Road surface hidden disease detection method, device and equipment and storage medium

Abstract

The invention discloses a road surface hidden disease detection method, device and equipment and a storage medium, and the method comprises the steps: firstly, adjusting a road radar image, inputting the adjusted road radar image into a Darknet-53 feature extraction network to obtain feature maps which are reduced by 4, 8, 16 and 32, carrying out the transverse connection of the feature maps which are reduced by 32 times and the feature maps which are reduced by 16 times, and obtaining a first feature map; inputting the image to a first convolution block for convolution to obtain a first scale feature map, sequentially obtaining a second scale feature map and a third scale feature map, and finally performing blind disease detection by using the first scale feature map, the second scale feature map and the third scale feature map to obtain a detection result. Through multi-scale feature extraction and feature fusion on different scales, the detection capability of the model to small targets is enhanced, and the detection precision and efficiency of small targets with different shapes in a road radar image are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, in particular to a road surface hidden disease detection method, device, equipment and storage medium. BACKGROUND

[0002] As a non-destructive testing method, ground penetrating radar (GPR) has been widely used in highway hidden disease detection and gradually become the mainstream detection method due to its high efficiency, real-time continuity, high resolution and strong anti-interference. However, the main bottleneck of current ground penetrating radar in application is the data interpretation stage.

[0003] The radar image generated after ground penetrating radar detection needs to be manually interpreted by professional personnel to identify different disease types and their positions. Common disease types include multiple internal voids in the base layer, interlayer voids, interlayer loose and loose structure, etc. However, due to the huge amount of data of radar image, manual interpretation is not only inefficient, but also the interpretation accuracy depends on the experience and technical level of the operator, which makes it difficult to guarantee the stability and consistency of the interpretation results, and the cost is relatively high. Therefore, it is of great application value to develop an efficient and stable intelligent interpretation technology.

[0004] In recent years, YOLOv3 algorithm as a one-stage target detection method has achieved remarkable success in the field of optical images, with the advantages of high detection accuracy and fast speed. However, there are significant differences between highway radar images and optical images: the aspect ratio is uneven, the hidden disease features are small and complex, which makes the YOLOv3 algorithm limited in processing such images, and cannot meet the needs of effective detection of highway radar images. SUMMARY

[0005] The purpose of the present application is to overcome the technical defects of the prior art, and provide a road surface hidden disease detection method, device, equipment and storage medium, which enhances the model's detection ability for small targets by multi-scale feature extraction and feature fusion at different scales, and improves the detection accuracy and efficiency of small and various-shaped targets in highway radar images.

[0006] The purpose of the present application is achieved by the following technical solutions: In a first aspect, the present application provides a road surface hidden disease detection method, which is applied to an improved YOLOv3 network, the improved YOLOv3 network includes a Darknet-53 feature extraction network, and the method comprises: The highway radar image is input into the Darknet-53 feature extraction network after adjustment to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times, and a feature map reduced by 32 times. The feature map reduced by 32 times is subjected to convolution upsampling and then horizontally connected with the feature map reduced by 16 times to obtain a first feature map, and the first feature map is input into a first convolution block for convolution to obtain a first scale feature map. The first feature map is subjected to convolution upsampling and then horizontally connected with the feature map reduced by 8 times to obtain a second feature map, and the second feature map is input into a second convolution block for convolution to obtain a second scale feature map. The second feature map is subjected to convolution upsampling and then horizontally connected with the feature map reduced by 4 times to obtain a third feature map, and the third feature map is input into a third convolution block for convolution to obtain a third scale feature map. The first scale feature map, the second scale feature map, and the third scale feature map are used for concealed disease detection to obtain a detection result, which includes the position and category information of the boundary box of the corresponding scale.

[0007] In a possible implementation, the Darknet-53 feature extraction network includes two convolution layers, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block, and the step of inputting the highway radar image into the Darknet-53 feature extraction network after adjustment to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times, and a feature map reduced by 32 times includes: The highway radar image is input into the two convolution layers, the first residual block, the first convolution layer, and the second residual block in sequence after adjustment to obtain a feature map reduced by 4 times. The feature map reduced by 4 times is input into the second convolution layer and the third residual block in sequence to obtain a feature map reduced by 8 times. The feature map reduced by 8 times is input into the third convolution layer and the fourth residual block in sequence to obtain a feature map reduced by 16 times. The feature map reduced by 16 times is input into the fourth convolution layer and the fifth residual block in sequence to obtain a feature map reduced by 32 times.

