Road disease detection method, device, equipment and medium

Abstract

The invention discloses a road disease detection method, device and equipment and a medium, and relates to the field of road disease detection.The method comprises the steps that a feature map of an asphalt pavement road image is input into a target detection model, the feature map is decomposed into different frequency bands through a frequency selection module, and frequency components of the different frequency bands are obtained; determining a frequency balance characteristic based on the frequency component; determining a first convolutional feature map through a first convolution module according to the frequency balance feature, the feature map and an expansion rate corresponding to each pixel of the feature map; determining a second convolutional feature map according to the convolution kernel weight through a second convolution module, so as to carry out the convolution of the first convolutional feature map and the second convolutional feature map, and obtaining a target convolutional feature map; and through a target three-dimensional detection head constructed according to the target perception attention, determining bounding box coordinates used for marking road diseases and disease categories based on the target feature map after convolution so as to complete road disease detection. Therefore, the accuracy of road disease detection is improved.

Description

TECHNICAL FIELD

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

[0002] Road disease detection is a key link of road maintenance work, and accurate disease identification and statistics are an important way to improve road maintenance efficiency, ensure traffic safety and prolong the service life of roads, and play a key role in intelligent transportation and digital construction of infrastructure. At present, road disease detection still mainly relies on manual visual inspection, and there are significant pain points such as "mood-based" disease analysis, "experience-based" disease analysis, and "eye-based" disease analysis. Many colleges, enterprises and research institutes have developed AI (Artificial Intelligence) detection algorithms for road disease detection, such as yolo (You Only Look Once, a single-stage target detection algorithm), rcnn (Region-based Convolutional Neural Network, region-based convolutional neural network), etc. However, in the face of complex road disease detection business scenarios, different image resolutions, different disease scales, and mixed interference, these algorithms are difficult to generalize, and the disease identification accuracy in business application is only 50%~60%, and it is difficult to identify small diseases.

[0003] Early research on road disease detection mainly focused on machine learning methods based on feature engineering. Threshold segmentation method, edge detection method, region growing algorithm and other methods were first used to extract visual features of road images, and made important explorations in automatic disease detection. Supervised learning algorithms such as support vector machine and random forest first use traditional methods to extract visual features, and then send the extracted visual features to machine learning algorithms for image classification. Although these methods have improved the detection efficiency to some extent, it is difficult for rule-based methods to cope with complex disease characteristics, and the capture ability is limited, gradually difficult to meet the needs of modern road disease detection; with the improvement of computer computing power and the accumulation of data, deep learning-based methods have been widely used in road disease detection, but in the face of the unique problems of road diseases, how to balance the large receptive field and high-frequency information, and realize high-precision and high-accuracy disease identification is still a problem difficult to solve in the industry. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a road disease detection method, device, equipment and medium, which can solve the problem of uneven distribution of difficult-to-identify and easy-to-identify diseases in actual roads, improve the processing capacity of the model for difficult-to-distinguish samples, effectively extract and identify multi-scale diseases, improve the recognition accuracy and recall rate of the algorithm, and reduce disease misidentification and misidentification. The specific scheme is as follows:

[0005] In a first aspect, the present application discloses a road disease detection method, comprising:

[0006] obtaining an asphalt pavement road image, inputting a feature map of the asphalt pavement road image into a target detection model, decomposing the feature map into different frequency bands through a frequency selection module in the target detection model to obtain frequency components of different frequency bands, and determining a frequency balance feature based on the frequency components;

[0007] determining a first convolution feature map after the first convolution module in the target detection model according to the frequency balance feature, the feature map, and the dilution rate corresponding to each pixel of the feature map;

[0008] determining a second convolution feature map after the second convolution module in the target detection model according to the kernel weight, so as to convolve the first convolution feature map and the second convolution feature map to obtain a target convolution feature map;

[0009] determining the bounding box coordinates and the disease category for labeling the road disease based on the target convolution feature map through a target three-dimensional detection head constructed according to a target perception attention in the target detection model, so as to complete the road disease detection; the target perception attention includes scale perception attention, spatial perception attention, and task perception attention.

[0010] Optionally, before the feature map of the asphalt pavement road image is input into the target detection model, the method further comprises:

[0011] obtaining a data set; the data set includes a plurality of asphalt pavement road disease pictures vertically downward;

[0012] artificially labeling the data set through a labeling box to obtain a labeled data set; the labeling box includes a labeling box satisfying a preset small size condition and a labeling box satisfying a preset large size condition.

[0013] Optionally, before the feature map of the asphalt pavement road image is input into the target detection model, the method further comprises:

[0014] determining an inflation factor based on the size relationship between the intersection over union between the prediction box and the real box and the target threshold;

[0015] determining a target loss function according to the product of the inflation factor and a binary cross entropy loss function based on logarithmic likelihood estimation;

[0016] constructing an initial detection model through the frequency selection module, the first convolution module, the second convolution module, and the target three-dimensional detection head;

[0017] Train the initial detection model based on the labeled data set and the target loss function to obtain the target detection model.

[0018] Optionally, the feature map is decomposed into different frequency bands by a frequency selection module in the target detection model to obtain frequency components of different frequency bands, and a frequency balance feature is determined based on the frequency components, including:

[0019] The feature map is decomposed into different frequency bands by a frequency selection module in the target detection model to obtain frequency components of different frequency bands based on a frequency component acquisition formula; the frequency component acquisition formula is:

[0020] ;

[0021] wherein, is the frequency component; is an inverse fast Fourier transform; F is a fast Fourier transform; is a binary mask;

[0022] An initial frequency balance feature of each point in the feature map is determined based on the frequency component by a frequency balance feature determination formula; the initial frequency balance feature determination formula is:

[0023] ;

[0024] wherein, is the initial frequency balance feature of each point in the feature map; is a selection map of the bth frequency band, H and W are the height and width of the feature map respectively; B-1 is the number of pre-set frequency thresholds; is the frequency component corresponding to each point;

[0025] The frequency balance feature of the feature map is determined based on the initial frequency balance feature of each point.

