CN-121981972-A - YOLOV 11-based highway asphalt pavement defect detection method, YOLOV-based highway asphalt pavement defect detection device and medium

CN121981972ACN 121981972 ACN121981972 ACN 121981972ACN-121981972-A

Abstract

The invention provides a method, a device and a medium for detecting defects of an expressway asphalt pavement based on YOLOV < 11 >, and belongs to the technical field of image processing. The method comprises the steps of obtaining defect data of the expressway asphalt pavement, marking the defect data to generate corresponding labels, combining the marked image data and the labels into a data set, dividing the data set into a training set, a verification set and a test set, constructing YOLOV a statistical space attention model, training the YOLOV statistical space attention model by using the training set, obtaining an optimal model through tuning super parameters of the verification set, collecting video data of the expressway asphalt pavement defect to be detected in real time in a moving process by a camera, transmitting the video data to the YOLOV statistical space attention model after training to detect defects, and outputting position, size and category information of the defects. Solves the problems of poor accuracy, low efficiency and the like of the defect detection of the expressway asphalt pavement under the surface condition of the complex asphalt pavement.

Inventors

WANG YONGCHAO
ZHENG LINJIANG
CHEN XINGZHOU

Assignees

重庆大学

Dates

Publication Date: 20260505
Application Date: 20260107

Claims (10)

1. The method for detecting the defects of the expressway asphalt pavement based on YOLOV a 11 is characterized by comprising the following steps of: Obtaining defect data of the expressway asphalt pavement, marking the defect data, and generating a corresponding label; Combining the marked image data and the label into a data set, and dividing the data set into a training set, a verification set and a test set; Constructing YOLOV11 statistical space attention model, wherein the model comprises a backbone network, a neck network and a head network, wherein the neck network is added with NNeam modules, the NNeam modules input output characteristics of a backbone network C2PSA module, the output characteristics of a neck network third C3k2 module are spliced with the output characteristics of the NNeam modules after convolution treatment and then input into a detection head through a neck network fourth C3k2 module, and the NNeam modules are statistical neural network attention modules irrelevant to parameters; training the YOLOV statistical space attention model by using a training set, and obtaining an optimal model by verifying set tuning super parameters; The camera collects video data of the defects of the highway asphalt pavement to be detected in real time in the moving process, the video is transmitted into a YOLOV statistical space attention model after training to detect the defects, and the position, the size and the category information of the defects are output.
2. The method for detecting the defects of the asphalt pavement of the expressway based on YOLOv as set forth in claim 1, wherein the P5 characteristics of the backbone network are sequentially processed by SPPF and C2PSA modules to obtain upstream characteristics; the upstream features are firstly up-sampled in a neck network, then the up-sampled P4 features with the same scale as a main network are spliced in a channel dimension and are input into a first C3k2 module of the neck network to be fused to obtain P4 output features, the P4 output features are up-sampled again, are spliced with the P3 features of the main network and are fused with a second C3k2 module of the neck network to obtain P3 output features, the P3 output features are down-sampled to the P4 scale through convolution and are spliced with the P4 output features, the third C3k2 module of the neck network is fused to obtain new P4 output features, the new P4 output features are down-sampled to the P5 scale through convolution and are spliced with the upstream features of the C2PSA, the P5 output features are fused with a fourth C3k2 module of the neck network, and finally the P3 output features, the new P4 output features and the P5 output features are input into a detection head.
3. The method for detecting defects of an expressway asphalt pavement based on YOLOv a according to claim 2, wherein the NNeam module sequentially includes: the space statistics and normalization unit is used for calculating space mean and variance channel by channel for the input features, introducing a stabilizing term for normalization to obtain basic attention energy, and further carrying out first-stage recalibration on the input features to obtain features x1; The phased multi-mode energy and gating unit is used for sequentially carrying out the processing of a geometric feature stage, a steady statistics stage and a frequency domain/topology/fractal stage based on the feature x1, carrying out parameter-free fusion and gating on corresponding energy in each stage, and sequentially obtaining features x2, x3 and x4; and the two-stage non-parameter fusion and final gating unit is used for performing stage fusion and Copula related fusion on the basic attention energy and the energy obtained by fusing each stage to obtain final fused attention energy, and gating the characteristic x4 through a Sigmoid function to obtain module output.
4. The method for detecting the defects of the asphalt pavement of the expressway based on YOLOV a in claim 3, wherein the geometric feature stage comprises the steps of calculating gradient energy, laplacian energy, structural tensor anisotropy and HessianBlobness based on the feature x1, carrying out unified normalization or bounding treatment on the energies, carrying out parameter-free fusion to obtain geometric feature energy, and carrying out gating on the feature x1 by using a Sigmoid function to obtain a feature x2.
5. The method for detecting defects of an expressway asphalt pavement based on YOLOV a 11 as set forth in claim 4, wherein the robust statistics stage includes calculating MAD robust scale, SURE risk approximation and Wasserstein residual error based on the feature x2, performing parameter-free fusion to obtain robust statistical energy, and gating the feature x2 with a Sigmoid function to obtain a feature x3.
6. The method for detecting the defects of the asphalt pavement of the expressway based on YOLOV, according to claim 5, wherein the frequency domain/topology/fractal stage comprises the steps of calculating high-frequency energy ratio, topology activity and fractal self-similarity indexes based on the characteristic x3, carrying out parameter-free fusion to obtain high-frequency energy, and carrying out gating on the characteristic x3 by using a Sigmoid function to obtain the characteristic x4.
7. The method for detecting defects of an expressway asphalt pavement based on YOLOV a claim 6, wherein in the two-stage non-parametric fusion and final gating unit, the stage fusion includes: respectively calculating energy average values of four stages of basic attention energy, geometric characteristic energy, steady statistical energy and high-frequency energy; based on the energy mean value, calculating the maximum entropy weight and the accompanying distribution weight of the energy mean value of each stage through a maximum entropy weight function and an accompanying distribution function respectively; respectively fusing the maximum entropy weight and the concomitantly distributed weight of each stage in parallel to obtain the fused weight of each stage; And weighting and summing the basic attention energy, the geometric feature energy, the steady statistical energy and the high-frequency energy by using the fusion weight to obtain the stage fusion energy.
8. The method for detecting defects in an expressway asphalt pavement based on YOLOV a of claim 7, wherein said correlation modeling is Copula correlation fusion, comprising: Performing PIT normalization processing on basic attention energy, geometric feature energy, steady statistical energy and high-frequency energy to obtain uniformly distributed variables; Calculating rank correlation coefficients among the uniformly distributed variables, and determining an average value; according to the average value, parameter mapping without parameters ArchimedeanCopula is carried out, if the average value is more than or equal to 0.5, the mapping type is Gumbel, otherwise, the mapping type is Clayton Aggregation is carried out by utilizing the selected mapping type, and relevant fusion energy is obtained; and averaging the stage fusion energy and the related fusion energy to obtain final fusion attention energy.
9. A highway asphalt pavement defect detection apparatus based on YOLOV a 11 comprising a processor and a memory storing program instructions, wherein the processor is configured to perform the highway asphalt pavement defect detection method based on YOLOV a according to any one of claims 1-8 when the program instructions are run.
10. A computer readable storage medium, having stored thereon a computer program which when executed by a processor implements the method of YOLOV-based highway asphalt pavement defect detection according to any one of claims 1-8.

