CN112964712A

CN112964712A - Method for rapidly detecting state of asphalt pavement

Info

Publication number: CN112964712A
Application number: CN202110159801.2A
Authority: CN
Inventors: 但汉成; 柏格文; 祝志恒; 山宏宇
Original assignee: Central South University
Current assignee: Central South University
Priority date: 2021-02-05
Filing date: 2021-02-05
Publication date: 2021-06-15

Abstract

A method for rapidly detecting the state of an asphalt pavement comprises the following steps: step one, drilling a core on a road to obtain a cylindrical road test piece; secondly, shooting the side surface of the cylindrical road test piece by using an area-array camera around one circle of the cylindrical road test piece, matching, aligning and reconstructing all photos, and synthesizing panoramic images to obtain a synthetic image; marking the composite; step three, identifying the synthesized image by adopting an artificial neural network, and optimizing by adopting a Gaussian weight method and a secondary marking method to obtain an identification result image; performing binarization calculation and processing on the identification result image, filling cracks and cavities, and performing particle segmentation on adhesion among the aggregate particles; and analyzing the morphological characteristics and spatial position distribution of the aggregate particles in the identification result after the identification result is processed. The invention realizes the rapid, accurate and quantitative analysis of various parameter performances of the asphalt mixture, thereby achieving the purpose of rapidly detecting the road state.

Description

Method for rapidly detecting state of asphalt pavement

Technical Field

The invention belongs to the technical field of civil engineering roads, and particularly relates to a method for rapidly detecting the state of an asphalt pavement.

Background

The improvement of the national transportation level can not leave the construction of a road network, after large-scale capital construction of a large amount of roads has passed, the impact of the road reservation amount on a detection technology and a maintenance technology can not be small and varied, and the engineering quality and the road surface diseases of the road surface need to be evaluated quickly and accurately. The method for judging the engineering quality by utilizing the image is one of visual and effective methods, establishes a relationship between the visual image and the performance or damage of the geotechnical building, and is applied to the aspects of damage detection, state monitoring, performance prediction and the like. In the field of road surfaces, image recognition is used as a computer-aided method to improve research and engineering efficiency. At present, image acquisition methods for indoor tests are few, and CT scanning is often adopted to judge the change condition of internal cracks or gaps after a mechanical property test is carried out on an asphalt mixture test piece; after the mixture test piece is sliced, an image is shot and binarization processing is carried out on the mixture test piece, so that the internal information of the test piece can be obtained, the morphological characteristics of the aggregate and the damage of the mixture such as cracks are analyzed, and the relation with the mechanical property is established. The leading-edge research theory and the method have a plurality of bottlenecks in practical engineering application, including the problems of complicated flow, strong specialization, expensive equipment, slow speed and the like. Taking the existing road surface disease detection method as an example, a detection person can distinguish the position, the scale and the type of the disease one by one in a road surface image shot by a road surface detection vehicle, which is one of the most common modes in the current road detection, namely, machine acquisition is cooperated with manual judgment, but the existing method can not accurately judge the specific conditions of road surface parameters and defect diseases. The core drilling and sampling of the pavement are mostly used for various indoor strength tests and extraction tests to reflect the quality indexes of the pavement, and although the condition of pavement diseases can be accurately judged to a certain extent, the on-site evaluation is difficult to carry out due to a plurality of reasons. In short, the development of road surface detection technology aims at more effective information acquisition, faster determination method, more accurate analysis means and less human cost consumption.

Therefore, it is necessary to design a method for rapidly detecting the state of the asphalt pavement.

Disclosure of Invention

The invention aims to provide a method for rapidly detecting the state of an asphalt pavement, and aims to solve the problems that the specific situations of pavement parameters and defect diseases cannot be accurately judged by the conventional common road detection method in the background art, and the field evaluation is difficult to perform by the conventional accurate judgment methods.