[0026] Optionally, a first convolutional feature map is determined by a first convolution module in the target detection model according to the frequency balance feature, the feature map, and the dilation rate corresponding to each pixel of the feature map, including:

[0027] The feature map is subjected to discrete Fourier transform processing based on the frequency balance feature by a first convolution module in the target detection model to obtain a frequency component matrix; the amplitude of the frequency component matrix is used to represent the frequency distribution of the feature map in the horizontal and vertical directions;

[0028] determine a frequency analysis result according to the frequency component matrix; the frequency analysis result comprises a first expansion rate corresponding to a high-frequency region of the feature map; a second expansion rate corresponding to a low-frequency region of the feature map; the first expansion rate is less than the second expansion rate;

[0029] determine the expansion rate corresponding to each pixel of the feature map by optimizing a convolution layer parameter; wherein an optimization formula of the convolution layer parameter is:

[0030] ;

[0031] wherein, is the convolution layer parameter; is the expansion rate corresponding to the pixel of the position p of the feature map; and respectively represent the pixel with the highest high-frequency power and the pixel with the lowest high-frequency power;

[0032] determine a first post-convolution feature map based on the frequency analysis result and the expansion rate corresponding to each pixel of the feature map by a first post-convolution feature map determination formula; the first post-convolution feature map determination formula is:

[0033] ;

[0034] wherein, is the first post-convolution feature map; K is a convolution kernel size; is a convolution kernel parameter; is a calculation step; is the expansion rate corresponding to the pixel of the position p of the feature map.

[0035] Optionally, the determining a second post-convolution feature map according to a convolution kernel weight by the second convolution module in the target detection model comprises:

[0036] determine the convolution kernel weight based on a convolution kernel weight determination formula by the second convolution module in the target detection model; the convolution kernel weight determination formula is:

[0037] ;

[0038] wherein, is the convolution kernel weight; is a low-frequency part of the convolution kernel; is a high-frequency part of the convolution kernel; and respectively are weights of the low-frequency part and the high-frequency part of the convolution kernel;

[0039] determine a second post-convolution feature map according to the convolution kernel weight by a second post-convolution feature map determination formula; the second post-convolution feature map determination formula is:

[0040] ;

[0041] in, This is the feature map after the second convolution; K is the kernel size. The weights are the convolution kernel weights; The step size is calculated; p is the position.

[0042] Optionally, before determining the bounding box coordinates and disease category for labeling road defects based on the target convolutional feature map using the target 3D detection head constructed according to target perception attention in the target detection model, the method further includes:

[0043] The target 3D detection head is constructed based on scale-aware attention, spatial-aware attention, and task-aware attention; wherein the modeling formula for the target 3D detection head is:

[0044] ;

[0045] in, This is a feature representation that is enhanced layer by layer through a multi-level attention mechanism; The feature map of the target 3D detection head is input; Perceive attention for the task; For spatial awareness attention; Scale-aware attention.

[0046] The method involves using a 3D target detection head constructed based on target perception attention in the target detection model to determine the bounding box coordinates and disease categories for labeling road defects based on the convolutional feature map of the target, thereby completing road defect detection. This includes:

[0047] The target convolutional feature map is subjected to a pooling operation to obtain pooled features;

[0048] The pooled features are segmented to obtain segmented features;

[0049] The segmented features are fused to obtain the fused features;

[0050] The fused features are subjected to robustness and consistency enhancement processing to obtain the processed features; the robustness and consistency enhancement processing includes any one or a combination of feature normalization and standardization processing, attention mechanism weighting processing, feature dimensionality reduction and compression processing, and feature alignment processing.

[0051] The processed features are convolved to obtain the first convolved feature map;

[0052] perform multi-scale feature fusion on the target post-convolution feature map and the first post-convolution feature map to obtain a first post-fusion feature map;

[0053] perform convolution on the post-fusion feature map to obtain a second post-convolution feature map;

[0054] perform up-sampling on the second post-convolution feature map to obtain an up-sampled post-convolution feature map;

[0055] perform fusion on the up-sampled post-convolution feature map and the first post-fusion feature map to obtain a second post-fusion feature map;

[0056] determine, based on the second post-fusion feature map, a bounding box coordinate and a disease category for labeling road diseases through a target three-dimensional detection head constructed according to target perception attention in the target detection model, to complete road disease detection.

[0057] In a second aspect, the present application discloses a road disease detection device, comprising:

[0058] a frequency balance feature determination module configured to acquire an asphalt pavement road image, input a feature map of the asphalt pavement road image into a target detection model, decompose the feature map into different frequency bands through a frequency selection module in the target detection model to obtain frequency components of the different frequency bands, and determine frequency balance features based on the frequency components;

[0059] a first post-convolution feature map determination module configured to determine a first post-convolution feature map through a first convolution module in the target detection model according to the frequency balance features, the feature map, and a dilation rate corresponding to each pixel of the feature map;

[0060] a target post-convolution feature map acquisition module configured to determine a second post-convolution feature map through a second convolution module in the target detection model according to convolution kernel weights, so as to perform convolution on the first post-convolution feature map and the second post-convolution feature map to obtain a target post-convolution feature map;

[0061] a road disease detection module configured to determine, based on the target post-convolution feature map, a bounding box coordinate and a disease category for labeling road diseases through a target three-dimensional detection head constructed according to target perception attention in the target detection model, to complete road disease detection; the target perception attention comprises scale perception attention, spatial perception attention, and task perception attention.

[0062] In a third aspect, the present application discloses an electronic device, comprising:

[0063] a memory configured to store a computer program;

[0064] A processor is configured to execute a computer program to implement the road disease detection method as described above.

[0065] In a fourth aspect, the present application discloses a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is executed by a processor, the road disease detection method as described above is implemented.

