Road recessive disease identification method and system based on three-dimensional radar map

Abstract

The invention discloses a road recessive disease identification method and system based on a three-dimensional radar map. The method comprises the steps that an initial improved model is constructed based on YOLOv8, a target CutSimAM attention mechanism is added to the tail end of a head network, a feature extraction module is C2fEFAttention, a loss function is a SlideLoss sliding weighting function, and a bounding box loss function is a D-IoU bounding box loss function; acquiring a radar disease map transmitted by the three-dimensional ground penetrating radar, and preprocessing the radar disease map to obtain a data set for training; training the initial improved model based on the training data set to obtain a target improved model; and inputting a to-be-detected radar image of a to-be-detected road into the target improved model, and identifying recessive disease information of the to-be-detected road. According to the method, the YOLOv8 model is improved through the CutSimAM attention mechanism, the C2fEFAttention, the SlideLoss sliding weighting function and the D-IoU bounding box loss function, and the lightweight model is obtained and applied to road hidden disease recognition, so that high-precision and high-efficiency real-time detection and recognition are realized.

Description

Technical Field

[0001] The present application relates to the field of radar monitoring technology, and in particular to a method and system for identifying hidden road defects using three-dimensional radar images. Background Art

[0002] At present, my country's expressway network has become increasingly complete. With the continuous increase in service life, under the influence of heavy loads and complex environments, asphalt pavement structures have gradually been damaged, such as structural cracks, poor interlayer quality, looseness and other defects. Transportation infrastructure has gradually entered the stage of large-scale maintenance. With the development of new non-destructive testing technologies, ground penetrating radar provides a solution for the refined identification of internal road surface defects. The A-scan and B-scan data collected by two-dimensional ground penetrating radar determine the structural morphology and spatial position of the underground medium through radio wave waveform, amplitude intensity and other characteristics, while the three-dimensional ground penetrating radar (3D Ground Penetrating Radar) uses a multi-channel three-dimensional antenna array to collect C-scan data, combining time domain and frequency domain characteristic parameters, and can generate stereo images of road transverse and longitudinal sections and horizontal sections. The technology has achieved a transformation from "point to surface", greatly improving the detection efficiency and detection accuracy, and can achieve the purpose of rapid diagnosis and evaluation of defects. Compared with traditional methods such as core sampling and ultrasonic testing [4], fast, efficient and non-destructive three-dimensional ground penetrating radar detection technology has become one of the most widely used technical means for road disease detection.

[0003] With the rapid development of deep learning technology, applying digital and intelligent technologies to three-dimensional ground penetrating radar data analysis will help improve radar detection efficiency and achieve real-time detection of hidden road defects.

[0004] At present, the application of deep learning technology in the research process of road hidden disease identification still faces many problems, such as the lack of real ground penetrating radar map data, different models of ground penetrating radar have different hardware equipment parameters and software interpretation methods, which makes it difficult to match radar disease map data; the ground penetrating radar signal characteristics are complex, and there are interference information such as aliasing waves, noise, and clutter. When manually identifying and labeling, there are differences in the understanding of the morphology of road hidden diseases, which may lead to misjudgment, and when faced with large-scale data, manual interpretation is inefficient; there is a lack of targeted detection methods, data processing and image analysis technology. The recognition accuracy of conventional single-stage algorithms is low, the recognition efficiency of two-stage algorithms is low, the model size is large, and it is difficult to embed equipment for real-time detection.

[0005] Therefore, how to achieve high-precision real-time detection of hidden road defects using three-dimensional ground penetrating radar is an urgent problem to be solved. Summary of the invention

[0006] In order to solve the above problems, the embodiments of the present application provide a method and system for identifying hidden road defects based on three-dimensional radar images.

[0007] In the first aspect, in order to solve the above technical problems, the present application provides a method for identifying hidden road defects using a three-dimensional radar map, comprising:

[0008] An initial improved model is constructed based on the YOLOv8 model, wherein a target Cut_SimAM attention mechanism is added to the end of the head network of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function;

[0009] Acquire a radar damage map transmitted by a three-dimensional ground penetrating radar, and preprocess the radar damage map to obtain a training data set;

[0010] Training the initial improved model based on the training data set to obtain a target improved model;

[0011] The radar image of the road to be tested is input into the target improvement model to identify the hidden disease information of the road to be tested.

