A method for constructing a tunnel void state radar spectrum image recognition model and a method for recognizing tunnel void state radar spectrum image

Abstract

The present invention discloses a tunnel void state radar spectrum image recognition model construction method and a tunnel void state radar spectrum image recognition method, belonging to the field of image recognition technology. The method extracts the time domain abstract semantic features from shallow to deep in the tunnel void state radar spectrum image through a convolutional neural network, and constructs a multi-source time-frequency domain complete global observation feature space of the tunnel void state radar spectrum image by fusing the frequency domain grayscale statistical features extracted by Fourier transform, and then performs feature splicing on different feature vectors to obtain the final feature vector; the final feature vector is input into the SCN classifier to obtain a heterogeneous deep feature network model, so that the method of the present invention is closer to the human visual model, fully utilizes the time domain and frequency domain characteristics of the image, and has strong adaptability and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition technology and deep learning technology, and specifically relates to a method for constructing a tunnel void state radar spectrum image recognition model and a tunnel void state radar spectrum image recognition method. Background Art

[0002] With the country's massive investment in infrastructure such as roads and railways, the number of bridges and tunnels built has increased day by day. However, with the continuous increase in the number of operating tunnels and the continuous increase in the operating time, a large number of tunnels have entered overhaul due to the existence of voids, which seriously affects the life of the lining structure and driving safety. Therefore, it is crucial to study the identification method of voids behind the lining for the safety impact of tunnel structures and preventive measures. Among them, the geological radar detection of tunnel lining uses electromagnetic waves in the form of broadband short pulses, which are sent from the lining surface to the inside of the tunnel lining through the antenna transmitter, and then reflected by the target body (voids, etc.) in the lining or the interface of different linings and returned to the lining surface, and received by the radar antenna receiver. By processing and interpreting the received radar signal, the purpose of detecting the target body in the tunnel lining is achieved. It can be widely used in non-destructive testing of the construction quality of railway, highway, and hydropower tunnel linings. It can not only effectively evaluate the quality of existing tunnel linings, but also find the quality defects of new tunnel linings, and also control the construction quality of tunnel linings during the construction stage.

[0003] However, traditional methods for tunnel cavity radar wave image detection are based on manual detection, which requires inspectors to manually identify the image. There is a possibility of improper detection by personnel, resulting in the disadvantage of insufficient accuracy. In recent years, with the increasing application of machine learning in image recognition, it has become possible to automatically identify tunnel cavity radar wave images.

[0004] For example, the prior art discloses an automatic tunnel disease identification device based on panoramic image CNN. First, a panoramic image of the inner wall of the tunnel is quickly acquired through a panoramic visual sensor, and then the panoramic image is processed, mainly including panoramic image expansion, image preprocessing, binarization processing, etc. to extract suspected disease areas; finally, a convolutional neural network is used to automatically detect and classify the diseases. However, the image features extracted by this method are relatively simple, and the recognition accuracy of the constructed model is low. Summary of the invention

[0005] 1. Problems to be solved

[0006] In view of the problem that the image feature extraction in the existing tunnel cavity radar wave image detection is relatively simple, resulting in the low recognition accuracy of the constructed model. The present invention provides a tunnel cavity state radar spectrum image recognition model construction method and a tunnel cavity state radar spectrum image recognition method, by obtaining first image information, wherein the first image information includes a radar spectrum image without tunnel cavity and a radar spectrum image of tunnel cavity; performing two-dimensional Fourier transformation and grayscale processing on the first image information to obtain second image information, performing grayscale value statistics on the pixels of the second image information, and sequentially splicing to form a first feature vector; inputting the first image information into a heterogeneous deep feature network for convolution, and outputting the obtained second feature vector; performing feature fusion splicing on the first feature vector and the second feature vector to obtain a third feature vector; then inputting the third feature vector into an SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, saving the current heterogeneous deep feature network model. The time domain abstract semantic features of the tunnel void state radar spectrum image from shallow to deep are extracted by convolutional neural network, and the frequency domain grayscale statistical features extracted by Fourier transform are integrated to construct a multi-source time-frequency domain complete global observation feature space of the tunnel void state radar spectrum image. This makes the method of the present invention closer to the human visual model, fully utilizes the time domain and frequency domain characteristics of the image, and has strong adaptability and robustness.