[0066] Firstly, the asphalt pavement image is acquired, and a feature map of the asphalt pavement image is input into a target detection model. The feature map is decomposed into different frequency bands by a frequency selection module in the target detection model to obtain frequency components of the different frequency bands, and a frequency balance feature is determined based on the frequency components. A first convolution module in the target detection model determines a first post-convolution feature map according to the frequency balance feature, the feature map, and a dilation rate corresponding to each pixel of the feature map. A second convolution module in the target detection model determines a second post-convolution feature map according to a convolution kernel weight, so as to convolve the first post-convolution feature map and the second post-convolution feature map to obtain a target post-convolution feature map. A target three-dimensional detection head constructed according to a target perception attention in the target detection model determines a bounding box coordinate for labeling a road disease and a disease category based on the target post-convolution feature map, so as to complete road disease detection. The target perception attention includes a scale perception attention, a spatial perception attention, and a task perception attention. It can be seen that, in view of the problem that multiple scales (different sizes and great differences) of diseases are difficult to be considered, a frequency adaptive dilation strategy is introduced, a spatial dilation rate is dynamically adjusted according to a local frequency component, appropriate convolution kernels are allocated to diseases of different sizes, and the balance between a receptive field and high-frequency information is achieved. Secondly, in view of the characteristics of different categories of diseases in terms of geometric shape and position, a three-dimensional dynamic detection head based on an attention mechanism is introduced to realize explicit cooperation of spatial geometric information and category information, so that the model can exert the inherent potential of disease categories and spatial positions. In this way, the balance between the receptive field and the high-frequency information is achieved, the model can better process diseases of different sizes, the accuracy of model detection is improved, the detection efficiency is greatly improved, and the labor cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0068] Figure 1 A road disease detection method disclosed by the present application is shown in a flowchart.

[0069] Figure 2 A data labeling schematic diagram disclosed by the present application;

[0070] Figure 3 A model convolution process schematic diagram disclosed by the present application;

[0071] Figure 4 A detection head schematic diagram disclosed by the present application;

[0072] Figure 5 (a) is a prior art recognition result schematic diagram disclosed by the present application;

[0073] Figure 5 (b) is a recognition result schematic diagram of the method of the present application disclosed by the present application;

[0074] Figure 6 A road disease detection network structure schematic diagram disclosed by the present application;

[0075] Figure 7 A road disease detection device structure schematic diagram disclosed by the present application;

[0076] Figure 8 A schematic diagram of an electronic device structure disclosed by the present application. DETAILED DESCRIPTION

[0077] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0078] Currently, road disease detection is mainly based on manual visual inspection, and there are significant pain points such as "mood-based", "experience-based", and "eye power-based" disease analysis. Many colleges, enterprises, and research institutes have developed AI detection algorithms for road disease detection, such as yolo, rcnn, etc. However, in the face of complex road disease detection business scenarios, different image resolutions, different disease scales, and mixed interference, these algorithms are difficult to generalize, and the disease identification accuracy in business applications is only 50%~60%, and it is difficult to identify small diseases. In order to solve the above technical problems, the present application discloses a road disease detection method, device, equipment and medium, which can solve the problem of uneven distribution of difficult-to-identify and easy-to-identify diseases in actual roads, improve the processing capability of the model for difficult-to-distinguish samples, effectively extract and identify multi-scale diseases, improve the algorithm recognition accuracy and recall rate, and reduce disease misidentification and misidentification.

[0079] Referring to Figure 1 The embodiments of the present application disclose a road disease detection method, which comprises:

[0080] Step S11, an asphalt pavement road image is acquired, a feature map of the asphalt pavement road image is input into a target detection model, the feature map is decomposed into different frequency bands by a frequency selection module in the target detection model to obtain frequency components of the different frequency bands, and a frequency balance feature is determined based on the frequency components.

[0081] In this embodiment, the asphalt pavement road image is first acquired, and then a feature map of the asphalt pavement road image is input into a target detection model. However, before the feature map of the asphalt pavement road image is input into the target detection model, a data set is acquired; the data set includes a plurality of asphalt pavement road disease pictures vertically shot downward; the data set is manually labeled by a labeling box to obtain a labeled data set; the labeling box includes a labeling box satisfying a preset small size condition and a labeling box satisfying a preset large size condition. The data set is from the asphalt pavement road disease pictures vertically shot downward, the labeling mode adopts a combination of large and small labeling boxes, the labeling effect is as shown in Figure 2 illustrated, the sensitivity of the model to small-scale features is improved, and the learning of global features by the model is emphasized, so that the disease recognition accuracy can be effectively improved, and misidentification of interference objects can be avoided. In order to solve the problems that the existing algorithm is difficult to generalize in the road detection scene, the disease recognition accuracy is low, and the misidentification and missed identification are serious, corresponding innovations are made to the model. The present application determines an inflation factor based on the size relationship between the intersection over union between the predicted box and the real box and the target threshold; determines a target loss function according to the product of the inflation factor and a binary cross-entropy loss function based on logarithmic likelihood estimation; constructs an initial detection model through the frequency selection module, the first convolution module, the second convolution module and the target three-dimensional detection head; and trains the initial detection model based on the labeled data set and the target loss function to obtain the target detection model. Since there is an imbalance problem between easy and difficult samples in the data distribution in the road defect detection scene. The number of simple samples in the data set is relatively large, and the number of difficult samples is small, so that the model tends to focus on simple samples and cannot accurately classify difficult samples. To solve this problem, the Slide Loss weight function is introduced into the road disease detection task to optimize the function, and the closeness of the detection box and the target box is measured to cope with the challenge of sample imbalance. The specific implementation process is as follows: when calculating the loss process in the training stage, the function assigns a larger inflation factor to the difficult-to-classify samples, guiding the model to pay more attention to the difficult-to-classify samples, thereby improving the discrimination ability of the model for difficult samples and improving the detection accuracy. Specifically, the determination of the inflation factor depends on the intersection over union between the predicted box and the real box (Intersection over Union, IoU), for the IoU threshold, samples with IoU close to the threshold μ are regarded as difficult samples, and samples with large distance from μ are divided into easy samples. For difficult classification samples, the inflation factor The weight function is calculated by the function f(x):

[0082] ;

[0083] The mean value of all bounding box IoU is used as the value of μ. According to the relationship between the sample intersection ratio (IoU) value and the threshold μ, different weight functions are used to weight the original composite loss function BCEWithLogitsLoss, and then back propagation is performed:

[0084] ;

[0085] Wherein, is the target loss function; is the inflation factor; is the binary cross-entropy loss function based on the logit estimate (the original composite loss function BCEWithLogitsLoss).

