Strip line detection phase-unraveling method and system based on FLNet convolutional neural network
By detecting stripe lines using the FLNet convolutional neural network and combining it with ring integrals and undirected graph repair, error points are identified and processed, solving the phase unwrapping errors caused by noise in traditional methods and achieving high-precision phase unwrapping.
Patent Information
- Application Number
- CN202310852055.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-12
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2043-07-12
AI Technical Summary
Traditional stripe detection algorithms are sensitive to noise, which can easily lead to errors in phase unwrapping results.
We employ FLNet-based convolutional neural network for stripe detection, combine ring integrals and undirected graphs to repair broken stripe lines, and identify and process error points caused by noise.
This effectively eliminates the influence of noise on phase unwrapping, improving the accuracy and consistency of phase unwrapping.
Smart Images
Figure CN117761691B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar imaging technology, and in particular to a method and system for phase unwrapping of stripe line detection based on FLNet convolutional neural network. Background Technology
[0002] Phase unwrapping is crucial for Interferometric Synthetic Aperture Radar (InSAR) data processing. Stripe detection algorithms achieve phase unwrapping by detecting stripes formed at adjacent pixels where phase transitions occur within the wrapped phase. Specifically, based on the number of stripes along the integration path between the reference pixel and the current pixel, the product of this number and 2π is added to the phase difference of the current pixel, thus unwrapping all pixels. Traditional stripe unwrapping-based algorithms are sensitive to noise or other interference factors. Noise can destroy stripes or create new ones, potentially leading to incorrect phase unwrapping results. Summary of the Invention
[0003] In view of this, the purpose of this invention is to provide a stripe line detection phase unwrapping method and system based on FLNet convolutional neural network. This method is based on two-dimensional phase unwrapping of stripe line detection and belongs to the data processing method of interferometric synthetic aperture radar in radar imaging technology.
[0004] To achieve the above objectives, the present invention provides the following technical solution:
[0005] The present invention provides a stripe line detection phase unwrapping method based on FLNet convolutional neural network, which includes the following steps:
[0006] Constructing an FLNet convolutional neural network;
[0007] Stripe detection is performed on the input image using an FLNet convolutional neural network.
[0008] Stripes obtained by repairing detection through ring integral;
[0009] Untangling via path integral;
[0010] Error points in stripe lines are identified using the FLNet convolutional neural network.
[0011] Handle error points.
[0012] Furthermore, the detected stripe lines are repaired according to the following steps:
[0013] Use the four corner points of the pixel as graph nodes;
[0014] Connect all nodes on the stripe line except for the points on the image boundary, and determine the node that is connected to only one node as the breakpoint.
[0015] Connect two breakpoints using the nearest neighbor principle, and then assign values to pixels near the connection line using the principle that the closed-loop integral is 0.
[0016] If the breakpoint is far away, connect the shortest path based on the distance between the nodes.
[0017] Furthermore, the path integral unwrapping is performed in the following manner:
[0018] By using the unwrapping results of the horizontal and then vertical paths, the upper left corner is selected as the reference point. When passing through the stripe line, 2π is added or subtracted according to the direction of the stripe line to achieve unwrapping.
[0019] Furthermore, the FLNet convolutional neural network includes an input convolution module, several HDCRES modules, and an output convolution module; the HDCRES module includes several dilated convolution modules connected in sequence; the input convolution module performs convolution operations on the input image, then processes it through several HDCRES modules, and finally completes the classification through the convolution operation of the output convolution module.
[0020] Furthermore, the HDCRES module comprises three dilated convolution operations, and the dilation rate of the dilated convolution module is set by using a zigzag dilation rate.
[0021] Furthermore, the input convolution module is a convolution module with 3×3 convolution operation, or the output convolution module is a convolution module with 3×3 convolution operation.
[0022] Furthermore, the FLNet convolutional neural network performs stripe detection on the input image according to the following steps;
[0023] First, the stripe lines are detected using the FLNet convolutional neural network, the absolute phase and the entangled phase are generated using the SRTM DEM, and the horizontal and vertical phase jump boundaries are detected using a preset filter.