[0012] The beneficial effects are:

[0013] In the technical solution provided in the embodiment of the present application, on the basis of the baseline model YOLOv8 model, an initial improved model is constructed, and the target Cut_SimAM attention mechanism is added to the head network end of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function. The target Cut_SimAM attention mechanism can make the edge of the feature information clearer, thereby focusing on the defect area in the image, and the C2f_EFAttention and SlideLoss sliding weighted functions can improve the detection speed of the model, so the initial well model constructed can improve the detection accuracy and operation efficiency of the model. After that, the radar defect map transmitted in real time by the three-dimensional ground penetrating radar is obtained, and the radar defect map is preprocessed to obtain a training data set. The initial improved model is trained based on the training data set to obtain the target improved model, that is, a lightweight model for the recognition of hidden road defects in the three-dimensional radar map is obtained, so as to achieve high-precision and high-efficiency real-time detection and recognition of hidden road defects.

[0014] In a second aspect, the present invention provides a road hidden disease recognition system using three-dimensional radar maps, including a model building unit, an acquisition unit, a training unit, and a recognition unit;

[0015] A model building unit, used to build an initial improved model based on the YOLOv8 model, wherein a target Cut_SimAM attention mechanism is added to the end of the head network of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function;

[0016] An acquisition unit, used for acquiring a radar defect map transmitted by a three-dimensional ground penetrating radar, and preprocessing the radar defect map to obtain a training data set;

[0017] A training unit, used for training the initial improved model based on the training data set to obtain a target improved model;

[0018] The identification unit is used to input the radar image of the road to be tested into the target improvement model to identify the hidden disease information of the road to be tested.

[0019] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and together with the specification, are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0021] Figure 1 It is a flow chart of a method for identifying hidden road defects using three-dimensional radar images, shown in an exemplary embodiment of the present application;

[0022] Figure 2 is a structural schematic diagram of an initial improved model in an exemplary embodiment of the present application;

[0023] Figure 3 is a schematic diagram of a target Cut_SimAM attention mechanism in an exemplary embodiment of the present application;

[0024] Figure 4 is a schematic diagram of C2f_EFAttention in an exemplary embodiment of the present application;

[0025] Figure 5 It is a comparative diagram of the output layer heat map of the verification model;

[0026] Figure 6 It is a comparative schematic diagram of the detection effect diagram of the verification model;

[0027] Figure 7 It is a block diagram of a road hidden disease identification system of a three-dimensional radar map shown in an exemplary embodiment of the present application;

[0028] Figure 8 It is a structural diagram of a computer system suitable for implementing an electronic device of an embodiment of the present application. DETAILED DESCRIPTION

[0029] Here, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims.

[0030] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0031] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0032] The term "multiple" as used in this application refers to two or more than two. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.

[0033] In order to solve the problems faced by the existing application of deep learning technology in the identification of hidden road defects, such as low recognition accuracy, low efficiency, large model size, and difficulty in embedding in equipment to achieve real-time detection, the embodiments of the present application propose a method and device for identifying hidden road defects using three-dimensional radar maps, an electronic device, and a computer-readable storage medium, which mainly involve the technology for identifying hidden road defects using three-dimensional radar maps included in road radar monitoring technology. These embodiments will be described in detail below.

[0034] First see Figure 1 , Figure 1This is a flow chart of a method for identifying hidden road defects using three-dimensional radar maps, shown in an exemplary embodiment of the present application. The method can be specifically executed by a server, which can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, and is not limited here.

[0035] like Figure 1 As shown, in an exemplary embodiment, the method for identifying hidden road defects using a three-dimensional radar map may include steps S101 to S104, which are described in detail as follows:

[0036] Step S101, an initial improved model is constructed based on the YOLOv8 model, a target Cut_SimAM attention mechanism is added to the end of the head network of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function.

[0037] The initial improved model constructed in this application adds an improved target Cut_SimAM attention mechanism to the head network, uses C2f_EFAttention in the backbone network to replace the C2f feature extraction module in the original network, uses SlideLoss sliding weighted function to replace the BCE loss function in the original network, and uses D-IoU bounding box loss function to replace the C-IoU bounding box loss function in the original network. Figure 2 As shown, Figure 2 It is a structural diagram of an initial improved model in an exemplary embodiment of the present application.

[0038] Step S102, obtaining a radar defect map transmitted by a three-dimensional ground penetrating radar, and preprocessing the radar defect map to obtain a training data set.

