Underground cable radar image detection method based on scale knowledge and deep learning

Abstract

A kind of underground cable radar image detection method based on scale knowledge and deep learning, comprising: S100: determine the scale size of underground cable in radar image;S200: calculate Gaussian kernel template according to scale size;S300: Gaussian kernel template is embedded in convolutional neural network framework, and the first convolution kernel greater than 1x1 in convolutional neural net is valued with the template parameter value calculated;S400: lock the parameters of first convolution layer, keep the parameters of this layer unchanged;S500: other parameters in network model are adjusted using training data, until the lowest solution of loss function minimum is obtained;S600: test data set is substituted into the model obtained by training, and the performance of the model is verified, steps S500-S600 are repeated until the preset training end condition is reached;S700: the model parameter with highest detection accuracy on test data set is determined as the final result.The present application reduces the amount of parameters required, and reduces the demand for training data amount.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar image detection, and particularly relates to a radar image detection method for underground cables based on scale knowledge and deep learning. BACKGROUND

[0002] B-scan radar image is the most common type of radar signal, and most engineering applications are based on the information provided by the radar image. Since the underground medium is not uniform, and various possible impurities exist, the useful information of the radar image is often hidden in the noisy image background, and even manual detection is quite difficult.

[0003] Since deep learning has very strong learning and expression ability, it can theoretically express any complex relationship, so the detection target recognition in the radar image often relies on deep learning methods. However, deep learning is a universal method, and a special deep learning model needs to be trained for a specific task. In order to train the model (i.e. to obtain the model parameters), a large amount of labeled training data is required, and the collection and labeling of such data is quite time-consuming, and the amount of data is often limited in practical applications. SUMMARY

[0004] In view of the above problems, the present application is proposed in order to provide a method and system for counting urban transient personnel and permanent population based on big data, which overcomes the above problems or at least partially solves the above problems.

[0005] In order to solve the above technical problems, the present application discloses the following technical solutions:

[0006] The present application discloses a radar image detection method for underground cables based on scale knowledge and deep learning, comprising:

[0007] S100: determining the scale size of the underground cable in the radar image;

[0008] S200: calculating a Gaussian kernel template according to the scale size;

[0009] S300: embedding the Gaussian kernel template into a convolutional neural network framework, and using the calculated template parameter value to assign values to the first convolutional kernel larger than 1x1 in the convolutional neural network;

[0010] S400: locking the parameters of the first convolutional layer, and keeping the parameters of the layer unchanged;

[0011] S500: adjusting other parameters in the network model using training data until the lowest solution of the loss function is obtained;

[0012] S600: Substitute the test data set into the trained model to verify the performance of the model, repeat steps S500-S600 until the preset training end condition is reached;

[0013] S700: Determine the model parameter with the highest detection accuracy on the test data set as the final result.

[0014] Further, in S200, the method for calculating the Gaussian kernel template according to the scale size is:

[0015] S201. Taking the center of the template as the coordinate origin, the horizontal right as the positive direction of the x-axis, and the vertical upward as the positive direction of the y-axis, the coordinates (x, y) of each position in the template are obtained;

[0016] S202. Assuming that the scale of the underground cable in the radar image is σ, substitute the coordinates (x, y) of each position into the following formula:

[0017]

[0018] S203. Then the sum of all position coordinate values is calculated:

[0019]

[0020] S204. Then S·G(x, y) is the value of the template at (x, y).

[0021] Further, in S300, the convolutional neural network framework adopts an 18-layer Resnet network.

[0022] Further, the calculated Gaussian convolution template is used to replace the convolution template with a size of 3*3 and a channel number of 64 in the uppermost layer of the 18-layer Resnet network.

[0023] Further, the specific method of S300 is:

[0024] S301. Train the Resnet18 model, only substitute a set of data, do not run the iterative optimization process, save the output model parameters, and obtain the data format of the model;

[0025] S302. Find the parameter storage position of the convolution template with a size of 3*3 and a channel number of 64 in the first layer of the Resnet18 model, and replace the corresponding value with the calculated value of the Gaussian convolution;

[0026] S303. Lock the convolution template parameters of this layer, which remain unchanged in the subsequent learning process;

[0027] S304. Iterative optimization to determine other model parameters in the network.

[0028] Further, in S600, steps S500-S600 are repeated until a preset training end condition is reached, the training end condition being that the number of training times reaches a given number or the error precision reaches a given precision.