[0024] The detected phase jump boundaries are then overlapped to obtain stripe lines, which are used as labels for training the network. Noise is then added to the wrapped phase as input to the network.
[0025] The present invention provides a stripe detection phase unwrapping system based on FLNet convolutional neural network, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method.
[0026] The beneficial effects of this invention are as follows:
[0027] The present invention provides a stripe line detection and phase unwrapping method and system based on FLNet convolutional neural network, which belongs to the data processing method of interferometric synthetic aperture radar in radar imaging technology. The FLNet convolutional neural network is used to detect stripe lines with entangled phase to achieve phase unwrapping.
[0028] In the FLNet structure, a hybrid dilated convolution method is used to replace downsampling to avoid resolution loss, and residual connections are used to prevent the network from becoming too deep. Since FLNet's prediction of entangled phase stripes has high detection accuracy but cannot guarantee complete accuracy, this embodiment proposes using undirected graphs and ring integrals to repair broken stripes. FLNet is used to identify error points caused by noise, and by processing the phase of these error points, the influence of these error points on phase unwrapping is essentially eliminated.
[0029] Compared with traditional stripe detection methods, the deep learning-based stripe detection method proposed in this paper can effectively eliminate phase unwrapping errors caused by noise damaging stripes.
[0030] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0031] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:
[0032] Figure 1 This is a flowchart of phase unwrapping based on FLNet stripe detection.
[0033] Figure 2 This is a block diagram of the FLNet principle.
[0034] Figure 3 This is a structural diagram of the HDCRES module.
[0035] Figure 4 To simulate the entanglement phase and the corresponding stripe lines. (a) Noise entanglement phase image; (b) Corresponding stripe line image.
[0036] Figure 5 The broken stripe lines and their corresponding phase unwrapping results are shown. (a) The broken stripe lines predicted by FLNet; (b) The unwrapping results using the stripe lines in (a).
[0037] Figure 6 This is an image of broken stripes.
[0038] Figure 7 This is a schematic diagram of an undirected graph and a ring integral used to repair stripes. (a) Undirected graph obtained from stripes and pixel boundaries; (b) Ring integral graph.
[0039] Figure 8 The repaired stripe lines and the corresponding phase unwrapping results are shown. (a) Repaired stripe lines; (b) Results of unwrapping the stripe lines in (a).
[0040] Figure 9 These are the error points caused by stripe line errors.
[0041] Figure 10 These are the error points in the simulation. (a) Absolute phase image with error points; (b) Values and locations of the corresponding error points.
[0042] Figure 11 Error point identification and post-processing; (a) Error points predicted by FLNet; (b) Final phase unwrapping results.
[0043] Figure 12 The unwrapping results for the two study regions are shown in (a)-(b). Absolute phase diagrams of the two study regions are shown in (c)-(d). Wrapping phase diagrams of the two study regions are shown in (e)-(f). Unwrapping results of the stripes in the two study regions are shown in (g)-(h). Unwrapping error diagrams of the two study regions are shown in (i)-(j). The unwrapping results of the two study regions are shown in (j). Detailed Implementation
[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.
[0045] like Figure 1 As shown, the stripe line detection phase unwrapping method based on FLNet convolutional neural network provided in this embodiment includes the following steps:
[0046] Constructing an FLNet convolutional neural network;
[0047] Stripe detection is performed on the input image using an FLNet convolutional neural network.
[0048] Stripes obtained by repairing detection through ring integral;
[0049] Untangling via path integral;
[0050] Error points in stripe lines are identified using the FLNet convolutional neural network.
[0051] Handle error points.
[0052] This embodiment uses an undirected graph and a ring integral to repair the stripe lines. The specific process is as follows:
[0053] Use the four corner points of the pixel as graph nodes;
[0054] Connect all nodes on the stripe line except for the points on the image boundary, and determine the node that is connected to only one node as the breakpoint.