[0086] In this way, using the Slide Loss loss function effectively solves the problem of uneven disease distribution and enhances the model's ability to handle difficult samples.

[0087] After the target detection model is constructed, the frequency component acquisition formula is used to decompose the feature map into different frequency bands through the frequency selection module in the target detection model, and the frequency components of different frequency bands are obtained. The frequency component acquisition formula is:

[0088] ;

[0089] Wherein, is the frequency component; is the inverse fast Fourier transform; F is the fast Fourier transform; is a binary mask.

[0090] The initial frequency balance feature of each point in the feature map is determined based on the frequency component through the frequency balance feature determination formula. The initial frequency balance feature determination formula is:

[0091] ;

[0092] Wherein, is the initial frequency balance feature of each point in the feature map; is the selection map of the bth frequency band, H and W are the height and width of the feature map respectively. the frequency component corresponding to each point.

[0093] determining the frequency balance feature of the feature map based on the initial frequency balance feature of each point.

[0094] That is, the frequency selection module in the target detection model of the present application balances high-frequency and low-frequency information by dynamically adjusting the weights of different frequency components, thereby enhancing the receptive field. The frequency selection module first decomposes the input feature map into different frequency bands: . Wherein is the inverse fast Fourier transform, which is used to convert the frequency domain signal back to the spatial domain. is a binary mask used to extract components in a specific frequency range:

[0095] ;

[0096] Wherein and are predefined frequency thresholds for dividing different frequency bands, respectively taken from B+1 pre-set frequency thresholds , and then the frequency selection balances the weights of different frequency components through spatial transformation and reweighting:

[0097] ;

[0098] Wherein, is the frequency balance feature learned after frequency selection, is the selection map of the bth frequency band, which is used to dynamically reweight the frequency components in different frequency bands in space. Preferably, the frequency range is divided into 4 frequency bands at an octave interval, i.e. .

[0099] Step S12, determining a first convolution feature map after the first convolution module in the target detection model according to the frequency balance feature, the feature map and the expansion rate corresponding to each pixel of the feature map.

[0100] In this embodiment, after determining the frequency balance feature, the feature map is processed by the first convolution module in the target detection model based on the frequency balance feature to obtain a frequency component matrix; the amplitude of the frequency component matrix is used to represent the frequency distribution of the feature map in the horizontal and vertical directions; a frequency analysis result is determined according to the frequency component matrix; the frequency analysis result includes a first expansion rate corresponding to a high-frequency region of the feature map; a second expansion rate corresponding to a low-frequency region of the feature map; the first expansion rate is smaller than the second expansion rate; the expansion rate corresponding to each pixel of the feature map is determined by optimizing the convolution layer parameter; wherein the optimization formula of the convolution layer parameter is:

[0101] ;

[0102] wherein, is the convolution layer parameter; is the expansion rate corresponding to the pixel of the feature map at position p; and represent the highest and lowest pixels of the high-frequency power respectively; the first post-convolution feature map is determined based on the frequency analysis result and the expansion rate corresponding to each pixel of the feature map by the first post-convolution feature map determination formula; the first post-convolution feature map determination formula is:

[0103] ;

[0104] wherein, is the first post-convolution feature map; K is the convolution kernel size; is the convolution kernel parameter; is the calculation step; is the expansion rate corresponding to the pixel of the feature map at position p.

[0105] In order to reasonably assign different expansion rates to different local parts, first, the frequency of the input feature is analyzed, and for the feature map , the discrete Fourier transform is used for processing, which can be represented as:

[0106] ;

[0107] wherein, For the output of DFT (Discrete Fourier Transform), H and W are its height and width, h and w are the indices of the feature map, the normalized frequencies in the height and width dimensions are given by |u| and |v| respectively, |u| is the spatial frequency of the image in the horizontal direction (width direction, i.e. x-axis), |v| is the spatial frequency of the image in the vertical direction (height direction, i.e. y-axis), and X(h, w) is the frequency balanced feature of the feature map calculated by the frequency selection module. Among them, the amplitude |F(u, v)| of F(u, v) reflects the frequency distribution of the feature map in the horizontal u and vertical v directions. High frequency components correspond to edges, textures and other details, and low frequency components correspond to global structures. Through DFT analysis, the model can quantify the frequency characteristics of different regions (such as smaller receptive fields for high frequency regions to preserve details, and larger receptive fields for low frequency regions to capture global information). Therefore, based on the results of frequency analysis, the model balances between large receptive field and effective bandwidth, thereby selecting different dilation rates for different local information. Since the input feature map has undergone a spatial transformation, the optimal dilation rate for each pixel may be different, therefore, the adaptive dilation rate strategy assigns different dilation rates to each pixel to achieve a better balance. The specific convolution operation can be described as:

[0108] ;

[0109] wherein, is the first feature map after the first convolution; K is the size of the convolution kernel; is the convolution kernel parameter; is the calculation step; the receptive field can be expanded by the dynamic dilation rate is the dilation rate corresponding to the pixel of the position p of the feature map, which is predicted by the convolution layer parameter θ, the parameter is θ and the non-negativity of the dilation rate is ensured by adding a ReLU layer. The dynamic dilation rate is the key to balance the receptive field and high frequency information. Therefore, determining the dynamic dilation rate can be regarded as an optimization problem, that is, maximizing the difference between the receptive field and the lost high frequency information. Regarding the description of the receptive field, the local of feature p is called window size s, , the receptive field RF(p) of the convolution layer is positively correlated with , which can be described as:

[0110] ;

[0111] wherein, K is the size of the convolution kernel, in order to measure the loss of high frequency information caused by the increase of the dilation rate, the high frequency power is further calculated:

[0112] ;

[0113] ​So far, the model self-learning optimization objective can be described as:

[0114]

[0115] However, since the frequency set has discreteness, and the calculation of HP is non-differentiable. Therefore, the present application chooses to directly optimize , that is, increasing the expansion rate at the position where HP(p) is low to encourage a larger receptive field, and reducing the expansion rate at the position where HP(p) is high to reduce the loss of frequency information, θ can be expressed as follows:

[0116]

[0117] wherein, and represent the pixel with the highest high-frequency power and the pixel with the lowest high-frequency power, respectively.