[0007] 2. Technical solution

[0008] To solve the above problems, the present invention adopts the following technical solutions.

[0009] A first aspect of the present invention provides a method for constructing a tunnel cavity state radar spectrum image recognition model, comprising:

[0010] S102: Acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity;

[0011] S104: performing a two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, performing grayscale value statistics on pixel points of the second image information, and sequentially splicing them to form a first feature vector;

[0012] S106: Inputting the first image information into a heterogeneous deep feature network for convolution, and outputting a second feature vector;

[0013] S108: performing feature fusion and splicing on the first feature vector and the second feature vector to obtain a third feature vector;

[0014] S110: Input the third feature vector into the SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, save the current heterogeneous deep feature network model.

[0015] Preferably, the step S110 further includes: when constructing a heterogeneous deep feature network, defining an artificially set error threshold and the maximum number of iterations , the number of initialization iterations is ;

[0016] definition Heterogeneous deep feature network The root mean square error of the network output of the iteration is used to determine or Is it true? If so, save the current heterogeneous deep feature network model.

[0017] , Indicates the number of samples contained in the first image information.

[0018] Preferably, the step S110 further includes: when or When it is not true, the weights and biases of the heterogeneous deep feature network from the fully connected layer to the input layer are updated according to the back-propagation algorithm of gradient descent, and the back-propagation process from the output layer of the heterogeneous deep feature network to the fully connected layer is calculated using formulas (1-1)-(1-3):

[0019] (1-1)

[0020] (1-2)

[0021] (1-3)

[0022] in, Heterogeneous deep feature network The root mean square error of the network output of the iteration; represents the network's expected output set, , is the expected output of the heterogeneous deep feature network, Heterogeneous deep feature network The actual output of the iteration network; Indicates network The feature set of the fully connected layer of the network at the time of the update, , Indicates the first image information The image is input into the heterogeneous deep feature network. The third eigenvector obtained by the iterative network; Represents the SCN classifier The weight of each node to the output layer; Represents the fully connected layer to the SCN classifier The weight of each node; Represents the fully connected layer to the SCN classifier The bias of each node; Represents the SCN classifier The activation function of each node; Indicates the number of nodes of the SCN classifier; Indicates the number of samples contained in the first image information.

[0023] Preferably, the step S104 further includes:

[0024] Locating the position of at least one feature in the first image information; the at least one feature is determined by a tunnel cavity state detection model obtained by training a YOLO model;

[0025] An image containing at least one feature in the first image is cropped, and two-dimensional Fourier transform and grayscale processing are performed to obtain second image information.

[0026] Preferably, the second image information is input into a heterogeneous deep feature network, and a 7×7 size is obtained through a convolution operation, and the number of convolution kernels is feature space; then input into two network branches, one of which is a 7×7 average pooling layer and the number of neurons is The other two branches are 4×4 average pooling layers and the number of neurons is The outputs of the two fully connected layers are then concatenated to obtain a dimension of The second eigenvector of .

[0027] A second aspect of the present invention provides a method for recognizing a tunnel cavity state radar spectrum image, comprising:

[0028] Acquire an image to be detected;

[0029] The image to be detected is input into the tunnel void state radar spectrum image recognition model constructed by the tunnel void state radar spectrum image recognition model construction method, the feature image of the image is determined by the convolutional neural network model, and the recognition result of the image is determined according to the feature image.

[0030] A third aspect of the present invention provides a device for constructing a tunnel cavity state radar spectrum image recognition model, comprising:

[0031] A training data acquisition unit, configured to acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity;

[0032] A feature extraction unit, configured to perform a two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, perform grayscale value statistics on pixel points of the second image information, and sequentially concatenate the pixels to form a first feature vector;

[0033] A convolution unit, used for inputting the first image information into a heterogeneous deep feature network for convolution, and outputting a second feature vector obtained;

[0034] a feature fusion unit, configured to perform feature fusion and splicing on the first feature vector and the second feature vector to obtain a third feature vector; and

[0035] The model building unit is used to input the third feature vector into the SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, the current heterogeneous deep feature network model is saved.

[0036] A fourth aspect of the present invention provides a storage medium that non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, instructions for constructing a tunnel void state radar spectrum image recognition model can be executed, or instructions for recognizing a tunnel void state radar spectrum image can be executed according to the method.