[0055] Connect two breakpoints using the nearest neighbor principle. If the breakpoints are far apart, connect the shortest path based on the distance between the nodes.
[0056] Then, using the principle that the closed-loop integral is 0, values are assigned to pixels near the connecting line, such as... Figure 7 As shown in (b), the paths in four directions need to pass through existing stripe lines and stripe lines to be connected. The solid red lines are existing stripe lines, and the dashed red lines are stripe lines to be connected. If the path passes from a pixel with a value of -1 to a pixel with a value of 1, the integral result is -1. If the path passes from a pixel with a value of 1 to a pixel with a value of -1, the integral result is 1. The integral result of the other paths is 0. According to this standard, the integral results of the four paths are added together, and the final result is required to be 0. This is used to determine the pixel direction of the path to be repaired. The values of two adjacent pixels are assigned to repair the stripe lines. That is, the integral result of the path of the stripe line to be repaired needs to be the opposite of the integral result of the path of the existing stripe lines. Therefore, the pixel order passed by the corresponding path is the opposite of the pixel order passed by the path of the existing stripe lines.
[0057] In this embodiment, path integral unwrapping is performed in the following manner:
[0058] Based on the unwrapping results of the horizontal and then vertical paths, the top left corner is selected as the reference point. If the path passes from a pixel with a value of 1 through a pixel with a value of -1, the corresponding wrapping phase needs to be increased by 2π. If the path passes from a pixel with a value of -1 through a pixel with a value of 1, the corresponding wrapping phase needs to be decreased by 2π.
[0059] like Figure 2 As shown, the FLNet convolutional neural network is used to detect stripes in the entangled phase to achieve phase untangling. The FLNet convolutional neural network includes an input convolutional module, several HDCRES modules, and an output convolutional module for feature extraction. HDCRES can expand the receptive field and enhance feature extraction performance without reducing resolution. The input convolutional module is used for preliminary extraction of input data features. The several HDCRES modules are used to extract features from the data. The output convolutional module is used for dimensionality reduction. The FLNet convolutional neural network provided in this embodiment has five HDCRES modules.
[0060] The FLNet convolutional neural network provided in this embodiment is a convolutional neural network for stripe detection and error point recognition. To avoid resolution loss, hybrid dilated convolution is used instead of downsampling to obtain a sufficient receptive field. Residual connections are added to HDC, and HDCRES is obtained by improving upon HDC. Figure 3 As shown, the HDCRES module includes several dilated convolution modules connected in sequence. The dilation rate of the dilated convolution modules is set according to a preset method. A zigzag dilation rate is used to avoid the extraction of discontinuous information. By setting different dilation rates, dilated convolution can obtain receptive fields of different sizes without losing resolution and without changing the number of parameters. The zigzag dilation rate in this embodiment can be set according to actual conditions or obtained by other methods. The selected dilation rate cannot have a common divisor greater than 1.
[0061] In this embodiment, the HDCRES module consists of three dilated convolution operations with dilation rates of 1, 2, and 3. In this embodiment, HDC can avoid information loss when dilated convolution is used continuously, and residual connections can prevent the network from becoming too deep. In this embodiment, HDC and residual connections are combined to obtain HDCRES.
[0062] Therefore, the input image first undergoes a 3×3 convolution operation, then passes through five HDCRES blocks, and finally undergoes another 3×3 convolution operation to complete the classification. The specific implementation process is as follows:
[0063] First, FLNet is used to detect stripe lines. Absolute phase and wrapped phase are generated using SRTM DEM. A 64×64 sliding window is used to truncate the SRTM DEM with a step size of 64. The SRTM DEM is converted into absolute phase, and then the absolute phase is wrapped to obtain a noise-free wrapped phase. A preset filter is used to detect the horizontal and vertical phase jump boundaries in the wrapped phase. In this embodiment, the preset filter uses 1×2 and 2×1 filters. When the product of two pixels in the filter is negative and the absolute values of the two pixels are both greater than the set threshold, it is considered that a phase jump has occurred, and the horizontal and vertical jump boundaries can be detected.