[0039] In this embodiment, the radar damage map can be obtained by using a vehicle-mounted three-dimensional array radar to detect municipal roads, and the radar damage map is preprocessed to obtain a training data set.

[0040] Step S103, training the initial improved model based on the training data set to obtain a target improved model.

[0041] After completing the construction of the initial improved model and obtaining the training data set, the initial improved model is trained based on the training data set to obtain a target improved model that is compatible with the identification of hidden road defects, so that it can be applied to the real-time detection of hidden road defects.

[0042] Step S104: inputting the radar image of the road to be tested into the target improvement model to identify the hidden disease information of the road to be tested.

[0043] In the actual application of the target improvement model, the radar image of the road to be tested is acquired in real time through the three-dimensional ground penetrating radar, and the image is input into the obtained target improvement model, so that the target improvement model is used to identify the hidden road defects in the radar image to detect the hidden defect information of the road to be tested.

[0044] As can be seen from the above, in the method provided in this embodiment, on the one hand, by constructing an initial improved model based on the baseline model YOLOv8 model, the target Cut_SimAM attention mechanism is added to the head network end of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function. The target Cut_SimAM attention mechanism can make the edge of the feature information clearer, thereby focusing on the diseased area in the image, and the C2f_EFAttention and SlideLoss sliding weighted functions can improve the detection speed of the model, so the constructed initial well model can improve the detection accuracy and computational efficiency of the model. On the other hand, after the initial improved model is constructed, the radar damage map transmitted in real time by the three-dimensional ground penetrating radar is obtained, and the radar damage map is preprocessed to obtain a training data set. The initial improved model is trained based on the training data set to obtain a target improved model, that is, a lightweight model for identifying hidden road defects in three-dimensional radar maps, so as to achieve high-precision and high-efficiency real-time detection and identification of hidden road defects.

[0045] In an exemplary embodiment provided by the present application, the SlideLoss sliding weighted function replaces the BCE (Binary Cross-Entropy Loss) loss function in the original network. The basic formula of the SlideLoss sliding weighted function is:

[0046]

[0047] The threshold μ is the average value of the intersection-over-union ratio of all sample bounding boxes. Samples larger than μ are positive samples, and samples smaller than μ are negative samples.

[0048] In an exemplary embodiment provided by the present application, the specific steps of preprocessing the radar disease atlas to obtain a training data set include:

[0049] The radar damage map is filtered and the filtered radar image is labeled and classified according to the damage type to obtain the initial data set; the damage types include poor interlayer, interlayer moisture, interlayer looseness, and loose structure;

[0050] The initial data set is enhanced to obtain a training data set.

[0051] In this embodiment, the Swedish ImpulseRadar vehicle-mounted three-dimensional array radar Raptor-80 can be used to collect radar damage maps, and then the radar damage maps are subjected to bandpass filtering, background removal, and spectrum analysis methods to remove clutter and noise. The parameters of the radar equipment are shown in Table 1.

[0052] Table 1: Parameters of Raptor-80 vehicle-mounted radar

[0053] Technical indicators Antenna center frequency Number of channels Sampling interval Time Window Dwell time Detection speed Technical Parameters 800MHz 24 2.5~5cm 25ns~75ns 3μs 30km / h

[0054] After completing the filtering process of the radar damage map, it is labeled and classified according to the damage type. For example, the labeling tool LabelMe is used to label the radar damage map after filtering, generate XML labeling information, and obtain the initial data set based on the radar damage map and the labeling information. The initial data set is further enhanced to obtain the training data set. Among them, the damage types include poor interlayer, interlayer water, loose interlayer, and loose structure, which correspond to the labeling information poor_l, water_l, loose_l, and loose_s respectively.

[0055] In another exemplary embodiment, the data augmentation processing includes blurring noise, sharpening processing, random brightness and contrast, random inversion, and image stitching, so the corresponding specific steps are: blurring noise, sharpening processing, random brightness and contrast, random inversion, and image stitching are performed on the initial data set in sequence to obtain a training data set.

[0056] The specific steps of blurring noise are as follows: blurring the images in the initial dataset using a generalized normal filter with randomly selected parameters. This transformation also adds multiplicative noise to the generated kernel before convolution, combining blurring and noise injection to enhance the data.