[0118] In step S13, a second convolution module in the target detection model determines a second convolution feature map according to convolution kernel weights, so as to convolve the first convolution feature map and the second convolution feature map to obtain a target convolution feature map.

[0119] In the embodiment, the convolution kernel weights are determined by the second convolution module in the target detection model based on a convolution kernel weight determination formula; the convolution kernel weight determination formula is:

[0120]

[0121] wherein, is the convolution kernel weight; is a low-frequency part of the convolution kernel; is a high-frequency part of the convolution kernel; and are weights of the low-frequency part and the high-frequency part of the convolution kernel, respectively; a second convolution feature map is determined according to the convolution kernel weight by a second convolution feature map determination formula; the second convolution feature map determination formula is:

[0122]

[0123] wherein, is the second convolution feature map; represents a pixel value at p in the feature map; K is a convolution kernel size; is the convolution kernel weight; is a calculation step; p is a position.

[0124] ​​​​The frequency-based adaptive dilated convolution achieves a delicate balance between effective bandwidth and receptive field by assigning an individual dilation rate to each pixel, while in traditional convolution operations, the weights of the convolution kernel are static, that is, fixed once the training is completed. Although such static convolution kernels can capture information of different frequencies, there is still room for enhancement when dealing with complex visual patterns. To further improve the effective bandwidth, the present application applies a dynamic weighting strategy to use adaptive dynamic convolution kernels to process and capture low-frequency and high-frequency information respectively, and the specific implementation process is as follows:

[0125] For a static convolution kernel, the weight decoupling can be described as , is the low-frequency part of the convolution kernel, which acts as a KxK low-frequency mean filter, is the high-frequency part of the convolution kernel, which can be defined as the difference between the convolution kernel and the low-frequency part.

[0126] The adaptive dynamic convolution module dynamically adjusts the weights of the low-frequency and high-frequency parts:

[0127] ;

[0128] and As the weights of the low-frequency part and the high-frequency part of the convolution kernel, they are predicted through the global pooling layer and the convolution layer. Adjusting the proportion of the two enables the network to flexibly focus on low-frequency background information or high-frequency detail information according to the complexity of the input features.

[0129] Therefore, the second convolution module, the adaptive dynamic convolution module in the present application, can be represented as:

[0130] ;

[0131] wherein, is the second convolution feature map; K is the size of the convolution kernel; is the weight of the convolution kernel; is the calculation step; p is the position.

[0132] Finally, as shown in Figure 3 , the first convolution feature map and the second convolution feature map are convolved to obtain the target convolution feature map .

[0133] In this way, since the road detection image resolution is large, such as 4000*2000, in the commonly used algorithm model, when the image is input to the model and resized to the specified size, the fine high-frequency information is easily lost, and it is difficult to balance the global information of the disease due to the limitation of the receptive field. To solve these problems, without significantly increasing the parameter amount of the model, the application creatively introduces frequency adaptive dilated convolution from the perspective of frequency spectrum analysis, dynamically adjusts the spatial expansion rate for different parts of features according to local frequency components, and dynamically allocates different inflation coefficients for different local feature maps, so as to better balance the larger receptive field and retain high-frequency information.

[0134] In step S14, the target three-dimensional detection head constructed according to the target perception attention in the target detection model is used to determine the boundary box coordinates for labeling the road disease and the disease category based on the target convolution feature map, so as to complete the road disease detection; the target perception attention includes scale perception attention, spatial perception attention and task perception attention.

[0135] In the embodiment, in the traditional target detection task, modeling the position and shape relationship between objects can better help object recognition, and the same is true for road surface diseases. Different categories of diseases have different shape and position characteristics, although the target detection head of the existing model has implicitly modeled the spatial and category relationship, but this modeling is difficult to fully play the synergistic effect of the two optimization objectives. The present application explicitly interacts the category judgment of the disease with the geometric information, fully plays the recognition information, optimizes the internal synergistic potential between the optimization objectives, and realizes higher target recognition accuracy, and the specific implementation manner is as follows. Multiple self-attention mechanisms are used to explicitly model between scale perception, spatial perception and task perception, as shown in Figure 4 , a three-dimensional explicit synergistic dynamic detection head is constructed. It can be represented as:

[0136] ;

[0137] Among them, is the feature representation enhanced layer by layer through the multi-level attention mechanism; is the feature map input to the target three-dimensional detection head; is the task perception attention; is the spatial perception attention; is the scale perception attention. The scale perception attention is applied in the horizontal scale; the spatial perception attention focuses on the spatial dimension of the feature; and the task perception attention is arranged in the channel dimension.

[0138] The specific process for determining the bounding box coordinates and the disease category for labeling the road disease based on the target convolution feature map through the target three-dimensional detection head constructed according to the target perception attention in the target detection model is as follows: performing a pooling operation on the target convolution feature map to obtain a pooled feature; segmenting the pooled feature to obtain a segmented feature; fusing the segmented feature to obtain a fused feature; performing robustness and consistency enhancement processing on the fused feature to obtain a processed feature; the robustness and consistency enhancement processing includes any one or a combination of several of feature normalization and standardization processing, attention mechanism weighting processing, feature dimension reduction and compression processing, and feature alignment processing; performing convolution on the processed feature to obtain a first convolution feature map; performing multi-scale feature fusion on the target convolution feature map and the first convolution feature map to obtain a first fused feature map; performing convolution on the fused feature map to obtain a second convolution feature map; performing up-sampling on the second convolution feature map to obtain an up-sampled feature map; fusing the up-sampled feature map and the first fused feature map to obtain a second fused feature map; determining the bounding box coordinates and the disease category for labeling the road disease based on the second fused feature map through the target three-dimensional detection head constructed according to the target perception attention in the target detection model to complete the road disease detection.