Description

YOLOV 11-based highway asphalt pavement defect detection method, YOLOV-based highway asphalt pavement defect detection device and medium Technical Field The invention relates to a method, a device and a medium for detecting defects of an expressway asphalt pavement based on YOLOV < 11 >, and belongs to the technical field of computer vision. Background The highway asphalt pavement defect detection is a core task for guaranteeing structural safety in the field of constructional engineering, and the durability and the reliability of the highway in high-altitude areas are widely concerned. In the field, deep learning technology is becoming a key means for improving defect detection precision and efficiency, and is used for realizing accurate positioning and quantitative analysis of defects of asphalt pavement and prediction of defect development trend. Along with the continuous improvement of the requirements of the construction industry on the health monitoring of the asphalt highway, more front-edge technologies are introduced in the field of the defect detection of the asphalt pavement of the modern highway, so that the limitation of the traditional detection method is broken through, and the more efficient and intelligent detection is realized. At present, the traditional highway asphalt pavement defect detection method is numerous. The method based on threshold segmentation divides pixels in the image into defect and non-defect areas by setting a fixed or self-adaptive threshold, is simple in calculation, and can rapidly realize segmentation for images with single background and obvious defect characteristics. However, the surface condition of the asphalt pavement in the actual engineering is complex, and various interference factors exist, and the method is extremely sensitive to the selection of the threshold value, and incorrect setting of the threshold value can cause misjudgment or missed judgment of the defect. The edge detection algorithm determines the defect edge by detecting the abrupt change of the gray value in the image, has a certain effect on the defect with a clear edge, but in the scene of the expressway asphalt pavement, the defect edge is often blurred, and a discontinuous edge detection result is easy to generate, so that the complete identification of the defect is influenced. In highway asphalt pavement defect detection, semantic segmentation algorithms attempt to classify each pixel in an image to accurately separate a defective region from a background region. Although finer defect contour information can be provided theoretically, in practical application, due to the change of illumination conditions of an asphalt pavement and the irregularity of defect forms, the semantic segmentation model is easy to be mixed in category, and a non-defect area is misjudged as a defect, or otherwise, the detection accuracy is limited. FASTERRCNN such region-based proposed target detection algorithms determine the location and size of defects by generating candidate regions that may contain defects, and classifying and regressing the regions. However, in the highway asphalt pavement defect detection task, the defect shapes are changeable and often interweaved with each other, the candidate region generated by FASTERRCNN is difficult to accurately cover all defects, and the calculation cost is high, and the efficiency is low when a large number of images are processed. Therefore, how to provide a method for detecting defects of asphalt pavement of highway based on YOLOV a to improve the accuracy and stability of the detection is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a method, a device and a medium for detecting defects of an expressway asphalt pavement based on YOLOV < 11 >, which solve the problems of poor accuracy, low efficiency and the like of detecting defects of the expressway asphalt pavement under the surface condition of a complex asphalt pavement. The invention aims to achieve the aim, and the aim is achieved by the following technical scheme: A method for detecting defects of an expressway asphalt pavement based on YOLOV comprises the following steps: Obtaining defect data of the expressway asphalt pavement, marking the defect data, and generating a corresponding label; Combining the marked image data and the label into a data set, and dividing the data set into a training set, a verification set and a test set; Constructing YOLOV11 statistical space attention model, wherein the model comprises a backbone network, a neck network and a head network, wherein the neck network is added with NNeam modules, the NNeam modules input output characteristics of a backbone network C2PSA module, the output characteristics of a neck network third C3k2 module are spliced with the output characteristics of the NNeam modules after convolution treatment and then input into a detection head through a neck network fourth C3k2 module, and th