[0008] In a possible implementation, the method further includes: The real box sizes in the concealed disease dataset are clustered using a K-Means clustering algorithm, an average overlap degree is used as an objective function, an optimal Anchor size is selected according to the relationship between the number of clusters and the average intersection over union, and the Anchor sizes corresponding to the first scale feature map, the second scale feature map, and the third scale feature map are determined.

[0009] In a possible implementation, the objective function is: wherein is the number of clusters, represents the number of samples in the th cluster center, is the real bounding box of the sample, is the cluster center, is the intersection over union of the real bounding box of the concealed disease and the cluster center box.

[0010] In a possible implementation, the method further comprises: adopting multi-class average precision to evaluate the overall detection precision of the improved YOLOv3 network on the concealed disease.

[0011] In a second aspect, the present application provides a road surface concealed disease detection device, which comprises: a feature extraction module, configured to input the adjusted highway radar image into a Darknet-53 feature extraction network to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times, and a feature map reduced by 32 times, respectively; a first processing module, configured to perform convolution upsampling on the feature map reduced by 32 times, and then perform horizontal connection with the feature map reduced by 16 times to obtain a first feature map, and input the first feature map into a first convolution block to perform convolution to obtain a first scale feature map; a second processing module, configured to perform convolution upsampling on the first feature map, and then perform horizontal connection with the feature map reduced by 8 times to obtain a second feature map, and input the second feature map into a second convolution block to perform convolution to obtain a second scale feature map; a third processing module, configured to perform convolution upsampling on the second feature map, and then perform horizontal connection with the feature map reduced by 4 times to obtain a third feature map, and input the third feature map into a third convolution block to perform convolution to obtain a third scale feature map; a concealed disease detection module, configured to use the first scale feature map, the second scale feature map, and the third scale feature map to perform concealed disease detection to obtain a detection result, wherein the detection result comprises a bounding box position and class information corresponding to a scale.

[0012] In a possible implementation, the Darknet-53 feature extraction network comprises two convolution layers, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block, and the feature extraction module is configured to: input the adjusted highway radar image into the two convolution layers, the first residual block, the first convolution layer, and the second residual block in sequence to obtain the feature map reduced by 4 times; The feature map reduced by 4 times is sequentially input into the second convolutional layer and the third residual block to obtain a feature map reduced by 8 times; The feature map reduced by 8 times is sequentially input into the third convolutional layer and the fourth residual block to obtain a feature map reduced by 16 times; The feature map reduced by 16 times is sequentially input into the fourth convolutional layer and the fifth residual block to obtain a feature map reduced by 32 times.

[0013] In a possible implementation, the apparatus further includes: The Anchor size generation module is configured to use a K-Means clustering algorithm to cluster the real box sizes in the concealed disease data set, use the average overlap degree as a target function, select an optimal Anchor size according to the relationship between the number of clusters and the average intersection-over-union ratio, and determine the Anchor sizes corresponding to the first scale feature map, the second scale feature map, and the third scale feature map.

[0014] In a third aspect, the present application further provides a computer device, which includes a processor and a memory, and the memory stores a computer program, the computer program is loaded and executed by the processor to implement the road concealed disease detection method according to any one of the first aspect.

[0015] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is loaded and executed by a processor to implement the road concealed disease detection method according to any one of the first aspect.

[0016] The above main scheme of the present application and each further selected scheme thereof can be freely combined to form multiple schemes, all of which are the schemes that can be adopted and claimed by the present application; and the present application can also be freely combined between each non-conflicting selection and other selections. Those skilled in the art can understand that there are many combinations according to the prior art and common knowledge after understanding the schemes of the present application, and all of them are the technical schemes claimed by the present application, which will not be enumerated here.