[0029] The above technical solution provided by the embodiments of the present application has at least the following beneficial effects:

[0030] The underground cable radar image detection method based on scale knowledge and deep learning disclosed by the present application comprises the following steps: S100, determining the scale size of the underground cable in the radar image; S200, calculating a Gaussian kernel template according to the scale size; S300, embedding the Gaussian kernel template into a convolutional neural network framework, and assigning the calculated template parameter value to the first convolution kernel larger than 1*1 in the convolutional neural network; S400, locking the parameters of the first convolution layer, and keeping the parameters of the layer unchanged; S500, adjusting other parameters in the network model by using training data until the lowest solution when the loss function is the smallest is obtained; S600, substituting a test data set into the model obtained by training to verify the performance of the model, and repeating steps S500-S600 until a preset training end condition is reached; and S700, determining the model parameter with the highest detection precision on the test data set as the final result. The present application introduces field knowledge into a deep learning framework, directly replaces the parameters of the first convolution layer to be trained with a Gaussian kernel template based on a convolutional neural network framework, thereby introducing the scale knowledge of the detection target into the deep learning model, reducing the amount of parameters to be obtained, and reducing the demand for the amount of training data.

[0031] The technical solutions of the present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0033] Figure 1 For the flowchart of the underground cable radar image detection method based on scale knowledge and deep learning in the embodiment 1 of the present application;

[0034] Figure 2 For the coordinate diagram of each point in the template in the embodiment 1 of the present application;

[0035] Figure 3 For the Resnet18 network framework diagram in the embodiment 2 of the present application;

[0036] Figure 4 For the underground cable image data diagram generated by gprMax in the embodiment 2 of the present application. DETAILED DESCRIPTION

[0037] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be thoroughly understood and fully conveyed to those skilled in the art.

[0038] To solve the problems in the prior art, an underground cable radar image detection method based on scale knowledge and deep learning is provided.

[0039] Embodiment 1

[0040] An underground cable radar image detection method based on scale knowledge and deep learning, as Figure 1 , comprises:

[0041] S100: determining the scale size of the underground cable in the radar image;

[0042] S200: calculating a Gaussian kernel template according to the scale size;

[0043] S300: embedding the Gaussian kernel template into a convolutional neural network framework, and using the calculated template parameter value to assign a value to the first convolutional kernel larger than 1x1 in the convolutional neural network;

[0044] S400: locking the parameters of the first convolutional layer, and keeping the parameters of the layer unchanged;

[0045] S500: adjusting other parameters in the network model using training data until the lowest solution of the loss function is obtained;

[0046] S600: substituting the test data set into the model obtained by training, verifying the performance of the model, repeating steps S500-S600 until the preset training end condition is reached;

[0047] S700: determining the model parameter with the highest detection accuracy on the test data set as the final result.

[0048] In the present embodiment S200, the method for calculating the Gaussian kernel template according to the scale size is:

[0049] S201. Taking the center of the template as the coordinate origin, the horizontal right as the positive direction of the x-axis, and the vertical upward as the positive direction of the y-axis, the coordinates (x, y) of each position in the template are obtained;

[0050] S202. Assuming that the scale of the underground cable in the radar image is σ, substituting the coordinates (x, y) of each position into the following formula:

[0051]

[0052] S203. Then find the sum of all position coordinates:

[0053]

[0054] S204. Then S·G(x, y) is the value at (x, y) in the template.

[0055] The calculation method of Gaussian kernel template, such as Figure 2 As shown, the 3×3 model is taken as an example to illustrate as follows:

[0056] With the center of the template as the coordinate origin, the horizontal right direction is the positive direction of the x-axis, and the vertical upward direction is the positive direction of the y-axis. The coordinates of each point in the template are as follows: Figure 1 shown.

[0057] Assume that the scale of the underground cables in the radar image is σ, and substitute the coordinates of each position in the image into the following formula:

[0058]

[0059] If we substitute the coordinates of the upper left corner (-1, 1), and assume that the scale of the underground cables in the radar image is 5 (the data can be adjusted according to actual conditions), we get

[0060]

[0061] The same method is used to calculate the values ​​corresponding to all positions (9 in this example), and then find their sum

[0062]

[0063] Then S·G(x,y) is the value at (x,y) in the template.

[0064] In some limited embodiments, taking the most outstanding convolutional neural network Resnet18 as an example, assuming that the scale s of underground cables in radar images is greater than 5 pixels,

[0065] (1) Figure 3 As shown, first build an 18-layer Resnet network

[0066] Figure 3 In the example, “1*1conv,64” means that the convolutional template size in this layer is 1*1 and the number of channels is 64. Similarly, “3*3conv,128” means that the convolutional template size in this layer is 3*3 and the number of channels is 128.

[0067] (2) Replace the convolution template in the top “3*3conv,64” layer with the calculated Gaussian convolution template.

[0068] Specific operation methods:

[0069] ①Same as general deep learning, train Resnet18 model, the difference is only to replace a set of data, do not run the iterative optimization process, save the output model parameters, get the model data format;

[0070] ②Find Figure 3 the parameter storage location of the convolution template in the first "3*3conv, 64" layer, replace the corresponding value with the calculated value of the Gaussian convolution;

[0071] ③Lock the convolution template parameters of this layer, so that they remain unchanged in the subsequent learning process;

[0072] ④Iterative optimization to determine other model parameters in the network.