[0139] In addition, the computing resources of the present application are shown in Table 1. The quantitative indicators are shown in Table 2, and the visualization comparison of the recognition results of the existing method and the recognition results of the present application is shown in Figures 5(a) and 5(b). It can be seen that the present application has achieved excellent results in various indicators, and the performance of the model of the present application is particularly outstanding in the key indicator of mAP@0.5 (mean Average Precision at IoU=0.5, which represents the average value of the detection accuracy of the model for all classes when the Intersection over Union (IoU) threshold is 0.5). The actual ability of feature extraction is improved, thereby having higher detection accuracy and reliability in actual application, and being more in line with the requirements of target detection in complex scenes.

[0140] Table 1

[0141]

[0142] Table 2

[0143]

[0144] In addition, compared with the traditional manual inspection method, the present application can realize automatic detection of road diseases, realize the increase of the detection mileage from 20KM / person-day to 200 kilometers / person-day, and improve the detection efficiency by 10 times, thereby greatly improving the detection efficiency and reducing the labor cost.

[0145] In summary, the present application first acquires an asphalt pavement road image, inputs a feature map of the asphalt pavement road image into a target detection model, decomposes the feature map into different frequency bands through a frequency selection module in the target detection model to obtain frequency components of the different frequency bands, and determines a frequency balance feature based on the frequency components; determines a first convolution feature map based on the frequency balance feature, the feature map, and a dilution rate corresponding to each pixel of the feature map through a first convolution module in the target detection model; determines a second convolution feature map based on a convolution kernel weight through a second convolution module in the target detection model, so as to convolve the first convolution feature map and the second convolution feature map to obtain a target convolution feature map; determines a bounding box coordinate for labeling a road disease and a disease category based on the target convolution feature map through a target three-dimensional detection head constructed based on a target perception attention in the target detection model, so as to complete road disease detection; and the target perception attention includes a scale perception attention, a spatial perception attention, and a task perception attention. As can be seen, the present application introduces a frequency adaptive expansion strategy to dynamically adjust a spatial dilution rate according to a local frequency component, to match appropriate convolution kernels for diseases of different sizes, to balance a receptive field and high-frequency information. Secondly, a three-dimensional dynamic detection head based on an attention mechanism is introduced to realize explicit cooperation of spatial geometric information and category information, so that the model can exert the inherent potential of disease categories and spatial positions. In this way, the balance of the receptive field and the high-frequency information is achieved, the model can better handle diseases of different sizes, the accuracy of model detection is improved, the detection efficiency is greatly improved, and the labor cost is reduced.

[0146] Based on the previous embodiment, the present application discloses a road disease detection method, which can determine a bounding box coordinate for labeling a road disease and a disease category based on a target convolution feature map through a target three-dimensional detection head constructed based on a target perception attention in the target detection model, so as to complete road disease detection. Next, a specific road disease detection process will be described in detail.

[0147] As shown in Figure 6 , it includes a feature extraction network, a feature fusion network, and a target detection head.

[0148] In the feature extraction network, the frequency-based adaptive dilated convolution is based on the input feature map, the frequency selection module extracts the feature frequency information, and then the adaptive dilated convolution module dynamically adjusts the dilated rate of the convolution kernel (to adapt to different frequency features), and the output feature is connected to BatchNorm2d (Batch Normalization 2D, Batch Normalization) and SiLU (Sigmoid Linear Unit, activation function) to enhance the feature expression ability. Convolution module: regular convolution operation is performed on the output of the feature extraction network to extract basic features. Multi-scale fusion module: convolution layer: convolution operation is performed on the output of the feature extraction network to preliminarily extract features. Pooling layer: the dimensionality of the features is reduced and the features are aggregated through pooling operations (such as max pooling) to retain key information. Feature segmentation: the pooled features are segmented as needed to provide branch features for subsequent multi-scale fusion. Feature fusion module: the multi-scale branch features are fused (such as splicing, weighted summation) to integrate feature information of different scales. Fusion layer: the fused features are further processed to enhance the robustness and consistency of the features. Convolution layer: the output of the fusion layer is convolved again to refine the feature expression.

[0149] In the feature fusion network, the multi-scale fusion module: the outputs of the feature extraction network and the multi-scale fusion module are fused with multi-scale features (such as cross-scale splicing, attention mechanism fusion) to realize the cooperation of features of different scales. Convolution module: convolution operation is performed on the fused features to extract high-level features after fusion. Up-sampling layer: the features are up-sampled (such as deconvolution, interpolation) to restore the spatial resolution of the features to provide matching scale features for subsequent fusion. Fusion layer: the up-sampled features and the outputs of the multi-scale fusion module are fused to further integrate multi-scale information.

[0150] In the target detection head (three-dimensional explicit collaborative dynamic detection head), the multi-scale fusion module: the output of the feature fusion network is fused with multi-scale features to ensure that the detection head receives multi-scale rich features. Convolution module: convolution operation is performed on the fused features to extract the features required by the detection head. Fusion layer: the outputs of the convolution module and the multi-scale fusion module are fused to enhance the detection capability of the features. Multi-scale fusion module: the fused features are fused again with multi-scale features to further optimize the scale coverage of the features. Convolution module: convolution operation is performed on the final fused features to output the feature map required by the detection head.

[0151] In summary, the system takes the input feature map as the starting point, extracts the basic features through the feature extraction network, realizes multi-scale and multi-level feature fusion through the multi-scale fusion module and the feature fusion network, and finally completes the target detection by the three-dimensional explicit collaborative dynamic detection head. Through convolution, pooling, upsampling, fusion and other operations between modules, a complete process from feature extraction to detection output is constructed, realizing efficient processing and accurate detection of multi-modal information.

[0152] Referring to Figure 7 As shown in the figure, the embodiment of the present application discloses a road disease detection device, comprising:

[0153] The frequency balance feature determination module 11 is configured to obtain an asphalt pavement road image, input a feature map of the asphalt pavement road image into a target detection model, decompose the feature map into different frequency bands through a frequency selection module in the target detection model, obtain frequency components of the different frequency bands, and determine frequency balance features based on the frequency components.

[0154] The first convolution feature map determination module 12 is configured to determine a first convolution feature map based on the frequency balance features, the feature map, and the expansion rate corresponding to each pixel of the feature map through a first convolution module in the target detection model.