[0057] The specific steps of sharpening are: sharpen the initial data set and superimpose it with the original image, with the degree of sharpening being random within a specified range. This enhances the edges and details of the image, making the overall image clearer.

[0058] The specific steps of random brightness and contrast are: randomly increase or decrease the brightness and contrast of the initial data set within a specified range.

[0059] The specific steps of random flipping are: randomly flipping the initial data set, including left-right flipping and up-down flipping.

[0060] The specific steps of image stitching are: randomly select four different images from the initial data set, randomly crop and scale the selected images to adapt them to the size of the new stitched image. Stitch the four processed images into a new image with a size of 640×640. According to the stitched image, correct the coordinate information of the target box to correspond to the new image coordinate system.

[0061] In an exemplary embodiment provided by the present application, the specific steps of training the initial improved model to obtain the target improved model include:

[0062] Divide the training data set based on a preset division ratio to obtain corresponding training set, validation set, and test set;

[0063] Based on the training set, validation set and test set, the initial improved model is trained according to the preset training batches and training rounds to obtain the target improved model; wherein, the division ratio is 4:1:1, the training batches are 16, and the training rounds are 300.

[0064] In this embodiment, the division ratio for dividing the training data set is 4:1:1, and the basic hyperparameters of the training improvement model are set as follows: the training batch is 16 and the training round is 300. For example, the radar damage atlas collected by the three-dimensional ground penetrating radar includes 2451 radar damage atlases, of which 1600 are divided as training sets, 400 as validation sets, and 451 as test sets; the basic hyperparameters of the training improvement model are set as follows: the training batch is 16, the training round is 300, the optimizer selects AdamW, and the other hyperparameters are default.

[0065] In the embodiments provided in this application, the experimental environment configuration for training the initial improved model can be seen in Table 2.

[0066] Table 2: Experimental environment configuration

[0067]

[0068]

[0069] In an exemplary embodiment provided by the present application, in the process of training the initial improved model based on the training data set to obtain the target improved model, the specific steps of the target Cut_SimAM attention mechanism for data processing of the training data include:

[0070] Using the target Cut_SimAM attention mechanism, the input feature map of the input target Cut_SimAM attention mechanism is cut into four cut blocks;

[0071] Perform mean subtraction and normalization processing on each clipped image block, and activate the processed clipped image block through the Sigmoid function to obtain an activated image block;

[0072] Obtain the first characteristic matrix of the cut block and the second characteristic matrix of the activated block, perform a direct product of the second characteristic matrix and the second characteristic matrix to obtain the target block;

[0073] The features of the four target blocks are concatenated to obtain the output feature map of the head network in the target improvement model.

[0074] In this embodiment, the target Cut_SimAM attention mechanism is an improved Cut_SimAM attention mechanism for road hidden disease recognition, see Figure 3 , Figure 3 is a schematic diagram of the target Cut_SimAM attention mechanism in an exemplary embodiment of the present application. Figure 3 As shown in the figure, in the model, the improved Cut_SimAM attention mechanism, for the input feature map of the input attention mechanism, first cuts the input feature map into four cut blocks, then performs mean subtraction and normalization on each cut block, and activates the processed cut blocks through the Sigmoid function to obtain activated blocks; obtains the first feature matrix of the cut blocks and the second feature matrix of the activated blocks, and directly multiplies the second feature matrix with the second feature matrix to obtain the target block; features of the four target blocks are spliced ​​to obtain the output feature map of the head network in the target improved model.

[0075] It should also be noted that when the target improvement model performs detection and recognition, the target Cut_SimAM attention mechanism also follows the above data processing process.

[0076] In this way, the present application adds an improved Cut_SimAM attention mechanism to the head network through the above embodiments, so that the improved model is more focused on the identification of useful features, that is, the accuracy of the model is improved.

[0077] In an exemplary embodiment provided by the present application, C2f_EFAttention includes an EFAttention module; in the process of training the initial improved model based on the training data set to obtain the target improved model, the specific steps of data processing of the training data by C2f_EFAttention include:

[0078] Using C2f_EFAttention, perform a convolution on the input of C2f_EFAttention to obtain an intermediate feature map;

[0079] Split the intermediate feature map to obtain the first part of the feature map and the second part of the feature map;

[0080] The second part of the feature map is enhanced through the EFAttention module, and the processed second part of the feature map is concatenated with the first part of the feature map to obtain the output of the C2f_EFAttention module.

