Visual intelligence-based aged asphalt stripping method and system

Abstract

The invention belongs to the field of image recognition, and particularly relates to an aged asphalt stripping method and system based on visual intelligence, and the method comprises the steps: collecting light, medium and heavy and mixed aged asphalt pavement images, carrying out the preprocessing, dividing a data set, and marking the cracking features; extracting regional features through a sliding window, counting crack parameters and calculating a feature proportion coefficient; and constructing a multi-scale network containing convolution units with different scales, dynamically adjusting the unit ratio according to a ratio coefficient, fusing features, and outputting three types of crack segmentation maps and detection frames through parallel branches. And adopting a combined loss function training model, and dynamically adjusting the learning rate to convergence based on the verification set. Pre-processing a to-be-detected image, inputting into the model, and post-processing to output an aged asphalt stripping suggestion. According to the technology, the detection limitation of a traditional network in a mixed aging scene is solved, small-scale micro cracks and large-scale cracks are synchronously captured, the detection accuracy is improved, and suggestions fitting the reality are output.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, specifically relating to a method and system for stripping aging asphalt based on visual intelligence. Background Technology

[0002] In asphalt pavement aging detection, when mildly aged microcracks coexist with moderately to severely aged large-area cracks, the fixed receptive field of traditional convolutional neural networks (CNNs) is difficult to capture small-scale details and large-scale regional features simultaneously, leading to missed detections (microcracks) or over-detections (misjudgments of normal areas). For example, "microcrack detection" for mildly aged pavement requires small-scale convolutional kernels, but it may miss "large-area cracks" for moderately to severely aged pavement. "Region segmentation" for moderately to severely aged pavement requires large-scale receptive fields, but it may misjudge mildly aged microcracks as "noise".

[0003] Furthermore, traditional methods only extract features statically without considering the dynamic evolution and correlation of different cracks, resulting in low detection accuracy in complex scenarios and directly affecting the scientific nature of decisions regarding the stripping of aging asphalt. Summary of the Invention

[0004] The purpose of this invention is to provide a visual intelligence-based method and system for stripping aged asphalt, in order to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A visual intelligence-based method for stripping aged asphalt includes the following steps:

[0007] Images of asphalt pavement under mild, moderate, and mixed aging scenarios were collected. The asphalt pavement images were preprocessed and divided into training, validation, and test sets. Crack features of different degrees were labeled to generate annotation files.

[0008] Samples are extracted from the validation set, regional features are extracted through a sliding window, parameters of microcracks, medium-area cracks and large-area cracks are statistically analyzed, and feature proportion coefficients are calculated.

[0009] A multi-scale network containing convolutional units of different scales is constructed. The ratio of convolutional units of different scales is dynamically adjusted according to the feature ratio coefficient. After feature fusion, the segmentation map and detection box of three types of cracks are output through parallel branches.

[0010] The model is trained using a combined loss function, and the learning rate is dynamically adjusted based on the validation set index until the convergence criteria for microcrack, medium-scale crack segmentation, and large-area crack detection are met.

[0011] The target area image to be detected is preprocessed and then input into the model. The output results are then postprocessed to provide a suggestion on whether to strip the aged asphalt.

[0012] Furthermore, the asphalt pavement image preprocessing includes: using Gaussian filtering to remove high-frequency noise, using the CLAHE algorithm to improve local contrast, using feature matching to correct perspective transformation so that the tilt image error is less than a threshold, and cropping to uniform pixels.

[0013] Furthermore, the annotation of crack features of different degrees includes: pixel-level annotation of microcracks with a line width of not less than 1 pixel, polygon annotation of cracks with an area of ​​medium scale, and polygon annotation with center point annotation of cracks with a large area.

[0014] Furthermore, regional features are extracted using a sliding window method, and parameters for microcracks, medium-scale cracks, and large-scale cracks are statistically analyzed. This includes: extracting local regions using the sliding window method, and statistically analyzing the total length and average width of microcracks, the total area of ​​medium-scale cracks, and the total area of ​​large-scale cracks.

[0015] Furthermore, the feature proportion coefficients are calculated, including: the total length L1 and average width W1 of the microcracks, the total area S2 of the mesoscale cracks, and the total area S3 of the large-area cracks. The proportion of microcracks is defined as K1 = (L1 × W1) / total pixels of the image, the proportion of mesoscale cracks is K2 = S2 / total pixels of the image, and the proportion of large-area cracks is K3 = S3 / total pixels of the image. The proportion coefficients are calculated, with the first proportion coefficient α′ = K1 / K3 and the second proportion coefficient β = K2 / K3.