[0155] The target convolution feature map acquisition module 13 is configured to determine a second convolution feature map through a second convolution module in the target detection model based on convolution kernel weights, so as to convolve the first convolution feature map and the second convolution feature map to obtain a target convolution feature map.

[0156] The road disease detection module 14 is configured to determine the boundary box coordinates for labeling the road disease and the disease category based on the target convolution feature map through a target three-dimensional detection head constructed based on a target perception attention in the target detection model, so as to complete the road disease detection. The target perception attention includes scale perception attention, spatial perception attention, and task perception attention.

[0157] Since the embodiments of the device part correspond to the above-mentioned embodiments, the embodiments of the device part are described with reference to the embodiments of the above-mentioned method part, and will not be repeated here.

[0158] It can be seen that, in the face of the problem that multiple scales (size is uneven, difference is great) diseases are difficult to balance, the frequency adaptive expansion strategy is introduced, the spatial expansion rate is dynamically adjusted according to the local frequency component, and appropriate convolution kernel is allocated for diseases of different sizes, so as to realize the balance of receptive field and high-frequency information. Secondly, in the face of different types of diseases with different geometric shape positions, a three-dimensional dynamic detection head based on attention mechanism is introduced to realize the explicit cooperation of spatial geometric information and category information, so that the model can exert the inherent potential of disease category and spatial position. In this way, the balance of receptive field and high-frequency information is realized, the model can better process diseases of different sizes, the accuracy of model detection is improved, the detection efficiency is greatly improved, and the labor cost is reduced.

[0159] Further, the embodiment of the present application further discloses an electronic device, Figure 8 The electronic device 20 structure diagram shown in accordance with an exemplary embodiment, the contents in the figure cannot be considered as any limitation on the scope of use of the present application.

[0160] Figure 8 The electronic device 20 structure diagram provided by the embodiment of the present application. The electronic device 20, specifically can include: at least one processor 21, at least one memory 22, power supply 23, communication interface 24, input output interface 25 and communication bus 26. Wherein, the memory 22 is used for storing computer program, the computer program is loaded and executed by the processor 21, to realize the related steps in the road disease detection method disclosed by any of the preceding embodiments. In addition, the electronic device 20 in the embodiment of the present application can be an electronic computer.

[0161] In the embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create data transmission channel between the electronic device 20 and the external device, and the communication protocol followed is any communication protocol applicable to the technical solution of the present application, which is not limited here; the input output interface 25 is used to obtain external input data or output data to the outside world, and the specific interface type can be selected according to the specific application needs, which is not limited here.

[0162] In addition, the memory 22 as the carrier of resource storage can be read-only memory, random access memory, disk or optical disk, etc., and the resources stored thereon can include operating system 221, computer program 222, etc., and the storage mode can be temporary storage or permanent storage.

[0163] The operating system 221 is configured to manage and control the various hardware devices on the electronic device 20, and the computer program 222 can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of performing the road disease detection method disclosed in any of the preceding embodiments by the electronic device 20, the computer program 222 can further include computer programs capable of performing other specific tasks.

[0164] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the road disease detection method disclosed in the preceding embodiments. For the specific steps of the method, please refer to the corresponding content disclosed in the preceding embodiments, which will not be repeated here.

[0165] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0166] The skilled person can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of the two. In order to clearly show the interchangeability of hardware and software, the composition and steps of each example have been described in the above description. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0167] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of the two. The software modules can be placed in random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0168] Finally, it needs to be pointed out that in this document, relational terms such as first and second and the like can only be intended to distinguish one entity or operation from another entity or operation without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus including the stated element.

[0169] The above detailed description of the technical solutions provided by the present application has been provided, and the principles and implementation modes of the present application have been described by applying specific examples. The above description of the examples is only for the purpose of helping to understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description of the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method of detecting a road disease, characterized by, The method comprises the following steps: acquiring an asphalt pavement road image, inputting a feature map of the asphalt pavement road image into a target detection model, decomposing the feature map into different frequency bands through a frequency selection module in the target detection model to obtain frequency components of the different frequency bands, and determining a frequency balance feature based on the frequency components; determining a first convolution feature map based on the frequency balance feature, the feature map, and a dilation rate corresponding to each pixel of the feature map through a first convolution module in the target detection model; determining a second convolution feature map based on a convolution kernel weight through a second convolution module in the target detection model, so as to convolve the first convolution feature map and the second convolution feature map to obtain a target convolution feature map; determining a bounding box coordinate and a disease category for labeling road diseases based on the target convolution feature map through a target three-dimensional detection head constructed based on a target perception attention in the target detection model, so as to complete road disease detection; the target perception attention comprises a scale perception attention, a spatial perception attention, and a task perception attention.

2. The road disease detection method according to claim 1, characterized by, Before the feature map of the asphalt pavement road image is input into the target detection model, the method further comprises the following steps: acquiring a data set; the data set comprises a plurality of asphalt pavement road disease pictures vertically shot downward; performing artificial labeling on the data set through a labeling box to obtain a labeled data set; the labeling box comprises a labeling box satisfying a preset small size condition and a labeling box satisfying a preset large size condition.

3. The road disease detection method according to claim 2, characterized by, Before the feature map of the asphalt pavement road image is input into the target detection model, the method further comprises the following steps: determining an inflation factor based on a size relationship between an intersection over union between a prediction box and a true box and a target threshold; determining a target loss function based on a product of the inflation factor and a binary cross entropy loss function based on a logarithmic likelihood estimation; constructing an initial detection model through the frequency selection module, the first convolution module, the second convolution module, and the target three-dimensional detection head; training the initial detection model based on the labeled data set and the target loss function to obtain the target detection model.