[0017] The application discloses a road surface hidden disease detection method, device, equipment and storage medium. First, a highway radar image is adjusted and then input into a Darknet-53 feature extraction network to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times and a feature map reduced by 32 times. The feature map reduced by 32 times is horizontally connected with the feature map reduced by 16 times to obtain a first feature map, which is input into a first convolution block to perform convolution to obtain a first scale feature map. A second scale feature map and a third scale feature map are obtained in sequence. Finally, the first scale feature map, the second scale feature map and the third scale feature map are used for hidden disease detection to obtain a detection result. Through multi-scale feature extraction and feature fusion at different scales, the detection capability of the model for small targets is enhanced, and the detection precision and efficiency for small targets with different shapes in the highway radar image are improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some of the embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 A flowchart of a road surface hidden disease detection method according to an embodiment of the application is shown.

[0020] Figure 2 A schematic diagram of an improved YOLOv3 network according to an embodiment of the application is shown.

[0021] Figure 3 A graph of the correspondence between the number of clusters (k) and AvgIOU in an embodiment of the application is shown.

[0022] Figure 4 A schematic diagram of the intersection of a predicted box and a real box according to an embodiment of the application is shown. DETAILED DESCRIPTION

[0023] In the prior art, YOLOv3 algorithm, as a one-stage target detection method, has the advantages of high detection precision and fast speed, and has achieved great success in the field of optical graphics. However, the width-height ratio of highway radar graphics is not balanced and the hidden diseases are relatively small. The types of hidden diseases include a large number of internal pores in the base layer, interlayer voids, interlayer looseness and structural looseness. Therefore, the YOLOv3 algorithm is not sufficient to effectively detect highway radar images.

[0024] In order to solve the problems of the non-uniform aspect ratio of the highway radar image and the small target of the highway hidden disease, an application embodiment provides a pavement hidden disease detection method, device, equipment and storage medium. The improved YOLOv3 network is used for intelligent detection of the pretreated pavement hidden disease radar image. The fusion feature map and the boundary box size of the YOLOv3 network are modified. The improved YOLOv3 algorithm has higher detection accuracy than the original YOLOv3 algorithm, and improves the detection efficiency of the highway hidden disease. Next, it will be described in detail.

[0025] Please refer to Figure 1 , Figure 1 A flowchart of a pavement hidden disease detection method is shown. The method is applied to an improved YOLOv3 network, wherein the improved YOLOv3 network includes a Darknet-53 feature extraction network. The method includes the following steps: Step S1, the adjusted highway radar image is input into the Darknet-53 feature extraction network to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times and a feature map reduced by 32 times.

[0026] Figure 2 An improved YOLOv3 network is shown. Due to the problem of non-uniform aspect ratio of the highway radar image, in order to adapt to the requirements of the YOLOv3 algorithm, the input image needs to be adjusted first. The specific method is to resize the original image to a rectangle of 928x224, so that the model can effectively process images of different proportions. After that, the adjusted image is input into the feature extraction network based on the Darknet-53 architecture to obtain multi-scale feature representation. After two initial convolutional layers, a feature map with a dimension of 64 and a size of 464x112 is output.

[0027] Next, the idea of residual network (ResNet) is used to use 1x1 and 3x3 convolution kernels for dimension reduction and feature extraction at different levels of the network. In this way, four feature maps of different scales can be obtained, which are: a feature map of 232x56 (reduced by 4 times), a feature map of 116x28 (reduced by 8 times), a feature map of 58x14 (reduced by 16 times), and a feature map of 29x7 (reduced by 32 times).

[0028] The input initial feature map (size: 928x224) is input into two convolutional layers, and a 64-channel feature map with a size of 464x112 is obtained. The feature map is input into multiple convolutional layers and residual blocks, and a 232x56 feature map is extracted from the 464x112 feature map through further convolutional operations. The convolutional operations are continued, and a 116x28 feature map is extracted from the 232x56 feature map. The convolutional operations are performed again, and a 58x14 feature map is extracted from the 116x28 feature map. Finally, a 29x7 feature map is extracted from the 58x14 feature map.