[0016] Furthermore, the multi-scale network includes: an input layer connected to a multi-scale feature extraction module, which contains three sets of convolutional units. These three sets of convolutional units use 3×3, 7×7, and 15×15 convolutional kernels, respectively, to extract small, medium, and large-scale features. Each set of convolutional units is equipped with a BN layer, a ReLU activation function, and a 2×2 max pooling layer. The multi-scale feature extraction module is connected to a feature fusion layer, which adopts a cross-scale pyramid pooling structure. It performs 1×1, 2×2, 4×4, and 8×8 pooling on the feature maps of each scale output by the three sets of convolutional units, respectively. The pooling results are concatenated with the original feature maps and then compressed to 256 channels by a 1×1 convolution. The feature fusion layer is connected to the detection branch in the dual-branch architecture. The detection branch has three parallel branches that output micro-crack segmentation maps, medium-scale crack segmentation maps, and large-area crack detection boxes, respectively.

[0017] Furthermore, the dynamic adjustment of the ratio of convolutional units at different scales based on the feature proportion coefficient values ​​includes: First, the ratio of small-scale, medium-scale, and large-scale units has a default setting. Based on the calculated proportion coefficients, the first proportion coefficient α′ and the second proportion coefficient β, the number of small-scale, medium-scale, and large-scale convolutional units used in the multi-scale feature extraction module is adjusted respectively. When the first proportion coefficient α′ ≥ 2.0 and the second proportion coefficient β ≥ 1.0, the ratio of small-scale, medium-scale, and large-scale units is set to 4:3:1; when 0.5 < the first proportion coefficient α′ < 2.0 and 0.3 < the second proportion coefficient β < 1.0, the ratio of small-scale, medium-scale, and large-scale units is set to 2:2:2; when the first proportion coefficient α′ ≤ 0.5 and the second proportion coefficient β ≤ 0.3, the ratio of small-scale, medium-scale, and large-scale units is set to 1:3:4. In other cases, the ratio of small-scale, medium-scale, and large-scale units adopts the previously existing default setting.

[0018] Furthermore, the parallel branching output of segmentation maps and detection boxes for three types of cracks includes: the detection branch has three parallel branches. The first branch outputs a microcrack segmentation map through 1×1 convolution and Softmax activation function, and the first branch includes two categories: microcracks and background. The second branch outputs a mesoscale crack segmentation map through 1×1 convolution and Softmax activation function, and the second branch includes two categories: mesoscale cracks and background. The third branch outputs a large-area crack detection box through bounding box regression and category confidence calculation, and the third branch includes the detection box coordinates and the corresponding confidence score.

[0019] Furthermore, the model training method using combined loss functions includes: different loss functions are used for cracks of different degrees, and the total loss is calculated by weighted summation of the loss functions corresponding to different crack degrees.

[0020] An electronic device includes one or more processors, which execute the above-described vision-based intelligent method for stripping aged asphalt by running programs stored in a memory.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described visual intelligence-based method for stripping aged asphalt.

[0022] Beneficial effects include:

[0023] This invention effectively addresses the limitations of traditional convolutional neural networks in detecting both mildly aged microcracks and moderately severe large-area cracking. By constructing a multi-scale feature extraction network and dynamically adjusting the ratio of convolutional units at different scales, it achieves simultaneous capture of details from small-scale microcracks and features from large-scale cracked areas, avoiding the missed or over-detection problems caused by a single-scale receptive field. Combining cross-scale feature fusion with targeted parallel detection branches, it can accurately identify the distribution characteristics of microcracks, medium-scale cracks, and large-area cracks. By training and optimizing the model using combined loss functions, the detection accuracy for aging cracks of different degrees is further improved. Finally, through quantitative analysis of the detection results, it can output practical suggestions for aging asphalt stripping. This invention addresses the technical pain points of mixed scenarios in asphalt pavement aging detection, achieving accurate detection and scientific decision-making through multi-dimensional technological innovation. Specific effects are as follows:

[0024] Overcoming the limitations of mixed aging scene detection: By constructing a multi-scale feature extraction network with 3×3, 7×7, and 15×15 convolutional kernels, and dynamically adjusting the convolutional unit ratio (adaptively switching from 4:3:1 to 1:3:4) in combination with the feature ratio coefficient, the problem of traditional fixed-scale networks missing microcracks and over-detecting large-area cracks is solved. It achieves simultaneous and accurate capture of small-scale microcrack details and large-scale crack area features, significantly improving the detection integrity in mixed scenes.