[0081] In this embodiment, C2f_EFAttention is used in the backbone network to replace the C2f feature extraction module in the original network. Figure 4 , Figure 4 FIG. 1 is a schematic diagram of C2f_EFAttention in an exemplary embodiment of the present application. Figure 4 As shown in the figure, in the C2f_EFAttention model, the input data is convolved once to obtain an intermediate feature map, and the intermediate feature map is split to obtain the first feature map and the second feature map. Then, the second feature map is enhanced by the EFAttention module, and the processed second feature map is concatenated with the first feature map to obtain the output of the C2f_EFAttention module.

[0082] Among them, the EFAttention module reduces the feature map to a channel descriptor through global average pooling (GAP) at the channel layer, then processes the channel information through 1D convolution, outputs the attention weight through the Sigmoid function, and weights the input feature map; at the spatial layer, 1x1 convolution is applied to calculate the spatial weight, then the attention weight is output through the Sigmoid function, and the input feature map is weighted. Finally, the weighted results of the channel layer and the spatial layer are summed and output to enhance useful features.

[0083] It should also be noted that when the target improvement model is used for detection and recognition, C2f_EFAttention also follows the above data processing process.

[0084] In this way, through the above embodiments, the present application adopts C2f_EFAttention in the backbone network to replace the C2f feature extraction module in the original network, and applies the characteristics of C2f_EFAttention specifically to the data processing process of the feature map to construct an improved model to improve the detection speed of the model.

[0085] In an exemplary embodiment provided by the present application, after the target improvement model is obtained through training, its model performance needs to be tested so that the target improvement model can achieve the desired detection accuracy and efficiency. The specific steps include:

[0086] Obtaining evaluation indicators of the target improvement model, including detection accuracy evaluation indicators, detection efficiency evaluation indicators and lightweight evaluation indicators;

[0087] The model performance of the target improved model is obtained based on the detection accuracy evaluation index, detection efficiency evaluation index and lightweight evaluation index.

[0088] In another exemplary embodiment, the detection efficiency evaluation index is the number of detection frames per second (FPS, Frames PerSecond), the lightweight evaluation index is the number of parameters (Params, which refers to the total number of parameters that need to be trained in the network model) and the amount of calculation (FLOPs, floating point operations, meaning floating point operations) of the target improvement model; the detection accuracy evaluation index includes the recall rate R e (Recall), precision P r (Precision) and multi-category average precision (mAP), the expressions of precision, recall and multi-category average precision are:

[0089]

[0090] Among them, TP is a true positive example, FP is a false positive example, FN is a false negative example, and C is the number of disease types, C=4.

[0091] In this way, through the above-mentioned embodiments, after training the target improvement model, the present application tests its model performance by obtaining the number of detection frames per second, parameter quantity, calculation amount, recall rate, precision rate and multi-category average accuracy, so that the target improvement model can be further optimized based on the model performance of the detected target improvement model, thereby realizing high-precision and high-efficiency real-time detection and identification of hidden road diseases.

[0092] In an exemplary embodiment provided by the present application, in order to verify the effectiveness of the target improved model and the influence of each improved module in the model, YOLOv8 was used as the baseline model, and 7 groups of verification models were designed for ablation experiments. The module application of each group of verification models is shown in Table 3 below. The model YOLOv8-Sim represents the addition of the SimAM attention mechanism at the end of the backbone layer of YOLOv8, and the model YOLOv8-C represents the use of the Cut_SimAM module with a cutting function to replace the SimAM attention mechanism in YOLOv8-SimAM. YOLOv8-E means that the C2f_EFAttention module is used to replace the C2f module in the original network structure in the neck layer. The model YOLOv8-S means that the BCE loss function used in the original network is replaced by SlideLoss. The model YOLOv8-CE represents the comprehensive use of the Cut_SimAM module and the C2f_EFAttention module. The model YOLOv8-CES represents the comprehensive use of the Cut_SimAM module, the C2f_EFAttention module, and the SlideLoss sliding loss function.

[0093] In the training phase, the baseline model YOLOv8, the YOLOv8-Sim model with the SimAM attention mechanism, and the YOLOv8-C model with the Cut_SimAM module added are compared. The maximum values ​​of the channel features of the disease atlas of the three models are used as the pixel values ​​of the heat map and displayed in the form of a heat map. The visualization results are shown in Figure 2. Figure 5 As shown, Figure 5 This is a comparative diagram of the output layer heat map of the verification model, where red indicates that the model pays high attention to this area, and blue indicates that the model pays low attention to this area.