[0029] In one possible implementation, the Darknet-53 feature extraction network includes two convolutional layers, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block. The steps of inputting the highway radar image into the Darknet-53 feature extraction network after adjustment to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times, and a feature map reduced by 32 times include: The highway radar image is input into the two convolutional layers, the first residual block, the first convolutional layer, and the second residual block in sequence after adjustment to obtain a feature map reduced by 4 times. The feature map reduced by 4 times is input into the second convolutional layer and the third residual block in sequence to obtain a feature map reduced by 8 times. The feature map reduced by 8 times is input into the third convolutional layer and the fourth residual block in sequence to obtain a feature map reduced by 16 times. The feature map reduced by 16 times is input into the fourth convolutional layer and the fifth residual block in sequence to obtain a feature map reduced by 32 times.

[0030] Based on Figure 2 , first, the original highway radar image is adjusted to a size of 928x224 to adapt to the input requirements of the subsequent network. The adjusted image is input into the Darknet-53 architecture for feature extraction. This process is divided into multiple stages, including convolutional layer and residual block operations, and finally generates feature maps of different scales.

[0031] In the initial two convolutional layers, the input is the adjusted image (928x224). Through two initial convolutional layers, each convolutional layer can be connected with batch normalization (Batch Normalization) and an activation function (such as ReLU). The initial two convolutional layers are used for preliminary feature extraction, and output a 64-channel feature map with a size of 464x112.

[0032] In the first residual block, the input is a 64-channel feature map with a size of 464x112, and the output is a 128-channel feature map with a size of 232x56 (reduced by 4 times). The residual block performs feature enhancement and dimension reduction through skip connection and the combination of 1x1 and 3x3 convolution kernels.

[0033] The first convolutional layer continues to extract features, with an input of a 128-channel feature map with a size of 232x56, and an output of a feature map with a size of 232x56. The second residual block again performs feature enhancement, with an input of a 128-channel feature map with a size of 232x56, and an output of a 256-channel feature map with a size of 232x56 (reduced by 4 times), for subsequent multi-scale fusion.

[0034] The input of the second convolutional layer is a 256-channel feature map with a size of 232x56, and the output is a 512-channel feature map with a size of 116x28. The input of the third residual block is a 512-channel feature map with a size of 116x28, and the output is a 512-channel feature map with a size of 116x28 (reduced by 8 times).

[0035] The input of the third convolutional layer is a 512-channel feature map with a size of 116x28, and the output is a 1024-channel feature map with a size of 58x14. The input of the fourth residual block is a 1024-channel feature map with a size of 58x14, and the output is a 1024-channel feature map with a size of 58x14 (reduced by 16 times).

[0036] The input of the fourth convolutional layer is a 1024-channel feature map with a size of 58x14, and the output is a 2048-channel feature map with a size of 29x7. The input of the fifth residual block is a 2048-channel feature map with a size of 29x7, and the output is a 2048-channel feature map with a size of 29x7 (reduced by 32 times).

[0037] The Darknet-53 feature extraction network generates four feature maps of different scales in turn, which can better capture target information of different scales, thereby improving detection accuracy and robustness.

[0038] In step S2, the first feature map is obtained by performing convolution up-sampling on the feature map reduced by 32 times and horizontally connecting it with the feature map reduced by 16 times. The first feature map is input into the first convolutional block to perform convolution and obtain the first scale feature map.

[0039] In step S3, the second feature map is obtained by performing convolution up-sampling on the first feature map and horizontally connecting it with the feature map reduced by 8 times. The second feature map is input into the second convolutional block to perform convolution and obtain the second scale feature map.

[0040] Step S4, after the second feature map is convolved and up-sampled, it is transversely connected with the feature map reduced by 4 times to obtain a third feature map, and the third feature map is input into a third convolution block to obtain a third scale feature map.

[0041] The feature map reduced by 32 times is processed first, and the size of the feature map is (batchsize, 29, 7, 1024), which is a low-resolution feature map processed by the deepest layer of Darknet-53. One or more convolution operations are performed on the feature map to reduce the number of channels, and an up-sampling (Upsample) operation is used to enlarge the resolution of the feature map to the same as the feature map reduced by 16 times, and the output size becomes (batchsize, 58, 14, 512) of the feature map. The feature map is spliced (Concatenate) with the original feature map reduced by 16 times (the size is (batchsize, 58, 14, 512)), so as to combine high and low resolution features, and output a first feature map with a size of (batchsize, 58, 14, 1024). Then, a convolution block (including a convolution layer) is applied to the first feature map to further extract and integrate features, and a first scale feature map with a size of (batchsize, 58, 14, 256) is output.