[0025] Improve feature representation and detection accuracy: Employ a cross-scale pyramid pooling feature fusion layer to perform multi-scale pooling and concatenation compression on feature maps at various scales, fully exploring the correlation between local details and global distribution of cracks; Combined with three parallel detection branches (micro-crack segmentation, medium-scale crack segmentation, and large-area crack detection box output), and combined with a combined loss function, significantly improve the detection accuracy of cracks of different degrees.

[0026] Adapting to dynamically evolving aging characteristics: Innovatively introducing temporal evolution modeling and cross-scale coupling analysis, by calculating the microcrack propagation rate, crack expansion rate and connectivity, and combining fractional calculus to capture the "memory" of features, dynamic weights are generated to optimize the ratio of convolutional units, enabling the network to accurately adapt to the gradual evolution of cracks from microcracks to large-area cracks, avoiding adaptation bias caused by static feature analysis.

[0027] Scientifically sound peeling decision recommendations: By quantitatively analyzing the test results (indicators such as total microcrack length and total crack area), a graded decision standard is established. Recommendations such as daily maintenance, local repair, and milling peeling are provided for scenarios with no significant aging, mild, moderate, and severe aging, respectively, so as to achieve precise alignment between test results and engineering practice. Attached Figure Description

[0028] Figure 1This is a flowchart of the visual intelligence-based method for stripping aged asphalt according to the present invention.

[0029] Figure 2 This is a system block diagram of the visual intelligence-based aging asphalt stripping system of the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0031] This application discloses a visual intelligence-based method for stripping aged asphalt, such as... Figure 1 As shown, the steps include:

[0032] S1. Collect asphalt pavement images in light, moderate and heavy aging scenarios and mixed aging scenarios. Preprocess the asphalt pavement images, divide them into training set, validation set and test set, and annotate crack features of different degrees to generate annotation files.

[0033] S2. Extract samples from the validation set, extract regional features through a sliding window, statistically analyze the parameters of micro-cracks, medium-area cracks and large-area cracks, and calculate the feature proportion coefficient;

[0034] S3. Construct a multi-scale network containing convolutional units of different scales, dynamically adjust the ratio of convolutional units of different scales according to the feature ratio coefficient value, fuse features, and output segmentation maps and detection boxes of three types of cracks through parallel branches.

[0035] S4. The model is trained using a combined loss function, and the learning rate is dynamically adjusted based on the validation set index until the convergence criteria for microcrack, medium-scale crack segmentation and large-area crack detection are met.

[0036] S5. Input the preprocessed image of the target area into the model, postprocess the output results, and output a suggestion on whether to strip the aged asphalt.

[0037] For collecting images of asphalt pavement under light, moderate to heavy, and mixed aging scenarios, the asphalt pavement images were preprocessed and divided into training, validation, and test sets. In practice, a combination of vehicle-mounted high-definition cameras and drone aerial photography equipment was used to collect images, covering light aging scenarios (microcracks ≤0.5mm accounted for ≥70%), moderate to heavy aging scenarios (large-area cracks ≥0.1㎡ accounted for ≥70%), and mixed aging scenarios (microcracks and large-area cracks coexist, each accounting for ≥30%). The images collected for each scenario included lighting conditions such as strong light, backlight, and cloudy days, weather conditions such as dryness, humidity, and light water accumulation, and pavement material variables such as asphalt concrete and modified asphalt. A 5×5 convolution kernel was used during preprocessing. Gaussian filtering removes high-frequency noise, the CLAHE algorithm enhances local contrast, and feature matching corrects perspective transformation to ensure tilted image errors are below a threshold. After cropping to uniform pixels, the images are divided into training, validation, and test sets in a 7:2:1 ratio. For labeling crack features of varying degrees, LabelMe is used to generate annotation files. Microcracks are annotated at the pixel level with a line width of at least 1 pixel. Medium-scale cracks with an area of ​​0.01-0.1㎡ are annotated with polygons, and large-area cracks with an area of ​​at least 0.1㎡ are annotated with polygons and center points. After annotation, a JSON annotation file containing feature coordinates, area, and type information is generated.