[0094] It can be seen that compared with the baseline model, the SimAM attention mechanism re-weights the global features, enhances the model's feature extraction capability, and enhances the features of the originally unnoticed targets. Compared with the SimAM attention mechanism, the Cut_SimAM module makes the edges of feature information clearer, focuses on the diseased areas in the image, ignores the non-diseased area information, but also enhances the attention to some unlabeled areas, that is, suspected diseased areas.

[0095] During the verification process of this embodiment, 451 radar defect images of the verification set were used to perform an ablation experiment on the weight file saved in the training results. The results are shown in Table 3.

[0096] Table 3: Ablation experiments

[0097]

[0098] See also Figure 6 , Figure 6It is a comparative diagram of the detection effect diagram of the verification model. Figure 6 As shown in Table 3, the YOLOv8-Sim model with the SimAM attention mechanism has improved recognition accuracy, and mAP has increased by 1.3% compared to the YOLOv8na baseline model, but the FPS has decreased by about 15, which indicates that the processing speed of the model has increased by 8%. The YOLOV8-C model with the Cut_SimAM attention mechanism has no improvement in processing speed, but mAP has increased by 0.8% compared to YOLOv8-Sim. The YOLOv8-C model using the C2f_EFAttention module reduces the number of parameters by 0.5M and the amount of calculation by 1GFLOPs, which improves the detection speed of the model. The YOLOv8-S model using the SlideLoss sliding weighting function greatly improves the FPS without reducing the number of parameters. This shows that both the C2f_EFAttention module and SlideLoss help improve the computational efficiency of the model.

[0099] It can be seen that the method for identifying hidden road defects in three-dimensional radar maps provided in the embodiment of the present application adds an improved target Cut_SimAM attention mechanism to the head network, uses C2f_EFAttention in the backbone network to replace the C2f feature extraction module in the original network, uses SlideLoss sliding weighted function to replace the BCE loss function in the original network, and uses D-IoU bounding box loss function to replace the C-IoU bounding box loss function in the original network. After training, a target improved model YOLOv8-CES model is constructed. The YOLOv8-CES model combines the above-mentioned Cut_SimAM attention mechanism, C2f_EFAttention and SlideLoss sliding weighted function. The three optimization methods achieve the highest recognition accuracy, and its mAP is improved by 3.6% compared with the baseline model, and the processing speed is also increased by 16.7%. Among them, the C2f_EFAttention module reduces the number of parameters and the number of floating-point operations per second. The lightweight model is more advantageous for vehicle-mounted ground penetrating radars of embedded devices.

[0100] Figure 7 FIG. 7 is a block diagram of a road hidden disease identification system 700 of a three-dimensional radar map according to an exemplary embodiment of the present application. Figure 7 As shown, the system includes:

[0101] A model construction unit 701 is used to obtain an initial improved model based on the YOLOv8 model. The target Cut_SimAM attention mechanism is added to the head network end of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function;

[0102] An acquisition unit 702 is used to acquire a radar defect map transmitted by a three-dimensional ground penetrating radar, and pre-process the radar defect map to obtain a training data set;

[0103] A training unit 703 is used to train the initial improved model based on the training data set to obtain a target improved model;

[0104] The identification unit 704 is used to input the radar image of the road to be tested into the target improvement model to identify the hidden disease information of the road to be tested.

[0105] The system applies the road hidden disease identification method of the three-dimensional radar map provided by the present application, and constructs an initial improved model based on the baseline model YOLOv8 model through the model construction unit 701. The target Cut_SimAM attention mechanism is added to the head network end of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function. Then the acquisition unit 702 acquires the radar disease map transmitted in real time by the three-dimensional ground penetrating radar, and pre-processes the radar disease map to obtain a training data set. The training unit 703 trains the initial improved model based on the training data set to obtain a target improved model, that is, a lightweight model for road hidden disease identification of the three-dimensional radar map is obtained, so that high-precision and high-efficiency real-time detection and identification of road hidden diseases can be achieved through the recognition unit 704.

[0106] In another exemplary embodiment, the acquisition unit 702 is also used to filter the radar defect map, and label and classify the filtered radar image according to the defect type to obtain an initial data set; the defect types include poor interlayer, interlayer moisture, loose interlayer, and loose structure; the initial data set is data enhanced to obtain a training data set.