[0064] The horizontal and vertical transition boundaries are then overlapped to obtain complete stripe lines, which serve as labels for training the network. Noise is then added to the noise-free wrapped phase as input to the network. Figure 4 The image shows the noisy, wrapped phase and the corresponding stripe lines. The stripe line image consists of -1, 0, and 1; on the boundary from 1 to -1, the phase should be increased by 2π, otherwise subtracted by 2π. The image size is 64×64, and the number of images in the training and test datasets are 8000 and 800, respectively.
[0065] Figure 5Image (a) shows the striped lines predicted by the trained FLNet on the test set. It can be seen that FLNet can effectively predict striped lines with intertwined phases, achieving an average accuracy of 96.1% across 800 test images. While the detection accuracy is high enough, it cannot guarantee complete accuracy; therefore, some striped lines may be broken, such as... Figure 5 The location marked with a red box in (a).
[0066] exist Figure 6 In this case, when the wrapped phase is unfolded and the integration path crosses the broken stripe line, the final result will be increased or decreased by 2π compared to the actual result.
[0067] Figure 5 Figure (b) shows the unwrapping result through a path that is first horizontal and then vertical. The upper left corner is selected as the reference point. When the integration path passes through the top breakpoint region, the error will propagate to the right half of the unwrapping result.
[0068] This embodiment uses an undirected graph to repair stripe lines, such as Figure 7 As shown, Figure 7 This is a method for repairing undirected graphs and circular integrals. It treats the four corner points of a pixel as graph nodes, and connects all nodes along the stripe lines. Figure 7 In (a), a node connected to only one node, except for points on the image boundary, is a breakpoint, such as... Figure 7 In (a), points 1 and 2 are connected using the nearest neighbor principle, and then the pixels near the connection line are assigned values using the principle that the closed-loop integral is 0. Figure 7 (b)). For nodes that are far apart, find the shortest connection path. Figure 8 Image (a) shows the repaired stripes. (Compared to) Figure 5 Compared to (a), the breaks on the same stripe line are correctly connected, and the direction of the stripe line is also correct. Figure 8 (b) Figure 8 The results of the expansion in (a) are compared. Figure 5 In diagram (b), it can be seen that the error propagation problem has been solved.
[0069] Secondly, due to the noise inherent in the wrapped phase, the stripe lines themselves cannot be perfectly accurate, resulting in error points in the unwrapped phase image. Compared to the true phase value, these points increase or decrease by 2π, and are mainly distributed around the stripe lines. Figure 9 Therefore, FLNet is continued for error point identification. The absolute phase is obtained using SRTM DEM, and after adding noise, some points are randomly selected, with 2π increments or decrements made. Figure 10 In (a)), these random points are stored as training labels. Figure 10In (b), the label value is 1 when the error point increases by 2π, otherwise it is -1. The network input is the absolute phase of the error point, and the output is the position and value of the error point. The image size is 64×64, and the training and test dataset sizes are 8000 and 800, respectively. Figure 8 (b) and Figure 11 (a) Further post-processing is performed, that is, based on the identified error points, if the error point is 1, the corresponding pixel value is subtracted by 2π; if the error point is -1, the corresponding pixel value is added by 2π; if the error point is 0, no processing is required, thus obtaining the final phase unwrapping result. Figure 11 As can be seen from (b), most of the error points have been eliminated.
[0070] exist Figure 12 The image shows the phase unwrapping results for two other study areas. Looking at the phase unwrapping error image (fourth row), only a few pixels have errors, and these errors are integer multiples of 2π. Therefore, the rewrapped phase image (fifth row) is still consistent with the original wrapped phase image (second row), proving that this method can effectively guarantee the consistency of the unwrapped stripes.