[0038] For samples drawn from the validation set, regional features are extracted using a sliding window. Specifically, 30% of the samples are randomly selected from the validation set as analysis samples, with a total number of no less than 1000 images. A sliding window method with a window size of 256×256 and a stride of 128 is used to extract local regions. The total length and average width of microcracks, the total area of ​​mesoscale cracks, and the total area of ​​large-area cracks are statistically analyzed. For the statistical parameters of microcracks, mesoscale cracks, and large-area cracks, feature proportion coefficients are calculated. Specifically, the total length L1 and average width W1 of microcracks, the total area S2 of mesoscale cracks, and the total area S3 of large-area cracks are statistically analyzed. The proportions of microcracks are defined as K1 = (L1×W1) / total pixels of the image, the proportion of mesoscale cracks is K2 = (S2) / total pixels of the image, and the proportion of large-area cracks is K3 = (S3) / total pixels of the image. The proportion coefficients are calculated as follows: the first proportion coefficient α′ = K1 / K3, and the second proportion coefficient β = K2 / K3.

[0039] To construct a multi-scale network containing convolutional units of different scales, the implementation specifically involves building an input layer connected to a multi-scale feature extraction module. This module contains three sets of convolutional units, employing 3×3 convolutional kernels (stride 1, padding 1, 64 channels), 7×7 convolutional kernels (stride 1, padding 3, 64 channels), and 15×15 convolutional kernels (stride 1, padding 7, 64 channels) to extract small, medium, and large-scale features, respectively. Each set of units is equipped with a BN layer, a ReLU activation function, and a 2×2 max-pooling layer (stride 2). The block-connected feature fusion layer employs a cross-scale pyramid pooling structure. It performs 1×1, 2×2, 4×4, and 8×8 pooling on the feature maps output from the three units, respectively. The pooling results are then concatenated with the original feature maps and compressed to 256 channels using a 1×1 convolution. The phrase "performing 1×1, 2×2, 4×4, and 8×8 pooling operations on small-scale, medium-scale, and large-scale feature maps respectively" means that each scale of feature map must independently undergo these four pooling operations, rather than each scale corresponding to a separate pooling operation. Specifically, small-scale feature maps undergo 1×1, 2×2, 4×4, and 8×8 pooling sequentially, medium-scale feature maps undergo the same four pooling operations, and large-scale feature maps repeat this process. By performing multi-scale pooling on the feature maps at each scale, contextual information of different ranges can be captured (such as local details of microcracks and global distribution of large-area cracks); the feature fusion layer connects the detection branch in the dual branch, and the detection branch is set with 3 parallel branches, which output microcrack segmentation map, medium-scale crack segmentation map, and large-area crack detection box respectively.

[0040] For dynamically adjusting the ratio of convolutional units at different scales based on feature proportion coefficients, the specific implementation involves using the calculated proportion coefficients α′=K1 / K3 and β=K2 / K3. Firstly, there is a default setting for the ratio of small-scale, medium-scale, and large-scale units, such as a 2:2:2 ratio. Then, based on the calculated proportion coefficients α′ and β′, the number of small-scale, medium-scale, and large-scale convolutional units in the multi-scale feature extraction module, using 3×3, 7×7, and 15×15 convolutional kernels respectively, is adjusted. When the first proportion... When the coefficient α′≥2.0 and the second scaling factor β≥1.0, the ratio of small-scale units, medium-scale units, and large-scale units is set to 4:3:1; when 0.5<1.5<1.0<1.5<1.0<2.0<1.5<2.0<1.3<1.0<1.0<2.2:2<2.2<2.3<1.5<1.5<1.5<2.0 ...

[0041] For feature fusion, the implementation involves performing 1×1, 2×2, 4×4, and 8×8 pooling operations on the small-scale, medium-scale, and large-scale feature maps output by the multi-scale feature extraction module, respectively. The pooling results at each scale are then concatenated with the corresponding original feature maps to form a fusion feature with 448 channels. This fusion feature is then compressed to 256 channels using a 1×1 convolution to complete feature fusion. For outputting segmentation maps and detection boxes for three types of cracks through parallel branches, the implementation involves setting up three parallel branches in the detection branch. The first branch outputs a micro-crack segmentation map using a 1×1 convolution and a Softmax activation function, containing two categories: micro-cracks and background. The second branch outputs a medium-scale crack segmentation map using a 1×1 convolution and a Softmax activation function, containing two categories: medium-scale cracks and background. The third branch outputs a large-area crack detection box using bounding box regression and class confidence calculation, including the detection box coordinates (x1, y1, x2, y2) and the corresponding confidence score.