[0042] Then, one or more convolution operations are performed on the first feature map to reduce the number of channels, and an up-sampling (Upsample) operation is used to enlarge the resolution of the feature map to the same as the feature map reduced by 8 times, and the output size becomes (batchsize, 116, 28, 128) of the feature map. The feature map obtained in the previous step is spliced (Concatenate) with the original feature map reduced by 8 times (the size is (batchsize, 116, 28, 256)), so as to combine high and low resolution features. A second feature map with a size of (batchsize, 116, 28, 384) is output, and then a convolution block (including a convolution layer) is applied to the second feature map to further extract and integrate features, and a second scale feature map with a size of (batchsize, 116, 28, 128) is output.

[0043] The second feature map is subjected to one or more convolution operations to reduce the number of channels, and an upsample operation is used to increase its resolution to the same as the 4-fold reduced feature map. The output size becomes a feature map of (batchsize, 232, 56, 64), which is spliced with the original 4-fold reduced feature map (size (batchsize, 232, 56, 128)) to combine high and low resolution features, and a third feature map of (batchsize, 232, 56, 192) is output. A convolution block (including a convolution layer) is applied to the third feature map to further extract and integrate features, and a third scale feature map of (batchsize, 232, 56, 64) is output.

[0044] The network effectively fuses feature maps of different scales, uses high-resolution feature maps to supplement the information of low-resolution feature maps, thereby improving the accuracy and robustness of target detection. Step S5, using the first scale feature map, the second scale feature map and the third scale feature map to detect the concealed disease to obtain the detection result, the detection result includes the position and category information of the boundary box corresponding to the scale.

[0045] The first scale feature map (batchsize, 116, 28, 27), the second scale feature map (batchsize, 232, 56, 27) and the third scale feature map (batchsize, 464, 112, 27) generated by the above multi-scale detection are output, each grid predicts a bounding box, including position, category (4 types of concealed diseases) and confidence information.

[0046] The improved YOLOv3 network adjusts the feature fusion method, making the detection of small targets (such as road concealed diseases) more accurate. Specifically, by upsampling and horizontal connection, the resolution of the feature map is gradually restored, and high and low resolution features are combined, effectively improving the detection ability of the model for targets of different scales. Finally, the model uses three feature maps of different scales for detection, ensuring the comprehensiveness and accuracy of the detection results.

[0047] In one possible implementation, the method further includes: The K-Means clustering algorithm is used to cluster the real box sizes in the concealed disease dataset, the average overlap is used as the objective function, the optimal Anchor size is selected according to the relationship between the number of clusters and the average intersection over union, and the Anchor sizes corresponding to the first scale feature map, the second scale feature map and the third scale feature map are determined.

[0048] In YOLOv3, the anchor size dimension of network training is fixed, and the anchor size needs to be reselected due to the characteristics of the concealed disease dataset. The width and height of the real box in the concealed disease dataset are clustered using the K-Means clustering algorithm, and the average overlap (Avg IOU) is used as the objective function. The k value is selected as 1-15, and the concealed disease database is clustered and analyzed, Figure 3 The number of clusters (k) and the AvgIOU corresponding relationship diagram in the embodiment of the application is shown, and with the increase of k value, the change of the objective function is more and more stable. The Avg IOU value and the improved YOLOv3 network structure are combined, three feature layers are used for detection, three different size anchors are set in each layer, therefore 9 anchor boxes are selected, and table 1 shows the regenerated anchor size: Table 1

[0049] It can be seen that the adjusted anchor is more consistent with the characteristics of the concealed disease, and the width-height ratio is also consistent with the actual disease width-height ratio.

[0050] The objective function is: wherein k is the number of clusters, n is the number of samples in the th cluster center, is the real label box of the sample, is the cluster center, is the intersection over union of the real label box of the concealed disease and the cluster center box.