[0107] In another exemplary embodiment, the acquisition unit 702 is further used to perform blurring, sharpening, random brightness and contrast, random inversion, and image stitching on the initial data set in sequence to obtain a training data set.

[0108] In another exemplary embodiment, the training unit 703 is also used to divide the training data set based on a preset division ratio to obtain corresponding training sets, validation sets and test sets; based on the training set, validation set and test set, the initial improved model is trained according to preset training batches and training rounds to obtain a target improved model; wherein the division ratio is 4:1:1, the training batches are 16, and the training rounds are 300.

[0109] In another exemplary embodiment, the training unit 703 is also used to use the target Cut_SimAM attention mechanism to cut the input feature map of the input target Cut_SimAM attention mechanism into four cut blocks; perform mean subtraction and normalization processing on each cut block, and activate the processed cut block through the Sigmoid function to obtain an activated block; obtain the first feature matrix of the cut block and the second feature matrix of the activated block, and perform the direct product of the second feature matrix and the second feature matrix to obtain the target block; perform feature splicing on the four target blocks to obtain the output feature map of the head network in the target improved model.

[0110] In another exemplary embodiment, C2f_EFAttention includes an EFAttention module; the training unit 703 is further used to use C2f_EFAttention to perform a convolution on the input of C2f_EFAttention to obtain an intermediate feature map; split the intermediate feature map to obtain a first part feature map and a second part feature map; perform feature enhancement processing on the second part feature map through the EFAttention module, and feature concatenate the processed second part feature map and the first part feature map to obtain the output of the C2f_EFAttention module.

[0111] In another exemplary embodiment, the system further comprises:

[0112] An evaluation unit is used to obtain evaluation indicators of the target improvement model, the evaluation indicators include detection accuracy evaluation indicators, detection efficiency evaluation indicators and lightweight evaluation indicators; based on the detection accuracy evaluation indicators, the detection efficiency evaluation indicators and lightweight evaluation indicators, the model performance of the target improvement model is obtained.

[0113] It should be noted that the road hidden disease identification system of the three-dimensional radar map provided in the above embodiment and the road hidden disease identification method of the three-dimensional radar map provided in the above embodiment belong to the same concept, wherein the specific manner in which each module and unit performs the operation has been described in detail in the method embodiment, and will not be repeated here. In actual application, the road hidden disease identification device of the three-dimensional radar map provided in the above embodiment can distribute the above functions to different functional modules as needed, that is, divide the internal structure of the device into different functional modules to complete all or part of the functions described above, and this is not limited here.

[0114] An embodiment of the present application also provides an electronic device, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by one or more processors, the electronic device implements the method for identifying hidden road defects using three-dimensional radar images provided in the above-mentioned embodiments.

[0115] Figure 8 The structure diagram of the computer system suitable for implementing the electronic device of the embodiment of the present application is shown. It should be noted that: Figure 8 The computer system 800 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0116] like Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 802 or the program loaded from the storage part 808 to the random access memory (RAM) 803, such as executing the method in the above embodiment. In the RAM 803, various programs and data required for system operation are also stored. The CPU 801, ROM 802 and RAM 803 are connected to each other through a bus 804. The input / output (I / O) interface 805 is also connected to the bus 804.

[0117] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed so that a computer program read therefrom is installed into the storage section 808 as needed.

[0118] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication section 809, and / or installed from a removable medium 811. When the computer program is executed by a central processing unit (CPU) 801, various functions defined in the system of the present application are executed.

[0119] It should be noted that the computer-readable medium shown in the embodiment of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a 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. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. This propagated data signal can take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. A computer program contained on a computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0120] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0121] Another aspect of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for identifying hidden road defects using a three-dimensional radar map is implemented. The computer-readable storage medium may be included in the electronic device described in the above embodiment, or may exist independently without being assembled into the electronic device.

[0122] Another aspect of the present application also provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the road hidden disease identification method of the three-dimensional radar map provided in each of the above embodiments.

[0123] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for identifying hidden road defects using three-dimensional radar images, characterized in that: The method comprises: An initial improved model is constructed based on the YOLOv8 model, wherein a target Cut_SimAM attention mechanism is added to the end of the head network of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function; Acquire a radar damage map transmitted by a three-dimensional ground penetrating radar, and preprocess the radar damage map to obtain a training data set; Training the initial improved model based on the training data set to obtain a target improved model; The radar image of the road to be tested is input into the target improvement model to identify the hidden disease information of the road to be tested.

