate the binary map of the ROI, the area of ​​the connected region in the binary map of the ROI is calculated, and the area threshold that distinguishes the target from the corner reflection is set as the basis for the distinction between the corner reflection and the target.

[0079] In step S6, based on the binary map of the ROI slice obtained in step S5, the area value of the connected region near the center point of the ROI is selected. According to the set area feature threshold value, the area greater than the threshold value is considered as a suspected target, and the area less than the threshold value is considered as a corner reflection.

[0080] Simulation Experiment

[0081] The simulated SAR image has a pixel resolution of 1.5 meters. The scene consists of angular reflections and target vehicles. Images containing high-value targets are preprocessed using a power transform algorithm to convert 16-bit images to 8-bit images. The power transform parameter γ is set to 0.5. The maximum value in the image is found and its position is merged to obtain the ROI image. The ROI image is then subjected to mean filtering and adaptive threshold segmentation to generate a binary image. The area of ​​the connected region of the detected target within the ROI slice is calculated, and angular reflections are removed by setting a threshold.

[0082] Specific simulation imaging images are as follows: Figure 2 As shown in Table 1, the detection area feature data and identification table are presented.

[0083] Table 1 shows the area feature values ​​of the target and the corner reflection obtained after processing by the adaptive threshold segmentation method and the KSW entropy threshold segmentation method, respectively.

[0084] in Figure 2 The bright spots in the original 16-bit SAR image are the target vehicle and the corner reflection (the target vehicle is at the top, and the corner reflection is at the bottom); Figure 2 The inverted ROI slice and the target ROI slice are 48×48 ROI slices extracted from the original 16-bit SAR image after power transformation. Figure 2 The inverted corner and target ROI slice binary images obtained by adaptive threshold segmentation are obtained by the adaptive threshold segmentation method of this invention. Figure 2 The inverted corner and target ROI slice binary images obtained by KSW entropy threshold segmentation are obtained by conventional KSW entropy threshold segmentation.

[0085] Table 1

[0086]

[0087] The beneficial effects of this invention are as follows:

[0088] 1. Compared with existing target recognition methods based on deep learning, neural networks, etc., this invention is more applicable to DSP engineering application verification platforms. It can solve the problem of quickly detecting and distinguishing high-value targets and angular reflections in high-dynamic scenes on high-speed platforms without the need for a large number of training samples, and has good engineering application efficiency.

[0089] 2. Compared with the conventional KSW entropy threshold segmentation to obtain the ROI binary image and then use it to obtain the target feature vector to filter out inverted corners, this invention can improve the accuracy of inverted corner removal.

[0090] It should be noted that, in the embodiments of the present invention, the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the embodiments and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0091] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A high dynamic SAR image corner anti-aliasing method, characterized in that, The method comprises the following steps: Step S1, synchronous parallel operation is performed by using all the cores in the multi-core DSP processor, and high dynamic M-bit image data to be processed is converted into M / 2-bit image data through power transformation; Step S2, synchronous parallel operation is performed by using all the cores in the multi-core DSP processor, and the M / 2-bit image data after power transformation is subjected to a maximum value finding operation; Step S3, the master core in the multi-core DSP processor performs target pixel position merging on the maximum pixel value; Step S4, synchronous parallel operation is performed by using all the cores in the multi-core DSP processor, and the merged maximum pixel value is subjected to ROI slice segmentation; the master core in the multi-core DSP processor evenly distributes the selected n maximum value pixel points to all the cores in the multi-core DSP processor, each core takes an R×R ROI slice as the center of the maximum value pixel point, and n≤DN; Step S5, synchronous parallel operation is performed by using all the cores in the multi-core DSP processor, and the segmented ROI slice is subjected to ROI processing to generate an ROI binary image; Step S6, synchronous parallel operation is performed by using all the cores in the multi-core DSP processor, and the area value of a connected region in the ROI binary image is calculated to distinguish the target and the corner.