[0071] This embodiment also provides a stripe detection phase unwrapping system based on FLNet convolutional neural network, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method steps.
[0072] 1. This embodiment proposes a convolutional neural network called FLNet to detect tangled phase stripes to achieve phase untangling.
[0073] 2. In the FLNet architecture, this embodiment uses hybrid dilated convolution to replace downsampling to avoid resolution loss, and uses residual connections to prevent the network from becoming too deep.
[0074] 3. Although FLNet's prediction of entangled phase stripes has high detection accuracy, it cannot guarantee complete accuracy. This embodiment proposes using undirected graphs and ring integrals to repair broken stripes. In this embodiment, the undirected graph and ring integral methods are used for stripe repair, such as... Figure 1 As shown, Figure 1 There are steps involved in repairing striped lines. Figure 5 The phase unwrapping result is without stripe line repair. Figure 8 This is the result of untangling the striped lines after repair. Figure 5 and Figure 8 The before-and-after comparison shows the importance of this repair solution.
[0075] 4. This embodiment proposes to use FLNet to identify error points caused by noise. By processing the phase of the error points, the influence of the error points on phase unwrapping is basically eliminated.
[0076] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A stripe line detection phase unwrapping method based on FLNet convolutional neural network, characterized in that: Includes the following steps: Constructing an FLNet convolutional neural network; Stripe detection is performed on the input image using an FLNet convolutional neural network. Stripes obtained by repairing detection through ring integral; Untangling via path integral; The error points of the stripe lines are identified and processed using the FLNet convolutional neural network to obtain the output result. The detected stripe lines are repaired according to the following steps: Use the four corner points of the pixel as graph nodes; Connect all nodes on the stripe line except for the points on the image boundary, and determine the node that is connected to only one node as the breakpoint. Connect two breakpoints using the nearest neighbor principle, and then assign values to pixels near the connection line using the principle that the closed-loop integral is 0. If the breakpoints are far apart, connect the shortest path based on the distance between the nodes; Connect the breakpoints on the same stripe line, determine the pixel direction of the path to be repaired, and assign values to two adjacent pixels to repair the stripe line.
2. The stripe line detection phase unwrapping method based on FLNet convolutional neural network as described in claim 1, characterized in that: The path integral unwrapping is performed in the following manner: By using the unwrapping results of the horizontal and then vertical paths, the upper left corner is selected as the reference point. When passing through the stripe line, 2π is added or subtracted according to the direction of the stripe line to achieve unwrapping.
3. The stripe line detection phase unwrapping method based on FLNet convolutional neural network as described in claim 1, characterized in that: The FLNet convolutional neural network includes an input convolution module, several HDCRES modules, and an output convolution module. The HDCRES module includes several dilated convolution modules connected in sequence. The input convolution module performs convolution operations on the input image, which is then processed by several HDCRES modules, and finally classified by the convolution operation of the output convolution module.
4. The stripe line detection phase unwrapping method based on FLNet convolutional neural network as described in claim 3, characterized in that: The HDCRES module includes three dilated convolution modules, the dilation rate of which is set by using a zigzag dilation rate.
5. The stripe line detection phase unwrapping method based on FLNet convolutional neural network as described in claim 3, characterized in that: The input convolution module is a convolution module with a 3×3 convolution operation, or the output convolution module is a convolution module with a 3×3 convolution operation.
6. The stripe line detection phase unwrapping method based on FLNet convolutional neural network as described in claim 1, characterized in that: The FLNet convolutional neural network performs stripe detection on the input image according to the following steps; First, the stripe lines are detected using the FLNet convolutional neural network, the absolute phase and the entangled phase are generated using the SRTM DEM, and the horizontal and vertical phase jump boundaries are detected using a preset filter. The detected phase jump boundaries are then overlapped to obtain stripe lines, which are used as labels for training the network. Noise is then added to the wrapped phase as input to the network.
7. A stripe line detection and phase unwrapping system based on FLNet convolutional neural network, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 6.