[0037] 3. Beneficial effects

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention extracts the time domain abstract semantic features of the tunnel void state radar spectrum image from shallow to deep through a convolutional neural network, integrates the frequency domain grayscale statistical features extracted by Fourier transform, and constructs a multi-source time-frequency domain complete global observation feature space of the tunnel void state radar spectrum image, so that the method of the present invention is closer to the human visual model, fully utilizes the time domain and frequency domain characteristics of the image, and has strong adaptability and robustness;

[0040] 2. The present invention adopts a dual-branch network architecture to obtain multi-resolution feature maps of tunnel cavity state radar spectrum images, which enhances the adaptability of the time domain abstract semantic feature space to images of different resolutions, and fuses multi-scale features based on a dual-scale network architecture, thereby enhancing the distinguishability of the feature space for multi-category lining cavity images;

[0041] 3. The present invention adopts a random configuration network to replace the softmax layer in the traditional convolutional neural network, realizes the universal approximation of the multi-scale features of the radar spectrum image of the tunnel cavity state to the lining cavity, and enhances the learning and generalization capabilities of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and other purposes, features and advantages of the present application will become more apparent by describing the embodiments of the present application in more detail in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps. In the accompanying drawings:

[0043] Figure 1 A radar spectrum image A including a tunnel cavity state provided by an embodiment of the present invention;

[0044] Figure 2 Provides a radar spectrum image of a tunnel cavity state captured for an embodiment of the present invention;

[0045] Figure 3 A structural feature graph of a heterogeneous deep feature network provided by an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of a tunnel cavity state detection result obtained by a Yolo detection model provided in an embodiment of the present invention;

[0047] Figure 5 A schematic diagram of detection results of a heterogeneous deep feature network model for tunnel cavity state detection provided by an embodiment of the present invention;

[0048] Figure 6 A schematic flow chart of a method for constructing a radar spectrum image recognition model for tunnel void status provided by an embodiment of the present invention;

[0049] Figure 7 A schematic structural diagram of a device for building a tunnel cavity state radar spectrum image recognition model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0051] Exemplary Methods

[0052] like Figure 6 As shown, a method for constructing a tunnel cavity state radar spectrum image recognition model includes:

[0053] S102: Acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity.

[0054] In this example, the first image information is a set of radar spectrum images of tunnel cavity status obtained by geological radar, including radar spectrum images without tunnel cavity and radar spectrum images of tunnel cavity.

[0055] A set of possible implementations Before feature extraction, it is necessary to preprocess the set of radar spectrum images of tunnel void status acquired by geological radar to detect whether the computer computing power meets the calculation requirements of the original image size and to determine whether there is image blur and other problems. The radar spectrum images of tunnel void status are resized and denoised to obtain a set of preprocessed radar spectrum images of tunnel void status; then the preprocessed radar spectrum images of tunnel void status are classified into a set of radar spectrum images with tunnel voids and a set of radar spectrum images with tunnel voids. and radar spectrum image collection without tunnel cavity Please refer to Figure 1 This is a radar spectrum image with a tunnel cavity.

[0056] The purpose of grouping is to reduce the number of parameters while improving accuracy. By grouping, higher accuracy can be achieved under the same parameter scale, which is beneficial to not significantly increase the parameter scale of the model after increasing the number of parameters of the convolution kernel of the weight matrix, and better improve the recognition accuracy; while saving network parameters, better use of image local information.

[0057] A set of possible implementations , the radar spectrum image collection with tunnel voids and radar spectrum image collection without tunnel cavity Take each Manually annotate the images and obtain a collection of radar spectrum images with tunnel cavities middle The coordinate position of the rectangular bounding box where the tunnel void area of ​​the image is located is obtained, and then the set of manually annotated tunnel void state radar spectrum images is obtained as the training set for tunnel void state recognition. ; The radar spectrum image collection with tunnel cavity and radar spectrum image collection without tunnel cavity The rest of the tunnel void state radar spectrum image set is used as the test set for tunnel void state recognition. Using the training set The target detection YOLO model is trained to obtain a tunnel cavity state detection model, and then the tunnel cavity state detection model is used to detect the test set. Detection is performed to obtain the coordinate position of the rectangular bounding box of the radar spectrum image with tunnel holes in the test set. In this example, the YOLO model can also achieve classification and recognition. Figure 4 The test set is shown The effect diagram of the detection by the trained YOLO detection model; the highest detection accuracy that can be achieved by detecting radar spectrum images containing tunnel voids is 0.738.