[0051] In one possible implementation, the method further includes: The overall detection accuracy of the improved YOLOv3 network for the concealed disease is evaluated by using multi-class average precision.

[0052] The multi-class average precision (mean Average Precision, mAP) is used to evaluate the overall detection accuracy of the algorithm for the concealed disease, and the mAP is the average of the average precision (Average Precision, AP) of each class. The AP value is used to measure the detection effect of the algorithm for a certain class. The calculation of mAP includes intersection over union (Intersection over Union, IoU), precision, recall and AP. The IoU is the overlapping degree of the target region D predicted by the algorithm and the real label region G, and the IoU represents the ratio of their intersection and union.

[0053] Figure 4The intersection and union diagram of the predicted frame and the real frame proposed in the embodiment of the application is shown, wherein the red frame is the real marked area (Ground Truth), and the blue frame is the area predicted by the algorithm (Detection Result). The calculation method is: The greater the IoU value, the smaller the difference between the predicted area and the real area, and the more accurate the target position. In the target detection task, if the IoU of the predicted frame and the real marked frame of the algorithm is greater than a certain threshold, it can be considered that the prediction of the model is accurate. The algorithm selects IoU as 0.5 to calculate mAP.

[0054] In the target detection task, the sample data is divided into true positive (TP), false positive (FP), true negative (TN), and false negative (FN) according to the real label area and the predicted target area. Among them, TP represents the number of positive samples correctly identified; FP represents the number of negative samples incorrectly identified as positive samples; TN represents the number of negative samples correctly identified; FN represents the number of positive samples incorrectly identified as negative samples. The precision rate represents the proportion of the number of pictures correctly predicted to the total number of positive predictions; the recall rate represents the proportion of the number of pictures correctly predicted to the total number of labeled pictures. Their calculation methods are respectively: and .

[0055] When evaluating the algorithm, the precision and recall rates of the two indicators need to be considered. However, the precision and recall rates are negatively correlated. By setting different thresholds, the precision and recall rates of each class can be obtained, and the Precision-Recall curve is obtained by connecting these points. The area under the P-R curve balances the two indicators, and the calculation formula is The AP value ranges from 0 to 1, and the higher the AP value, the better the detection effect. mAP represents the average value of all class APs, and the calculation formula is . In the formula, n represents the number of hidden diseases, and represents the average precision of each class of disease.

[0056] In order to evaluate the accuracy of the improved method for hidden diseases, the improved YOLOv3 method and the traditional YOLOv3 method are compared, and Table 2 shows the AP value and mAP of the two models for hidden disease detection: Table 2

[0057] It can be seen that compared with the original YOLOv3, the AP value and the mAP value of the pavement hidden disease detection method based on ground penetrating radar and improved YOLOv3 are improved. The AP values of interlayer void, pore, interlayer loose and loose structure are increased to 88.5%, 66.7%, 72.3% and 77.8% respectively, and the mAP value is increased to 76.6%.

[0058] In summary, compared with the traditional YOLOv3 algorithm, the AP value of each type of disease is improved, and the detection accuracy is higher.

[0059] Compared with the prior art, the embodiments of the present application have the following beneficial effects: First, by fusing different scale feature maps, especially by upsampling and connecting the feature maps reduced by 32 times, 16 times and 8 times layer by layer, the detection accuracy of small targets is greatly enhanced.

[0060] Second, the K-Means clustering algorithm is used to regenerate 9 Anchor sizes more suitable for the characteristics of the hidden disease data set. This enables the model to more accurately match the real box distribution of the target object, significantly improving the accuracy of the detection box.

[0061] Third, the newly generated Anchor size and the optimized network structure ensure that the model can better cover the target area during detection, further improving the overall performance.