2. The method according to claim 1, characterized in that The preprocessing of the radar disease atlas to obtain a training data set includes: Filtering the radar damage map, and annotating and classifying the filtered radar image according to the damage type to obtain an initial data set; the damage types include poor interlayer, interlayer moisture, interlayer looseness, and loose structure; Perform data enhancement processing on the initial data set to obtain a training data set.

3. The method according to claim 2, characterized in that The performing data enhancement processing on the initial data set to obtain a training data set includes: The initial data set is subjected to blurring noise, sharpening processing, random brightness and contrast, random inversion and image splicing in sequence to obtain a training data set.

4. The method according to claim 1, characterized in that: The step of training the initial improved model based on the training data set to obtain a target improved model includes: Dividing the training data set based on a preset division ratio to obtain corresponding training set, validation set and test set; Based on the training set, the validation set and the test set, the initial improved model is trained according to preset training batches and training rounds to obtain a target improved model; wherein the division ratio is 4:1:1, the training batches are 16, and the training rounds are 300.

5. The method according to claim 1, characterized in that In the step of training the initial improved model based on the training data set to obtain a target improved model, the method includes: Using the target Cut_SimAM attention mechanism, cutting the input feature map input into the target Cut_SimAM attention mechanism into four cut blocks; Performing mean subtraction and normalization processing on each of the cut image blocks, and activating the processed cut image blocks through a Sigmoid function to obtain an activated image block; Acquire a first characteristic matrix of the cut block and a second characteristic matrix of the activated block, and perform a direct product of the first characteristic matrix and the second characteristic matrix to obtain a target block; The four target blocks are feature concatenated to obtain an output feature map of the head network in the target improved model.

6. The method according to claim 1, characterized in that The C2f_EFAttention includes an EFAttention module; In the step of training the initial improved model based on the training data set to obtain a target improved model, the method includes: Using the C2f_EFAttention, convolve the input of the C2f_EFAttention once to obtain an intermediate feature map; Splitting the intermediate feature map to obtain a first part feature map and a second part feature map; The second part of the feature map is subjected to feature enhancement processing by the EFAttention module, and the processed second part of the feature map and the first part of the feature map are subjected to feature splicing to obtain the output of the C2f_EFAttention module.

7. The method according to claim 1, characterized in that The expression of the SlideLoss sliding weighted function is: The threshold μ is the average value of the intersection-over-union ratio of all sample bounding boxes. Samples larger than μ are positive samples, and samples smaller than μ are negative samples.

8. The method according to claim 1, characterized in that The method further comprises: Obtaining evaluation indicators of the target improvement model, wherein the evaluation indicators include a detection accuracy evaluation indicator, a detection efficiency evaluation indicator, and a lightweight evaluation indicator; The model performance of the target improved model is obtained based on the detection accuracy evaluation index, the detection efficiency evaluation index and the lightweight evaluation index.

9. The method according to claim 8, characterized in that The detection efficiency evaluation index is per second The number of detection frames, the lightweight evaluation index is the number of parameters and the amount of calculation of the target improved model; The detection accuracy evaluation indicators include recall rate, precision rate and multi-category average precision. The expressions of precision rate, recall rate and multi-category average precision are respectively: Among them, Re is the recall rate, Pr is the precision rate, mAP is the multi-category average precision, TP is the true positive example, FP is the false positive example, FN is the false negative example, and C is the number of disease type categories, C=4.

10. A three-dimensional radar map road hidden disease recognition system, characterized in that: include: A model construction unit, used to obtain an initial improved model based on the YOLOv8 model, wherein a target Cut_SimAM attention mechanism is added to the head network end of the initial improved model, the feature extraction module is C2f_EFAttention, the loss function is SlideLoss sliding weighted function, and the bounding box loss function is D-IoU bounding box loss function; An acquisition unit, used for acquiring a radar defect map transmitted by a three-dimensional ground penetrating radar, and preprocessing the radar defect map to obtain a training data set; A training unit, used for training the initial improved model based on the training data set to obtain a target improved model; The identification unit is used to input the radar image of the road to be tested into the target improvement model to identify the hidden disease information of the road to be tested.

Images