[0058] Crop the training set The manually annotated rectangular bounding boxes and test set in All rectangular bounding boxes containing radar spectrum images of tunnel void states in the training set and test set The rest of the radar spectrum images that do not contain tunnel voids are normalized to obtain a radar spectrum image set of target tunnel voids; the radar spectrum image set of the target tunnel voids is then reclassified into a radar spectrum image set with tunnel voids, which is recorded as The radar spectrum image set with and without tunnel void is recorded as ; Take out from the radar spectrum image set g with tunnel cavity A radar spectrum image with tunnel voids is obtained, and a radar spectrum image without tunnel voids is taken from the radar spectrum image set v. The radar spectrum images without tunnel voids are used as training sets . Preferably, the training set here It can also be used as the first image information and input into a heterogeneous deep feature network for convolution.

[0059] S104: Perform two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, perform grayscale value statistics on pixel points of the second image information, and sequentially splice them to form a first feature vector.

[0060] In this example, the position of at least one feature in the first image information is located; the at least one feature here is one of the features determined by the tunnel cavity state detection model obtained by training the YOLO model, specifically refers to Figure 2 The position of the tunnel cavity area enclosed by the rectangular bounding box shown in the figure; the image containing at least this feature in the first image is cropped, and a two-dimensional Fourier transform and grayscale processing are performed to obtain the second image information. Those skilled in the art should understand that the features here can also be extracted according to other methods, such as local binary pattern (LBP) feature extraction.

[0061] A set of possible implementations, The radar spectrum images of the remaining tunnel void states in the radar spectrum image set g with tunnel voids and the radar spectrum image set v without tunnel voids are used as the test set .

[0062] S106: Input the first image information into a heterogeneous deep feature network, and output the obtained second feature vector.

[0063] In this example, the first image information is input into the heterogeneous deep feature network to obtain the second feature vector output by the network , the dimension is ; In this example, the dimension of the second eigenvector is .

[0064] A set of possible implementations, like Figure 3 The structure of the heterogeneous deep feature network shown is as follows:

[0065] The first layer is a 7×7 convolution layer, and the number of convolution kernels is The second layer is a 3×3 maximum pooling layer, and then two branches are generated.

[0066] The first branch consists of 4 parts. The first part consists of 4 3×3 convolutional layers. The number of convolution kernels is The second part consists of 4 3×3 convolutional layers, and the number of convolution kernels is The third part consists of 4 3×3 convolutional layers, and the number of convolution kernels is The fourth part consists of 4 3×3 convolutional layers, and the number of convolution kernels is indivual.

[0067] The second branch consists of 4 parts. The first part is a 1×1 convolution layer, followed by a 3×3 convolution layer and two 1×1 convolution layers, and finally a 3×3 convolution layer and a 1×1 convolution layer. The number of convolution kernels is The second part consists of a 1×1 convolution layer, followed by a 3×3 convolution layer and two 1×1 convolution layers, and finally a 3×3 convolution layer and a 1×1 convolution layer. The number of convolution kernels is The third part is composed of a 1×1 convolution layer, followed by a 3×3 convolution layer and two 1×1 convolution layers, and finally a 3×3 convolution layer and a 1×1 convolution layer. The number of convolution kernels is The fourth part consists of a 1×1 convolution layer, followed by a 3×3 convolution layer and two 1×1 convolution layers, and finally a 3×3 convolution layer and a 1×1 convolution layer. The number of convolution kernels is The feature outputs of the above two branches are fused and then input into two branches. The first branch is a 7×7 average pooling layer with a number of neurons. The second branch is a 4×4 average pooling layer with the number of neurons being The outputs of the two fully connected layers are then concatenated to obtain a dimension of The second eigenvector output.

[0068] S108: Perform feature fusion on the first feature vector and the second feature vector, input them into two pooling layers and two fully connected layers respectively, and then concatenate the outputs of the two fully connected layers to obtain a third feature vector.