[0062] A possible implementation of a pavement hidden disease detection device is given below, which is used to execute each execution step and the corresponding technical effect of the pavement hidden disease detection method shown in the above embodiments and possible implementations. The device comprises: The feature extraction module is configured to input the adjusted ground radar image into the Darknet-53 feature extraction network to obtain a feature map reduced by 4 times, a feature map reduced by 8 times, a feature map reduced by 16 times and a feature map reduced by 32 times. The first processing module is configured to perform convolution upsampling on the feature map reduced by 32 times, and then horizontally connect the feature map reduced by 16 times to obtain a first feature map. The first feature map is input into a first convolution block for convolution to obtain a first scale feature map. The second processing module is configured to perform convolution upsampling on the first feature map, and then horizontally connect the feature map reduced by 8 times to obtain a second feature map. The second feature map is input into a second convolution block for convolution to obtain a second scale feature map. The third processing module is configured to perform convolution upsampling on the second feature map, and then horizontally connect the feature map reduced by 4 times to obtain a third feature map. The third feature map is input into a third convolution block for convolution to obtain a third scale feature map. The concealed disease detection module is configured to perform concealed disease detection using the first scale feature map, the second scale feature map, and the third scale feature map to obtain a detection result, and the detection result includes a boundary box position and class information of a corresponding scale.

[0063] In a possible implementation, the Darknet-53 feature extraction network includes two convolutional layers, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block, and the feature extraction module is configured to: input the highway radar image to the two convolutional layers, the first residual block, the first convolutional layer, and the second residual block in sequence after adjustment to obtain a feature map reduced by 4 times; input the feature map reduced by 4 times to the second convolutional layer and the third residual block in sequence to obtain a feature map reduced by 8 times; input the feature map reduced by 8 times to the third convolutional layer and the fourth residual block in sequence to obtain a feature map reduced by 16 times; input the feature map reduced by 16 times to the fourth convolutional layer and the fifth residual block in sequence to obtain a feature map reduced by 32 times.

[0064] In a possible implementation, the apparatus further includes: The Anchor size generation module is configured to perform clustering on real box sizes in a concealed disease dataset using a K-Means clustering algorithm, adopt an average overlap degree as a target function, select an optimal Anchor size according to a relationship between a number of clusters and an average intersection over union, and determine Anchor sizes corresponding to the first scale feature map, the second scale feature map, and the third scale feature map.

[0065] The preferred embodiment provides a computer device, which can implement the steps in any of the road surface concealed disease detection methods provided in the embodiments of the present application, and thus can implement the beneficial effects of the road surface concealed disease detection methods provided in the embodiments of the present application. Details are described in the foregoing embodiments, which will not be described here again.

[0066] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor. Therefore, the embodiments of the present application provide a storage medium, which stores a plurality of instructions, and the instructions can be loaded by a processor to execute the steps in any of the road surface concealed disease detection methods provided in the embodiments of the present application.

[0067] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.

[0068] Due to the instructions stored in the storage medium, the steps in any pavement hidden disease detection method embodiment provided by the embodiments of the present application can be executed, thus achieving the beneficial effects of any pavement hidden disease detection method provided by the embodiments of the present application. Details are shown in the foregoing embodiments, which will not be repeated here.

[0069] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting hidden road surface defects, characterized in that, The method is applied to an improved YOLOv3 network, which includes the Darknet-53 feature extraction network. The method includes: After adjusting the highway radar images, they were input into the Darknet-53 feature extraction network to obtain feature maps reduced by 4 times, 8 times, 16 times, and 32 times respectively. After convolutional upsampling of the feature map down by a factor of 32, it is laterally concatenated with the feature map down by a factor of 16 to obtain the first feature map. The first feature map is then input into the first convolutional block for convolution to obtain the first scale feature map. After convolutional upsampling of the first feature map, it is laterally concatenated with the feature map scaled down by 8 times to obtain the second feature map. The second feature map is then input into the second convolutional block for convolution to obtain the second scale feature map. After convolutional upsampling of the second feature map, it is laterally connected with the feature map scaled down by 4 times to obtain the third feature map. The third feature map is then input into the third convolutional block for convolution to obtain the third scale feature map. The detection results of hidden diseases are obtained by using the first-scale feature map, the second-scale feature map and the third-scale feature map. The detection results include the bounding box position and category information of the corresponding scale.