2. The high dynamic SAR image corner detection method of claim 1, wherein, The step S1 comprises the following steps: Step S1.1, an N×N original image is segmented into WN×WN subgraphs with a size of W×W pixels, and the subgraphs are sequentially numbered as R1, R2,..., Rt, t=WN×WN; Step S1.2, the WN×WN subgraphs are evenly distributed to all the cores in the multi-core DSP processor; Step S1.3, f represents an input SAR image, the image size is N×N, and the image expression after power transformation is as follows: wherein c is a constant greater than 0, γ is an adjustable parameter, 0 < γ < 1, and the power transformation is a nonlinear transformation. When processing M-bit images, γ = 0.5 is taken, which can ensure that the maximum pixel value in the M-bit image is 2 M / 2 -1.

3. The high dynamic SAR image corner detection method of claim 2, wherein, The step S2 comprises the following steps: Step S2.1, all the cores in the multi-core DSP processor perform parallel operation on the N×N original image, and the value Max(i) of the maximum pixel point in each subgraph and the position (PosX(i), PosY(i)) of the maximum pixel point in the original image are calculated, i=1, 2,..., WN×WN; Step S2.2, the WN×WN maximum values are sorted, and the first DN maximum pixel points and the positions of the maximum pixel points in the original image are taken, and the positions are target potential maximum value points.

4. The high dynamic SAR image corner detection method of claim 3, wherein, In the step S3, the master core in the multi-core DSP processor merges the maximum pixel points in the first DN subgraphs according to the distance, the merging distance threshold is the target length / image resolution, and the distance between the maximum pixel points i and j is as follows: The distance d[i,j] between the maximum pixel points i and j is judged, if the distance is within the distance threshold, the two maximum pixel points are merged, the center coordinate value between the two maximum pixel points is assigned to the maximum pixel point i, the coordinate of the maximum pixel point j is cleared to 0, and the calculation formula is as follows: The number of iterations of the merge calculation process is The position of the merged pixel point is taken as the position of the center point of the region of interest ROI to be generated.

5. The high dynamic SAR image corner detection method of claim 4, wherein, The step S5 comprises the following steps: Step S5.1, mean filtering is performed on the ROI slice; According to the size of the scanning image slice, select the filter window size W f ×W f , according to the filter window size, calculate the half window length From the data point of the (1+W half ) row, (1+W half ) column of the ROI slice as the starting point, to the (m-W half ) row, (n-W half ) column data point as the end point, ensure that all data points in the filter window are in the ROI slice, calculate the mean M i,j of the cumulative sum of all data points in the filter window centered on the current data point: The filtered mean M is calculated (m+1),(n+1) The value of the center point in the current window is assigned, and all the points in the ROI that meet the conditions are sequentially traversed in this way, and the data points in the ROI slice that are not traversed use the original image pixel value; Step S5.2, adaptive threshold value threshold calculation is performed. First, find the maximum pixel value in the ROI slice after mean filtering, and calculate the segmentation threshold T0 of the corresponding ROI slice according to the maximum pixel value: In the formula, P max is the maximum value of the pixel amplitude in the ROI slice after mean filtering, and TH is the difference between the main lobe and the side lobe energy that is adaptively adjusted according to the detection power. Step S5.3, performing adaptive threshold segmentation on the filtered ROI slice; According to the selected segmentation threshold threshold T0, the gray image is segmented, and the pixel points with a gray value greater than the threshold T0 are valued as 2 M / 2 -1, otherwise 0, finally obtaining a binary image, valued as 2 M / 2 The pixel with a value of -1 is a pixel region corresponding to a potential target, which serves as a basis for subsequent target area feature extraction.

6. The high dynamic SAR image corner detection method of claim 5, wherein, In the step S6, select the connected region area value close to the ROI center point, set the area feature threshold value, and the region greater than the threshold value is considered as a suspected target, and the region less than the threshold value is considered as an angle reflection.

Images