[0069] Specifically, according to the second eigenvector and the first eigenvector Construct the third eigenvector (tunnel void state eigenvector) , the dimension is ; The dimension of the tunnel void state feature vector is ; Define the expected output of the heterogeneous deep feature network as , the tunnel void state feature vector Input SCN classifier to get heterogeneous deep feature network output , the network output With the expected output Subtract and get the network output error ;in, Output the number of categories for the heterogeneous deep feature network.

[0070] By adopting a dual-branch network architecture to obtain multi-resolution feature maps of the radar spectrum image of the tunnel cavity state, the adaptability of the time domain abstract semantic feature space to images of different resolutions is enhanced. Multi-scale features are fused based on the dual-scale network architecture, which enhances the distinguishability of the feature space for multi-category lining cavity images.

[0071] S110: Input the third feature vector into the SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, save the current heterogeneous deep feature network model.

[0072] The number of nodes of the SCN classifier in this example is , and a random configuration network is used to replace the softmax layer in the traditional convolutional neural network, realizing the universal approximation of the multi-scale features of the tunnel cavity state radar spectrum image to the lining cavity, enhancing the learning and generalization capabilities of the network. Before the third feature vector is input into the SCN classifier, it needs to be judged. Assign to After that, judge Is it established? If it is established, continue to the next step, otherwise return to step S102; wherein, Is the training set The number of samples included.

[0073] When building a heterogeneous deep feature network, pre-define the artificial error threshold and the maximum number of iterations , the number of initialization iterations is ;definition Heterogeneous deep feature network The root mean square error of the network output of the iteration is used to determine or Is it true? If so, save the current heterogeneous deep feature network model. , Indicates the number of samples contained in the first image information.

[0074] A possible implementation ,when or When it is not true, the weights and biases of the heterogeneous deep feature network from the fully connected layer to the input layer are updated according to the back-propagation algorithm of gradient descent, and the back-propagation process from the output layer of the heterogeneous deep feature network to the fully connected layer is calculated using formulas (1-1)-(1-3):

[0075] (1-1)

[0076] (1-2)

[0077] (1-3)

[0078] in, Heterogeneous deep feature network The root mean square error of the network output of the iteration; represents the network's expected output set, , is the expected output of the heterogeneous deep feature network, Heterogeneous deep feature network The actual output of the iteration network; Indicates network The feature set of the fully connected layer of the network at the time of the update, , Indicates the first image information The image is input into the heterogeneous deep feature network. The third eigenvector obtained by the iterative network; Represents the SCN classifier The weight of each node to the output layer; Represents the fully connected layer to the SCN classifier The weight of each node; Represents the fully connected layer to the SCN classifier The bias of each node; Represents the SCN classifier The activation function of each node; Indicates the number of nodes of the SCN classifier; Represents the number of samples contained in the first image information. It should be noted that the above formula in this example is re-derived by the applicant based on the heterogeneous neural network built by himself. Updating the weights and biases of the heterogeneous deep feature network from the fully connected layer to the input layer through the above formula will make the trained tunnel cavity state radar spectrum image recognition model more accurate.

[0079] A method for recognizing tunnel cavity state radar spectrum images, comprising:

[0080] S202: Acquire an image to be detected.

[0081] In this example, the test set T4 in step S104 may be selected to detect the performance of the heterogeneous deep feature network, or a radar spectrum image of the tunnel cavity to be detected may be input.

[0082] S204: Input the image to be detected into the tunnel void state radar spectrum image recognition model constructed by the tunnel void state radar spectrum image recognition model construction method as described above, determine the feature image of the image through the convolutional neural network model, and determine the recognition result of the image according to the feature image. , the effect diagram of detection based on heterogeneous deep feature network model is as follows Figure 5 As shown in the figure, the detection accuracy reaches 0.873. Therefore, the detection effect of the tunnel void state radar spectrum image recognition model is significantly improved compared with the YOLO model detection.