In order to achieve the purpose, the invention provides a method for rapidly detecting the state of an asphalt pavement, which comprises the following steps:

step one, drilling a core on a road to obtain a cylindrical road test piece;

secondly, shooting a plurality of photos on the side surface of a circle of the cylindrical road test piece by using an area-array camera, matching, aligning and reconstructing all the shot photos, and synthesizing panoramic images by using image synthesis software to obtain a synthetic image;

marking aggregate particles with the particle size larger than a preset standard in the synthetic image by using image marking software;

training and identifying the marked synthetic graph by adopting an artificial neural network, enabling the matching of the image blocks in the synthetic graph to be smoother by adopting Gaussian weight optimization, and increasing the identification accuracy of training by adopting a secondary marking method to obtain an identification result image;

performing binarization calculation and processing on the obtained identification result image, filling cracks and cavities in aggregate particles in the image after binarization calculation and processing in comparison with the original image, and performing particle segmentation on adhesion among the aggregate particles caused in the image binarization and filling process; and analyzing the morphological characteristics and spatial position distribution of aggregate particles in the processed identification result, representing the segregation and defects of the asphalt mixture in the cylindrical road test piece, and obtaining the parameter performance of each cylindrical road test piece so as to finish the aim of rapidly detecting the road state.

In a specific embodiment, in the second step, the aggregate particles with the particle size larger than the predetermined standard are aggregate particles with the particle size larger than 2.36 mm.

In a specific embodiment, in the second step, the area-array camera is a smart phone or a digital camera.

In a specific embodiment, in the third step, the artificial neural network preferably uses a U-NET network and a U-NET + + network.

In a specific embodiment, in the second step, the Image synthesis software is Image Composite Editor software; the image marking software is TeeAnnotting software or Labelme software.

In a specific embodiment, in the second step, the resolution of the composite map is 200 pixels per centimeter to 500 pixels per centimeter.

In a specific embodiment, in the second step, when shooting is performed around one circle of the cylindrical road test piece, the shooting rotation angle difference is 20-30 degrees, and 12-18 pictures are shot in total.

In a specific embodiment, the gaussian weight optimization is to reduce the translation distance of half of the sampling image blocks when the synthetic image is cut into small image blocks for training and recognition, so that overlapping portions exist between the small image blocks, and a grid effect is avoided; and using a Gaussian kernel as the weight when each image block region is superposed on the identification result of the small image block, taking the identification result of a new small image block for the part with the background weight smaller than the kernel weight, updating the part into a new kernel weight, and finally forming a complete identification result image.

In a specific implementation mode, the secondary marking method is that the recognition result image is reversely led into image marking software to obtain the contour of the aggregate particles after being segmented, fine modification is carried out on the recognition contour mark according to the artificially distinguished correct contour, and the modified marked image is expanded into a data set to be used for artificial neural network training recognition.

In a specific embodiment, in the fourth step, the binarization calculation and processing are performed by using the tsui method, the ISODATA method, or the minimal cut algorithm; the filling is manually judged and manually filled or assisted by ImageJ software Fillholes algorithm; the particle segmentation is realized by adopting an Adjustable washed plug-in ImageJ software.

Compared with the prior art, the invention has the following beneficial effects:

the invention uses advanced and portable image acquisition and reconstruction synthesis technology for acquiring the images of the road asphalt mixture, and realizes the rapid, accurate and quantitative analysis of various performances of the asphalt mixture by using the image recognition function of the trained neural network and the optimized image processing technology, thereby achieving the purpose of rapidly detecting the road state.

According to the method, the side surface of the core sample of the asphalt mixture is shot in an area-array camera imaging mode, and the effect of series picture reconstruction and synthesis is improved by introducing a method of matching a characteristic point cloud and a model. The area-array camera is utilized to realize the image reduction of the curved side surface of the asphalt mixture core sample, and the professional imaging function and level close to those of the industrial linear array camera are achieved by using convenient equipment.

The invention uses the U-NET + + migration of the medical segmentation network for the identification of the core sample of the asphalt mixture, thereby achieving better identification and segmentation effects.

According to the method, the matching and the smoothness of the output image are improved by introducing a Gaussian fuzzy algorithm; and the recognition efficiency of the asphalt mixture synthetic picture is optimized by adopting a secondary marking method. The trained T-U-NET + + has excellent identification accuracy and also has general applicability to asphalt mixtures with similar structures.

The method and the parameters for reasonably binarizing, filling and particle segmenting the identification result image have good representation effect on real surface information.

The present invention uses image processing results for rapid data acquisition and analysis. Including the slenderness ratio and roundness of aggregate particles, the uniformity and defect analysis of the asphalt mixture and the like. Particularly, in the nondestructive analysis of the asphalt mixture, the real space distribution information can be obtained, and a segregation weighting algorithm along the depth direction is provided, so that the advantages are obvious, and the manpower and material resources are saved.