[0073] Once the model parameters are determined, the model can be used to detect underground cables in radar images

[0074] As Figure 4 shown, to verify the effectiveness of the model, experiments were conducted on underground cable image data generated based on gprMax, and the results showed that the accuracy was as high as 98.7%, which was 1.4 percentage points higher than the result of directly training the Resnet network.

[0075] The underground cable radar image detection method based on scale knowledge and deep learning disclosed in the embodiment comprises: S100: determining the scale size of the underground cable in the radar image; S200: calculating a Gaussian kernel template according to the scale size; S300: embedding the Gaussian kernel template into a convolutional neural network framework, and assigning the calculated template parameter value to the first convolution kernel larger than 1*1 in the convolutional neural network; S400: locking the parameters of the first convolution layer, keeping the parameters of the layer unchanged; S500: adjusting other parameters in the network model using training data until the lowest solution of the loss function is obtained; S600: substituting the test data set into the model obtained by training to verify the performance of the model, repeating steps S500-S600 until the preset training end condition is reached; S700: determining the model parameter with the highest detection accuracy on the test data set as the final result. The present application introduces domain knowledge into the deep learning framework, and directly replaces the parameters of the first convolution layer to be trained with a Gaussian kernel template based on the convolutional neural network framework, thereby introducing the scale knowledge of the detection target into the deep learning model, reducing the amount of parameters to be obtained, and reducing the demand for the amount of training data.

[0076] It should be understood that the particular order in which the steps of processes presented in the disclosure have been presented can be rearranged. Furthermore, various aspects of the disclosure can be used alone or in various combinations. In the appended claims, means-plus-function or step-plus-function clauses can have their application interpreted merely as open-ended limits of a particular combination of steps that are done on the apparatus, as specified in 35 U.S.C. § 112(f). However, the steps of the claims are not to be interpreted as requiring their performance in the order recited.

[0077] In the detailed description above, various features are grouped together in single embodiments for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting a necessity to more features than are expressly identified in each claim. Rather, inventive functionality can be more effectively disclosed by grouping features in a single claim. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate preferred embodiment.

[0078] Those skilled in the art will further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0079] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0080] For a software implementation, the techniques described herein can be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. The software codes can be stored in memory units and executed by processors. The memory unit can be implemented within the processor or external to the processor, in which case it can be communicatively coupled to the processor via various means as is known in the art.

[0081] The above description includes examples of one or more embodiments. Of course, not all possible combinations of components or methods described above can be implemented to make up a particular embodiment, but one of ordinary skill in the art will recognize that many such combinations are possible. Accordingly, the present embodiments are intended to embrace all such alterations, modifications and variations that fall within the scope of the appended claims. Furthermore, to the extent that the term "includes" is used in either the detailed description or the claims, such term is intended to be interpreted as "including but not limited to." Additionally, where the description or the claims recite a range of values, it is intended that every possible value within the range be "incorporated" into the disclosure.

Claims

1. A radar image detection method for underground cables based on scale knowledge and deep learning, characterized in that: include: S100: determining the size of underground cables in the radar image; S200: Calculate a Gaussian kernel template according to the scale size, where the size of the Gaussian kernel template is 3*3. In S200, the method for calculating the Gaussian kernel template according to the scale size is: S201. With the center of the template as the coordinate origin, the horizontal right direction is the positive x-axis direction, and the vertical upward direction is the positive y-axis direction, the coordinates (x, y) of each position in the template are obtained; S202. Assume that the scale of underground cables in the radar image is , substitute the position coordinates (x, y) into the following formula: ； S203. Find the sum of all position coordinates: ； S204. As the template is located The value at S300: Embed the Gaussian kernel template into the convolutional neural network framework, and use the calculated template parameter value as the first one in the convolutional neural network that is greater than The convolution kernel is assigned; in S300, the convolution neural network framework adopts an 18-layer Resnet network; the calculated template parameter value is used to replace the first convolution layer parameter in the 18-layer Resnet network; the specific method of S300 is: S301. Train the Resnet18 model, substituting only one set of data without running the iterative optimization process, saving the output model parameters, and obtaining the model data format; S302. Find the parameter storage location of the first convolutional template with a size of 3*3 and 64 channels in the Resnet18 model, and replace its corresponding value with the calculated value of the Gaussian kernel template; S303. Lock the parameters of the first convolutional layer and keep them unchanged during the subsequent learning process; S304. Iterate and optimize to determine other model parameters in the network; S400: Lock the parameters of the first convolutional layer and keep the parameters of this layer unchanged; S500: Using the training data to adjust other parameters in the network model until the lowest solution with the minimum loss function is obtained; S600: Substitute the test data set into the trained model to verify the performance of the model, and repeat steps S500-S600 until the preset training end condition is met; S700: Determine the model parameters with the highest detection accuracy on the test data set as the final result.

2. The underground cable radar image detection method based on scale knowledge and deep learning according to claim 1, characterized in that: In S600, steps S500-S600 are repeated until a preset training end condition is reached, where the training end condition is that the number of training times reaches a given number or the error accuracy reaches a given accuracy.

Images