[0083] Exemplary Devices

[0084] like Figure 7 As shown, this example provides a tunnel cavity state radar spectrum image recognition model construction device based on the exemplary method, including

[0085] The training data acquisition unit 10 is used to acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity;

[0086] The feature extraction unit 20 is used to perform a two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, perform grayscale value statistics on the pixels of the second image information, and sequentially splice them to form a first feature vector; specifically, locate the position of at least one feature in the first image information; the at least one feature is determined by a tunnel cavity state detection model obtained by training the YOLO model; crop an image containing at least one feature in the first image, perform a two-dimensional Fourier transform and grayscale processing on the image to obtain second image information, perform grayscale value statistics on the pixels of the second image information, and sequentially splice them to form a first feature vector.

[0087] The convolution unit 30 is used to input the first image information into the heterogeneous deep feature network for convolution, and output the obtained second feature vector; in this example, the second image information is input into the heterogeneous deep feature network, and the size of 7×7 is obtained after the convolution operation, and the number of convolution kernels is feature space; then input into two network branches, one of which is a 7×7 average pooling layer and the number of neurons is The other two branches are 4×4 average pooling layers and the number of neurons is The outputs of the two fully connected layers are then concatenated to obtain a dimension of The second eigenvector of .

[0088] The feature fusion unit 40 is used to perform feature fusion on the first feature vector and the second feature vector to obtain a third feature vector.

[0089] The model building unit 50 is used to input the third feature vector into the SCN classifier to obtain the heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, the current heterogeneous deep feature network model is saved. When constructing a heterogeneous deep feature network, define an artificially set error threshold and the maximum number of iterations , the number of initialization iterations is ;definition Heterogeneous deep feature network The root mean square error of the network output of the iteration is used to determine or Is it true? If so, save the current heterogeneous deep feature network model. , Indicates the number of samples contained in the first image information.

[0090] when or When it is not true, the weights and biases of the heterogeneous deep feature network from the fully connected layer to the input layer are updated according to the back-propagation algorithm of gradient descent, and the back-propagation process from the output layer of the heterogeneous deep feature network to the fully connected layer is calculated using formulas (1-1)-(1-3):

[0091] (1-1)

[0092] (1-2)

[0093] (1-3)

[0094] in, Heterogeneous deep feature network The root mean square error of the network output of the iteration; represents the network's expected output set, , is the expected output of the heterogeneous deep feature network, Heterogeneous deep feature network The actual output of the iteration network; Indicates network The feature set of the fully connected layer of the network at the time of the update, , Indicates the first image information The image is input into the heterogeneous deep feature network. The third eigenvector obtained by the iterative network; Represents the SCN classifier The weight of each node to the output layer; Represents the fully connected layer to the SCN classifier The weight of each node; Represents the fully connected layer to the SCN classifier The bias of each node; Represents the SCN classifier The activation function of each node; Indicates the number of nodes of the SCN classifier; Indicates the number of samples contained in the first image information.

[0095] Exemplary computer program products and computer-readable storage media

[0096] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the tunnel void state radar spectrum image recognition method or the tunnel void state radar spectrum image recognition model construction method according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0097] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0098] In addition, an embodiment of the present application may also be a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the steps of the driving behavior decision-making method according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0099] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0100] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0101] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0102] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0103] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0104] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for constructing a radar spectrum image recognition model for tunnel void state, characterized in that: include: S102: Acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity; S104: performing a two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, performing grayscale value statistics on pixel points of the second image information, and sequentially splicing them to form a first feature vector; S106: Inputting the first image information into a heterogeneous deep feature network for convolution, and outputting a second feature vector obtained; S108: performing feature fusion and splicing on the first feature vector and the second feature vector to obtain a third feature vector; Step S108 specifically includes: performing feature fusion on the first feature vector and the second feature vector, inputting them into two pooling layers and two fully connected layers in sequence, and then concatenating the outputs of the two fully connected layers to obtain a third feature vector; S110: Input the third feature vector into the SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, save the current heterogeneous deep feature network model.

2. The method for constructing a tunnel void state radar spectrum image recognition model according to claim 1, characterized in that: The step S110 also includes: when constructing a heterogeneous deep feature network, defining an artificially set error threshold error and a maximum number of iterations μ max ,The number of initialization iterations is μ=1; Define e μ as the root mean square error of the output of the μ-th iteration network of the heterogeneous depth feature network, and judge whether e μ <error or μ>μ max holds. If it holds, save the current heterogeneous depth feature network model, where N represents the number of samples contained in the first image information, e i,μ Represents the network output error.