In addition to the objects, features and advantages described above, other objects, features and advantages of the present invention are also provided. The present invention is described in further detail below.

Detailed Description

The following is a detailed description of embodiments of the invention, but the invention can be implemented in many different ways, as defined and covered by the claims.

The image analysis of the pavement detection in the macro scale category mostly takes pavement images as analysis targets; the microscopic scale is used for obtaining information from aggregates, mixtures or pavement core samples. In the method for acquiring the surface image, the road surface can be spliced into a complete image by adopting a linear array camera just as the engineering inspection vehicle has the function of shooting the road surface image; the asphalt mixture is shot on the surface or a cutting plane of a test piece by a multi-purpose area-array camera, and if the surface is non-planar or uneven, the accuracy of the obtained information is adversely affected, so that indoor research has to adopt ray methods such as CT scanning and the like to reduce two-dimensional images or three-dimensional models of the asphalt mixture. In fact, in recent years, the hardware integration level and the imaging effect of the area-array camera are greatly improved, and compared with a line-array camera, clear pictures can be quickly obtained without professional skills, professional software and strict shooting conditions. The application of the advanced area-array camera in pavement detection is beneficial to realizing light-weight work, if the image information of a field pavement or an indoor test piece can be easily obtained only by shooting with widely popularized electronic equipment such as a mobile phone, a digital camera and the like, and the method has great value for promoting scientific research, construction, detection and other work of pavement mixture image analysis.

The deep learning becomes an important technical method of artificial intelligence machine learning due to the advantages of trainability, high efficiency, accuracy and the like, wherein a Convolutional Neural Network (CNN) is one of representative algorithms for deep learning image recognition, and on the basis of excellent image classification capability of the CNN, the improved FCN realizes the recognition function of a specific object. The road surface detection is essentially a large amount of image content discrimination work with similarity repeatability, which is perfectly fit with the advantages of the FCN framework, and in recent years, CNN/FCN mainly focuses on the algorithm and application research of road surface disease automatic identification, acquisition, processing and analysis of information after model reconstruction and the like in the road field. After the obtained road surface image is quickly shot and marked and trained, the identification and classification of the road surface diseases such as cracks, pits and the like are close to the result of naked eye judgment. Trained CNN/FCN models for image recognition of mixes enable the characterization of the mesostructure component distribution of the mix. These examples demonstrate that the ideal CNN/FCN model performs close to the level of a road engineer or researcher in a trained task on a mass learning basis, not only processing faster, but also without subjective error. The image analysis of the asphalt mixture is essentially the identification and segmentation of a three-phase system of gaps, asphalt mortar and aggregate, and generally, the method for distinguishing aggregate (coarse aggregate) from other phases is a reasonable method for image binary analysis. The U-NET artificial neural network is a neural network for biomedicine proposed in 2015, and is excellent in cell recognition and segmentation. An improved neural network algorithm U-NET + + (U-NETPlusPlus) based on U-NET is proposed in 2018, and aims to realize more accurate medical image segmentation and show better effects in macro-scale organ segmentation and micro-scale cell nucleus segmentation. The asphalt mixture test piece is used as a main object of a pavement indoor test, and the two-dimensional shape of the aggregate on the visible surface of the asphalt mixture test piece mainly has the characteristics of large particle size difference, irregular shape, random distribution, large quantity and the like, and is similar to the segmentation characteristics of medical images. Aggregate identification and segmentation of the asphalt mixture can be realized by migrating the U-NET or the U-NET + +.

The invention applies the technologies of image reconstruction algorithm, deep learning image identification and the like to the practical application of road direction. The invention utilizes portable and easily-obtained equipment, combines the advantage of low use threshold, realizes rapid and accurate image and data acquisition, realizes tool and reproducibility, and eliminates the bottleneck caused by factors such as high cost, high difficulty and the like.