4. The road disease detection method according to claim 1, characterized by, The method that the feature map is decomposed into different frequency bands through the frequency selection module in the target detection model to obtain frequency components of the different frequency bands and the frequency balance feature is determined based on the frequency components comprises the following steps: the feature map is decomposed into different frequency bands through the frequency selection module in the target detection model to obtain frequency components of the different frequency bands based on a frequency component acquisition formula; the frequency component acquisition formula is: ； wherein is the frequency component; is the inverse fast Fourier transform; F is the fast Fourier transform; is a binary mask; initial frequency balance features of each point in the feature map are determined based on the frequency components through a frequency balance feature determination formula; the initial frequency balance feature determination formula is: ； wherein, is the initial frequency balance feature of each point in the feature map; is the selection map of the bth frequency band, H and W are the height and width of the feature map respectively; B-1 is the number of pre-set frequency thresholds; is the frequency component corresponding to each point; the frequency balance feature of the feature map is determined based on the initial frequency balance features of each point.

5. The road disease detection method according to claim 1, characterized by, The method that the first convolution feature map is determined based on the frequency balance feature, the feature map, and a dilation rate corresponding to each pixel of the feature map through the first convolution module in the target detection model comprises the following steps: Discrete Fourier transform processing is performed on the feature map based on the frequency balance feature in the first convolution module of the target detection model to obtain a frequency component matrix; the amplitude of the frequency component matrix is used to represent the frequency distribution of the feature map in the horizontal and vertical directions; A frequency analysis result is determined according to the frequency component matrix; the frequency analysis result includes a first expansion rate corresponding to a high-frequency area of the feature map; a second expansion rate corresponding to a low-frequency area of the feature map; and the first expansion rate is smaller than the second expansion rate; An expansion rate corresponding to each pixel of the feature map is determined by optimizing a convolution layer parameter; wherein the optimization formula of the convolution layer parameter is: ； wherein, is the parameter of the convolution layer; is the dilation rate corresponding to the pixel of the position p of the feature map; and respectively represent the pixel with the highest power and the pixel with the lowest power. A first post-convolution feature map is determined based on the frequency analysis result and the expansion rate corresponding to each pixel of the feature map by using a first post-convolution feature map determination formula; wherein the first post-convolution feature map determination formula is: ； wherein, is the first feature map after the convolution; K is the size of the convolution kernel; is the convolution kernel parameter; is the calculation step; is the dilation rate corresponding to the pixel of the position p of the feature map.

6. The road disease detection method according to claim 1, characterized by, The second post-convolution feature map is determined according to the kernel weight by using a second convolution module in the target detection model, including: The kernel weight is determined based on a kernel weight determination formula by using the second convolution module in the target detection model; wherein the kernel weight determination formula is: ； wherein, is the kernel weight; is a low frequency portion of the kernel; is a high frequency portion of the kernel; and are weights of the low and high frequency portions of the kernel, respectively. The second post-convolution feature map is determined according to the kernel weight by using a second post-convolution feature map determination formula; wherein the second post-convolution feature map determination formula is: ； wherein, is the second convolutional feature map; K is the size of the convolution kernel; is the weight of the convolution kernel; is the calculation step; p is the position.

7. The road disease detection method according to claim 1, characterized by, Before determining the bounding box coordinates and the disease category for labeling the road disease based on the target post-convolution feature map, the target three-dimensional detection head constructed according to the target perception attention in the target detection model further includes: The target three-dimensional detection head is constructed based on scale perception attention, spatial perception attention and task perception attention; wherein the modeling formula of the target three-dimensional detection head is: ； wherein, is a feature representation enhanced layer by layer by a multi-level attention mechanism; is a feature map input to the target three-dimensional detection head; is the task-aware attention; is the spatial-aware attention; is the scale-aware attention.

8. The road disease detection method according to any one of claims 1 to 7, characterized by, The target three-dimensional detection head constructed according to the target perception attention in the target detection model determines the bounding box coordinates and the disease category for labeling the road disease based on the target post-convolution feature map to complete the road disease detection, including: A pooling operation is performed on the target post-convolution feature map to obtain a pooled feature; The pooled feature is segmented to obtain a segmented feature; The segmented feature is fused to obtain a fused feature; Robustness and consistency enhancement processing is performed on the fused feature to obtain a processed feature; the robustness and consistency enhancement processing includes any one or a combination of several of feature normalization and standardization processing, attention mechanism weighting processing, feature dimension reduction and compression processing, and feature alignment processing; Convolution is performed on the processed feature to obtain a first post-convolution feature map; Multi-scale feature fusion is performed on the target post-convolution feature map and the first post-convolution feature map to obtain a first fused feature map; Convolution is performed on the fused feature map to obtain a second post-convolution feature map; An up-sampling operation is performed on the second post-convolution feature map to obtain an up-sampled feature map; The up-sampled feature map and the first fused feature map are fused to obtain a second fused feature map; The target three-dimensional detection head constructed according to the target perception attention in the target detection model is used to determine the bounding box coordinates and the disease category for labeling the road disease based on the second fused feature map, so as to complete the road disease detection.

9. A road disease detecting apparatus characterized by comprising: The method comprises the steps of: The frequency balance feature determination module is configured to obtain an asphalt pavement image, input a feature map of the asphalt pavement image into a target detection model, decompose the feature map into different frequency bands through a frequency selection module in the target detection model, obtain frequency components of the different frequency bands, and determine frequency balance features based on the frequency components. The first convolutional feature map determination module is configured to determine a first convolutional feature map based on the frequency balance features, the feature map, and a dilation rate corresponding to each pixel of the feature map through a first convolutional module in the target detection model. The target convolutional feature map acquisition module is configured to determine a second convolutional feature map through a second convolutional module in the target detection model based on kernel weights, so as to convolve the first convolutional feature map and the second convolutional feature map to obtain a target convolutional feature map. The road disease detection module is configured to determine the bounding box coordinates and the disease category for labeling the road disease based on the target convolutional feature map through a target three-dimensional detection head constructed according to target perception attention in the target detection model, so as to complete the road disease detection.

10. An electronic device, comprising: The target perception attention comprises scale perception attention, spatial perception attention, and task perception attention. The memory is configured to store a computer program. The processor is configured to execute the computer program to implement the steps of the road disease detection method according to any one of claims 1 to 8.

11. A computer readable storage medium characterized by, The computer program is stored on the computer-readable storage medium and is executed by the processor to implement the steps of the road disease detection method according to any one of claims 1 to 8.

Images