[0042] For training the model using a combined loss function, specifically, a combined loss function is used: Dice loss with a weight of 0.3 is used for microcrack segmentation; Intersection over Union (IoU) loss with a weight of 0.3 is used for mesoscale crack segmentation; and CIoU loss with a weight of 0.4 is used for large-area crack detection. The total loss is calculated as: Total Loss = 0.3 × Dice Loss + 0.3 × IoU Loss + 0.4 × CIoU Loss. The learning rate is dynamically adjusted based on validation set metrics until the convergence criteria for microcrack, mesoscale crack segmentation, and large-area crack detection are met. In practice, the Adam optimizer with an initial learning rate of 0.0001 (β1 = 0.9, β2 = 0.999) was used for training. The training was conducted for 50 rounds with a batch size of 8. Every 5 rounds, the microcrack segmentation mIoU, mesoscale crack segmentation mIoU, and large-area crack detection mAP@0.5 of the validation set were calculated. When the above indicators did not improve for 8 consecutive rounds, the learning rate was decayed to 1 / 5 of the current value. After decaying 3 times, the training was terminated until the microcrack segmentation mIoU ≥ 0.85, the mesoscale crack segmentation mIoU ≥ 0.80, and the large-area crack detection mAP@0.5 ≥ 0.88 of the validation set.

[0043] For the target area image to be detected, after preprocessing, it is input into the model. Specifically, in the implementation, the target area image of the asphalt pavement is used. A Gaussian filter with a 5×5 convolution kernel and a standard deviation of 1.2 is used to remove high-frequency noise. The local contrast is improved by the CLAHE algorithm with cliplimit=2.0 and a grid size of 8×8. SIFT feature matching is used to correct the perspective transformation so that the tilt image error is ≤0.5°. The image is uniformly cropped to 1024×1024 pixels. Then, the preprocessed image is input into the trained multi-scale network.

[0044] Post-processing of the output results includes recommendations on whether to strip the aged asphalt. Specifically, this involves statistically analyzing the post-processed microcrack segmentation map, mesoscale crack segmentation map, and large-area crack detection frame results to calculate the total length of microcracks, the total area of ​​mesoscale cracks, and the total area of ​​large-area cracks. If the total length of microcracks is <500mm, the total area of ​​mesoscale cracks is <0.05㎡, and the total area of ​​large-area cracks is <0.1㎡, it is considered that there is no significant aging, and stripping is not required; only routine maintenance is recommended. If the total length of microcracks is ≥500mm and the total area of ​​mesoscale cracks is < If the total area of ​​large-scale cracks is less than 0.1㎡, it is considered mild aging. It is recommended to spray recycling agent or apply a thin overlay to the micro-cracked areas for repair, without stripping. If the total area of ​​medium-scale cracks is greater than or equal to 0.05㎡ and less than 0.1㎡, it is considered moderate aging. It is recommended to remove the loose mixture (including aged asphalt and aggregate) in the corresponding area for local stripping and repair. If the total area of ​​large-scale cracks is greater than or equal to 0.1㎡, or the total area of ​​medium-scale cracks is greater than or equal to 0.1㎡, it is considered severe aging. It is recommended to mill and strip all the aged asphalt mixture in the area and repave with new asphalt mixture.

[0045] In practice, the original method of calculating the "feature proportion coefficient" using only the static ratio of area, length, and pixels has limitations. Asphalt aging is a gradual process (microcracks may expand into mesoscale cracks, and then develop into large-area cracks). The original method does not capture the temporal evolution relationship of different features. The dense distribution of microcracks may accelerate the formation of large-area cracks, and the original method treats the three types of features as independent entities, failing to reflect their mutual influence. Therefore, in further implementation, this application proposes a new technology to replace the "feature proportion coefficient" calculation. Through temporal evolution modeling and cross-scale coupling analysis, the dynamic correlation and detection reliability of features are fused. This application proposes a method for adjusting the ratio of convolutional units based on dynamic weights, including the following steps:

[0046] For the validation set samples, feature time series are constructed according to the acquisition time series (images of the same road segment at different times) to capture the aging evolution trend:

[0047] Let the characteristic of time t (t = 1, 2, ..., T, where T is the time step) within the sliding window be:

[0048] Microcracks: Total length L1(t), average width W1(t), propagation rate v1(t) = [L1(t) - L1(t-1)] / Δt (Δt is the time interval, unit: days);

[0049] Mesoscale cracking: Total area S2(t), morphological expansion rate r2(t) = [S2(t) - S2(t-1)] / [S2(t-1)·Δt] (relative expansion rate);

[0050] Large-area cracks: Total area S3(t), connectivity c3(t) = actual number of connected regions / theoretical maximum number of regions (the higher the value, the more contiguous the cracks). L1(t-1): represents the total length of microcracks within the sliding window at time t-1 in the time series (i.e., the time before the current time t). S2(t-1): represents the total area of ​​mesoscale cracks within the sliding window at time t-1 in the time series.

[0051] Mutual information and partial correlation coefficients are used to measure the coupling strength between features (such as the influence of microcrack propagation on large-area cracking).

[0052] Define the coupling coefficient matrix C = [c i,j ] 3×3 , where c i,j Indicates the strength of the influence of the i-th feature on the j-th feature:

[0053] F i For the i-th type of feature (i=1: microcracks, i=2: mesoscale cracks, i=3: large-area cracks); I(F i ,F j Mutual information (ρ) measures the non-linear correlation between two features (the larger the value, the stronger the correlation). i,j The partial correlation coefficient is the net correlation coefficient after excluding interference from third-party features (|ρ i,j |∈[0,1]).

[0054] Considering the "memory" nature of aging characteristics (historical states influence the current state), fractional calculus is used to handle long-term effects: for the current value F of the i-th type of characteristic... i (T), whose fractional derivative (order α∈(0,1), representing the memory depth) is:

[0055] In the formula: Γ(·) is the gamma function, a is the initial time, and τ is the integration variable.

[0056] The final weights of the three types of features are calculated by combining the evolution rate, coupling degree, and fractional-order memory:

[0057] Introducing detection confidence θ i(For example, for microcracks, θ1 = 0.7 under backlight and θ1 = 0.95 under normal lighting), corrected eigenvalues:

[0058] Weighting formula:

[0059]

[0060] In the formula: The values ​​are the corrected eigenvalues. g1, g2, and g3 are the dynamic weights for microcracks, mesoscale cracks, and large-area cracks, respectively.

[0061] Adjusting the ratio of convolutional units based on dynamic weights:

[0062] Define new scaling factors, the first new scaling factor α”, the second new scaling factor β”: α” = g1 / g3, β” = g2 / g3, but because g1 / g2 / g3 contains evolution trends and coupling relationships, it can more accurately adapt to mixed aging scenarios (such as when microcracks propagate rapidly, automatically increasing the proportion of small-scale convolutional units).

[0063] The ratio of convolutional units is adjusted based on dynamic weights: First, the ratio of small-scale, medium-scale, and large-scale units has a default setting, such as 2:2:2. Then, based on the calculated ratio coefficients (first new ratio coefficient α”, second new ratio coefficient β”), the number of small-scale, medium-scale, and large-scale convolutional units using 3×3, 7×7, and 15×15 convolutional kernels respectively in the multi-scale feature extraction module is adjusted. When the first new ratio coefficient α” ≥ 2.0 and the second new ratio coefficient β” ≥ 1.0, the number of small-scale, medium-scale, and large-scale convolutional units is adjusted accordingly. The ratio of small-scale units, medium-scale units, and large-scale units is set to 4:3:1. When 0.5 < the first new scale factor α” < 2.0 and 0.3 < the second new scale factor β” < 1.0, the ratio of small-scale units, medium-scale units, and large-scale units is set to 2:2:2. When the first new scale factor α” ≤ 0.5 and the second new scale factor β” ≤ 0.3, the ratio of small-scale units, medium-scale units, and large-scale units is set to 1:3:4. In other cases, the ratio of small-scale units, medium-scale units, and large-scale units adopts the previously existing default setting.

[0064] This application also provides an embodiment of an electronic device. The electronic device is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors or processing units, memory, and buses connecting different components (including memory and processing units).

[0065] A bus refers to one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. Examples of these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MCA) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0066] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0067] The memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic devices may further include other removable / non-removable, volatile / non-volatile computer device storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media.