2. The method for detecting hidden road surface defects as described in claim 1, characterized in that, The Darknet-53 feature extraction network consists of two convolutional layers, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block. After adjusting the highway radar image, it is input into the Darknet-53 feature extraction network to obtain feature maps scaled down by a factor of 4, 8, 16, and 32, respectively. The steps include: After adjusting the highway radar image, it is sequentially input into two convolutional layers, the first residual block, the first convolutional layer, and the second residual block to obtain a feature map reduced by a factor of 4; The feature map, reduced by a factor of 4, is sequentially fed into the second convolutional layer and the third residual block to obtain a feature map reduced by a factor of 8. The feature map, reduced by 8 times, is sequentially fed into the third convolutional layer and the fourth residual block to obtain a feature map reduced by 16 times. The feature map, reduced by a factor of 16, is sequentially fed into the fourth convolutional layer and the fifth residual block to obtain a feature map reduced by a factor of 32.

3. The method for detecting hidden road surface defects as described in claim 1, characterized in that, The method further includes: The K-Means clustering algorithm was used to cluster the true bounding box sizes in the latent disease dataset. The average overlap was used as the objective function. The optimal anchor size was selected based on the relationship between the number of clusters and the average intersection-union ratio, and the anchor sizes corresponding to the first-scale feature map, the second-scale feature map, and the third-scale feature map were determined.

4. The method for detecting hidden road surface defects as described in claim 3, characterized in that, The objective function is: ,in It is the number of clusters. Indicates the first The number of samples in each cluster center These are the true bounding boxes of the samples. It is a cluster center. It is the intersection-union ratio of the true bounding boxes of hidden diseases and the cluster center boxes.

5. The method for detecting hidden road surface defects as described in claim 1, characterized in that, The method further includes: The overall detection accuracy of the YOLOv3 network for hidden diseases is improved by using multi-category average accuracy evaluation.

6. A device for detecting hidden road surface defects, characterized in that, The device includes: The feature extraction module is used to adjust the highway radar image and then input it into the Darknet-53 feature extraction network to obtain feature maps reduced by 4 times, 8 times, 16 times, and 32 times respectively. The first processing module is used to perform convolutional upsampling on the feature map that is reduced by 32 times and then horizontally connect it with the feature map that is reduced by 16 times to obtain the first feature map. The first feature map is then input into the first convolutional block for convolution to obtain the first scale feature map. The second processing module is used to perform convolutional upsampling on the first feature map and then horizontally connect it with the feature map that has been reduced by 8 times to obtain the second feature map. The second feature map is then input into the second convolutional block for convolution to obtain the second scale feature map. The third processing module is used to perform convolutional upsampling on the second feature map and then horizontally connect it with the feature map that has been reduced by 4 times to obtain the third feature map. The third feature map is then input into the third convolutional block for convolution to obtain the third scale feature map. The latent disease detection module is used to detect latent diseases using first-scale feature maps, second-scale feature maps, and third-scale feature maps to obtain detection results. The detection results include the bounding box position and category information at the corresponding scale.

7. The pavement hidden defects detection device as described in claim 6, characterized in that, The Darknet-53 feature extraction network consists of two convolutional layers, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a fifth residual block. The feature extraction module is used for: After adjusting the highway radar image, it is sequentially input into two convolutional layers, the first residual block, the first convolutional layer, and the second residual block to obtain a feature map reduced by a factor of 4; The feature map, reduced by a factor of 4, is sequentially fed into the second convolutional layer and the third residual block to obtain a feature map reduced by a factor of 8. The feature map, reduced by 8 times, is sequentially fed into the third convolutional layer and the fourth residual block to obtain a feature map reduced by 16 times. The feature map, reduced by a factor of 16, is sequentially fed into the fourth convolutional layer and the fifth residual block to obtain a feature map reduced by a factor of 32.

8. The pavement hidden defects detection device as described in claim 6, characterized in that, The device further includes: The Anchor Size Generation Module is used to cluster the true bounding box sizes in the latent disease dataset using the K-Means clustering algorithm. It uses the average overlap as the objective function and selects the optimal Anchor size based on the relationship between the number of clusters and the average intersection-union ratio, thus determining the Anchor sizes corresponding to the first-scale feature map, the second-scale feature map, and the third-scale feature map.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the method for detecting hidden road defects as described in any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the method for detecting hidden road defects as described in any one of claims 1-5.

Images