3. The method for constructing a tunnel void state radar spectrum image recognition model according to claim 2, characterized in that: In the step S110, it further includes: when e μ <error or μ>μ max does not hold, update the weights and biases of the heterogeneous depth feature network from the fully connected layer to the input layer according to the backpropagation algorithm of gradient descent, and calculate the backpropagation process from the output layer to the fully connected layer of the heterogeneous depth feature network using formulas (1-1)-(1-3): Among them, e μ is the root mean square error of the μth iteration network output of the heterogeneous deep feature network; T represents the expected output set of the network, T = {t1, t2, ...t N }, t i is the expected output of the heterogeneous deep feature network, t i,μ ' is the actual output of the μth iteration network of the heterogeneous deep feature network; F μ represents the feature set of the fully connected layer of the network when the network is updated for the μth time, F μ ={F 1,μ ,.F 2,μ ,...F N,μ }, F i,μ represents the third feature vector obtained by inputting the i-th image in the first image information into the heterogeneous deep feature network at the μ-th iteration; β j Represents the weight from the jth node to the output layer of the SCN classifier; represents the weight from the fully connected layer to the jth node of the SCN classifier; b j represents the bias from the fully connected layer to the jth node of the SCN classifier; k j represents the activation function of the jth node of the SCN classifier; L represents the number of nodes of the SCN classifier; N represents the number of samples contained in the first image information.

4. The method for constructing a tunnel void state radar spectrum image recognition model according to claim 1, characterized in that: The step S104 further includes: Locating the position of at least one feature in the first image information; the at least one feature is determined by a tunnel cavity state detection model obtained by training a YOLO model; An image containing at least one feature in the first image is cropped, and two-dimensional Fourier transform and grayscale processing are performed to obtain second image information.

5. The method for constructing a tunnel void state radar spectrum image recognition model according to claim 4, characterized in that: The step S106 further includes: The second image information is input into a heterogeneous deep feature network, and after a convolution operation, a feature space with a size of 7×7 and 8M convolution kernels is obtained; then it is input into two network branches, one of which is connected by a 7×7 average pooling layer and a fully connected layer with 8M neurons, and the other two branches are connected by a 4×4 average pooling layer and a fully connected layer with 32M neurons. The outputs of the two fully connected layers are then concatenated to obtain a second feature vector with a dimension of 40M.

6. A method for recognizing tunnel cavity status radar spectrum images, characterized in that: include: Acquire an image to be detected; The image to be detected is input into the tunnel void state radar spectrum image recognition model constructed by the tunnel void state radar spectrum image recognition model construction method as described in any one of claims 1 to 5, the feature image of the image is determined by the convolutional neural network model, and the recognition result of the image is determined according to the feature image.

7. A device for constructing a radar spectrum image recognition model for tunnel cavity status, characterized in that: include: A training data acquisition unit, configured to acquire first image information, wherein the first image information includes a radar spectrum image without a tunnel cavity and a radar spectrum image with a tunnel cavity; A feature extraction unit, configured to perform a two-dimensional Fourier transform and grayscale processing on the first image information to obtain second image information, perform grayscale value statistics on pixel points of the second image information, and sequentially concatenate the pixels to form a first feature vector; A convolution unit, used for inputting the first image information into a heterogeneous deep feature network for convolution, and outputting a second feature vector obtained; A feature fusion unit is used to perform feature fusion and splicing on the first feature vector and the second feature vector to obtain a third feature vector; specifically, the first feature vector and the second feature vector are subjected to feature fusion, and are respectively input into two pooling layers and two fully connected layers in sequence, and then the outputs of the two fully connected layers are spliced ​​to obtain the third feature vector; And a model building unit, used to input the third feature vector into the SCN classifier to obtain a heterogeneous deep feature network output; when the root mean square error of the iterative network output of the heterogeneous deep feature network meets the threshold, save the current heterogeneous deep feature network model.

8. A storage medium that non-transitorily stores computer-readable instructions, wherein: When the non-temporary computer-readable instructions are executed by a computer, instructions for constructing a tunnel void state radar spectrum image recognition model according to any one of claims 1 to 5 may be executed, or instructions for recognizing a tunnel void state radar spectrum image according to claim 6 may be executed.

Images