[0068] The electronic device can also communicate with one or more external devices (e.g., keyboard, pointing device, camera, etc.), may include a display, and may communicate with one or more devices that enable a user to interact with the electronic device, and / or with any device that enables the electronic device to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via an input / output (I / O) interface. Furthermore, the electronic device can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN) and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. The processor executes various functional applications and data processing by running programs stored in memory, such as implementing the visual intelligence-based aging asphalt stripping method provided in the above embodiments of the present invention, which includes: preprocessing asphalt pavement images, dividing them into training sets, validation sets, and test sets, and generating annotation files by labeling different degrees of crack features;

[0069] S2. Extract samples from the validation set, extract regional features through a sliding window, statistically analyze the parameters of micro-cracks, medium-area cracks and large-area cracks, and calculate the feature proportion coefficient;

[0070] S3. Construct a multi-scale network containing convolutional units of different scales, dynamically adjust the ratio of convolutional units of different scales according to the feature ratio coefficient value, fuse features, and output segmentation maps and detection boxes of three types of cracks through parallel branches.

[0071] S4. The model is trained using a combined loss function, and the learning rate is dynamically adjusted based on the validation set index until the convergence criteria for microcrack, medium-scale crack segmentation and large-area crack detection are met.

[0072] S5. Input the preprocessed image of the target area into the model, postprocess the output results, and output a suggestion on whether to strip the aged asphalt.

[0073] This application also provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, it implements the visual intelligence-based aging asphalt stripping method provided in the embodiments of the present invention. The method includes: preprocessing an asphalt pavement image, dividing it into a training set, a validation set, and a test set, and generating a labeling file by labeling different degrees of crack features.

[0074] S2. Extract samples from the validation set, extract regional features through a sliding window, statistically analyze the parameters of micro-cracks, medium-area cracks and large-area cracks, and calculate the feature proportion coefficient;

[0075] S3. Construct a multi-scale network containing convolutional units of different scales, dynamically adjust the ratio of convolutional units of different scales according to the feature ratio coefficient value, fuse features, and output segmentation maps and detection boxes of three types of cracks through parallel branches.

[0076] S4. The model is trained using a combined loss function, and the learning rate is dynamically adjusted based on the validation set index until the convergence criteria for microcrack, medium-scale crack segmentation and large-area crack detection are met.

[0077] S5. Input the preprocessed image of the target area into the model, postprocess the output results, and output a suggestion on whether to strip the aged asphalt.

[0078] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0079] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit programs for use by or in conjunction with an instruction execution system, system, or device.

[0080] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0081] This application also discloses a visual intelligence-based system for stripping aged asphalt, such as... Figure 2 As shown, it includes:

[0082] The image preprocessing module is used to standardize the acquired raw images, providing high-quality data for subsequent model input;

[0083] The data annotation and management module is used to perform feature annotation and dataset partitioning on preprocessed images.

[0084] The feature analysis and dynamic weight calculation module is used to extract samples from the validation set (30% randomly selected, with a total of ≥1000 samples) and calculate the dynamic weights used to adjust the network structure.

[0085] Multi-scale convolutional neural networks;

[0086] A multi-scale feature extraction and detection network module is used for a multi-scale convolutional neural network based on dynamic weight adaptive adjustment.

[0087] The model training and optimization module is used to train the network model based on the training set and dynamically optimize the parameters using the validation set.

[0088] The detection reasoning and decision suggestion module is used to assess the aging status of the image to be detected and output stripping suggestions.

Claims

1. A visual intelligence-based method for stripping aging asphalt, characterized in that, Including the following steps: Images of asphalt pavement under mild, moderate, and mixed aging scenarios were collected. The asphalt pavement images were preprocessed and divided into training, validation, and test sets. Crack features of different degrees were labeled to generate annotation files. Samples are extracted from the validation set, regional features are extracted through a sliding window, parameters of microcracks, medium-area cracks and large-area cracks are statistically analyzed, and feature proportion coefficients are calculated. A multi-scale network containing convolutional units of different scales is constructed. The ratio of convolutional units of different scales is dynamically adjusted according to the feature ratio coefficient. After feature fusion, the segmentation map and detection box of three types of cracks are output through parallel branches. The model is trained using a combined loss function, and the learning rate is dynamically adjusted based on the validation set index until the convergence criteria for microcrack, medium-scale crack segmentation, and large-area crack detection are met. The target area image to be detected is preprocessed and then input into the model. The output results are then postprocessed to provide a suggestion on whether to strip the aged asphalt.

2. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The preprocessing of asphalt pavement images includes: using Gaussian filtering to remove high-frequency noise, using the CLAHE algorithm to improve local contrast, using feature matching to correct perspective transformation so that the tilt image error is less than a threshold, and cropping to uniform pixels.

3. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The features of cracks of different degrees are marked as follows: microcracks are marked at the pixel level with a line width of not less than 1 pixel; cracks with an area of ​​medium scale are marked with polygons; and cracks with an area of ​​large area are marked with polygons plus center point.

4. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, Regional features were extracted using a sliding window method, and parameters for microcracks, medium-scale cracks, and large-scale cracks were statistically analyzed. This included: extracting local regions using the sliding window method, and statistically analyzing the total length and average width of microcracks, the total area of ​​medium-scale cracks, and the total area of ​​large-scale cracks.

5. The visual intelligence-based method for stripping aged asphalt according to claim 1, characterized in that, The characteristic proportion coefficients are calculated, including: the total length L1 and average width W1 of microcracks, the total area S2 of mesoscale cracks, and the total area S3 of large-area cracks. The proportion of microcracks is defined as K1 = (L1 × W1) / total pixels of the image, the proportion of mesoscale cracks is K2 = S2 / total pixels of the image, and the proportion of large-area cracks is K3 = S3 / total pixels of the image. The proportion coefficients are calculated, with the first proportion coefficient α′ = K1 / K3 and the second proportion coefficient β = K2 / K3.

6. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The multi-scale network includes: an input layer connected to a multi-scale feature extraction module, which contains three sets of convolutional units. These units use 3×3, 7×7, and 15×15 kernels to extract small, medium, and large-scale features, respectively. Each set of convolutional units is equipped with a BN layer, a ReLU activation function, and a 2×2 max pooling layer. The multi-scale feature extraction module is connected to a feature fusion layer, which uses a cross-scale pyramid pooling structure. It performs 1×1, 2×2, 4×4, and 8×8 pooling on the feature maps of each scale output by the three sets of convolutional units, respectively. The pooling results are concatenated with the original feature maps and then compressed to 256 channels using a 1×1 convolution. The feature fusion layer is connected to the detection branch in the dual-branch architecture. The detection branch has three parallel branches that output micro-crack segmentation maps, medium-scale crack segmentation maps, and large-area crack detection boxes, respectively.

7. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The dynamic adjustment of the ratio of convolutional units at different scales based on the feature proportion coefficient values ​​includes: First, the ratio of small-scale, medium-scale, and large-scale units has a default setting. Based on the calculated proportion coefficients, the first proportion coefficient α′ and the second proportion coefficient β, the number of small-scale, medium-scale, and large-scale convolutional units used in the multi-scale feature extraction module is adjusted respectively. When the first proportion coefficient α′ ≥ 2.0 and the second proportion coefficient β ≥ 1.0, the ratio of small-scale, medium-scale, and large-scale units is set to 4:3:1; when 0.5 < the first proportion coefficient α′ < 2.0 and 0.3 < the second proportion coefficient β < 1.0, the ratio of small-scale, medium-scale, and large-scale units is set to 2:2:2; when the first proportion coefficient α′ ≤ 0.5 and the second proportion coefficient β ≤ 0.3, the ratio of small-scale, medium-scale, and large-scale units is set to 1:3:

4. In other cases, the ratio of small-scale, medium-scale, and large-scale units adopts the previously existing default setting.

8. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The parallel branch outputs segmentation maps and detection boxes for three types of cracks, including: the detection branch has three parallel branches. The first branch outputs a microcrack segmentation map through 1×1 convolution and Softmax activation function, which includes two categories: microcracks and background. The second branch outputs a mesoscale crack segmentation map through 1×1 convolution and Softmax activation function, which includes two categories: mesoscale cracks and background. The third branch outputs a large-area crack detection box through bounding box regression and category confidence calculation, which includes the detection box coordinates and the corresponding confidence score.

9. The method for stripping aging asphalt based on visual intelligence according to claim 1, characterized in that, The model training method using combined loss functions includes: different loss functions are used for different degrees of cracks, and the total loss is calculated by weighted summation of the loss functions corresponding to different degrees of cracks.

Images