Example 1

The invention relates to a method for rapidly detecting the state of an asphalt pavement, which comprises the following steps:

step one, drilling a core on a road to obtain a cylindrical road test piece;

marking the aggregate with the particle size larger than a preset standard in the synthetic image by using image marking software;

Artificial Neural Networks (ANN) are a research hotspot in the field of Artificial intelligence since the 80 s of the 20 th century. The method abstracts the human brain neuron network from the information processing angle, establishes a certain simple model, and forms different networks according to different connection modes. It is also often directly referred to in engineering and academia as neural networks or neural-like networks. A neural network is an operational model, which is formed by connecting a large number of nodes (or neurons). Each node represents a particular output function, called the excitation function. Every connection between two nodes represents a weighted value, called weight, for the signal passing through the connection, which is equivalent to the memory of the artificial neural network. The output of the network is different according to the connection mode of the network, the weight value and the excitation function. The network itself is usually an approximation to some algorithm or function in nature, and may also be an expression of a logic strategy.

The cylindrical road test piece is an indoor formed Marshall test piece, a rotary compaction test piece or a core sample drilled from a road surface and the like, and can be used as a research object. The Marshall test method is used for forming cylinder test pieces with the diameter of 101.6mm and the height of 63.5mm, and the cylinder test pieces with various sizes can be obtained by rotary compaction, wherein the diameters are mainly 100mm and 150 mm. Because the side surface of the core sample is a cutting surface, the core sample is a unified research object, and an exposed cutting surface is obtained by drilling a core for a complete indoor forming test piece. The size of the test piece is not required, the test piece can be an upper layer core sample with the height of 4 cm-5 cm, or a comprehensive layer core sample with the height of 16 cm-20 cm, and the diameter has no special requirement.

When the area-array camera such as a smart phone or a digital camera is used for shooting the side surface of the asphalt mixture, the shot effective pixels are determined by each splicing part according to the proportion of the test piece in the picture, and the output resolution of the final composite image is determined by each splicing part.

The area-array camera adopts multi-angle photos to represent complete side surface characteristics, so that the area-array camera needs to shoot a circle around the side surface of the test piece. The matching and reconstructing effect of the images is determined by characteristic association points among the photos, each shot photo and other photos together represent part of the side surface of the mixed material, and when the number of the association points is sufficient, the matching and aligning method of the photos can restore the real positions of the association points. Practical practice has shown that when the difference in the rotational angle between adjacent photographs taken reaches 40 ° or more, i.e., 9 photographs or less in total, a situation in which reconstruction fails easily occurs. And considering that the mobile phone shooting mode is single-point focusing, the imaging effect of the two sides of the asphalt mixture in the picture is poor, and if the rotation angle difference is selected too much, the area of the part of the synthesized image is fuzzy. When the rotation angle difference is too small, the number of the photos is increased, so that the photo alignment calculation time is greatly prolonged, the imaging quality is not obviously improved, the reasonable shooting rotation angle difference is 20-30 degrees, and 12-18 photos are shot in total.

Since the matching between the photographs is determined by the correlation points, a slight change in the pitch angle and the lateral tilt do not affect the result of the synthesized image when photographed, but when a photograph is photographed at a special angle or the imaging effect is not good, it may be rejected and thus cannot be effectively used for matching alignment. In any event, it is essential to ensure that the part of the test piece in the picture is as clear as possible and has sufficient overlap (i.e., the point of association), and there may be slight differences between the shooting angle and the distance of the machine.

The matching points of photo alignment are divided into total key points and connecting points, the total key points are all relevant points for determining the matching model, and the connecting points are effective relevant points of each photo. And when in matching alignment, firstly acquiring the associated points on the photos, then quickly selecting and matching, then determining and restoring the positions of the matched points according to the matching result, and finally reversely calculating the camera positions for shooting the photos. For example, when the number of photos is 13 and the number of key points and connection points is limited to 10000 and 3000, the spatial distribution of the associated points often has a certain error, and a correct matching alignment result cannot be obtained; by properly increasing the number of photos and improving the upper limit of key points and connection points, the point cloud is uniformly and reasonably distributed, and the presented alignment result is good; usually, the matching result of the core sample of the lower layer with a large diameter is more dense. Under the prerequisite of guaranteeing photo quality and quantity, 10000 and 3000 are no less than respectively to preferred key point and tie point restriction number, when the test piece size increases or shoots the rotation angle difference and increase, can reach more ideal alignment effect through the quantity that suitably increases the restriction.

The point cloud reconstruction scene obtained after matching and aligning only reflects the relative position information of the photos, and a model with a standard size is required to perform coordinate positioning before splicing the photos, wherein the size of the model is equal to that of a shooting core sample. And after the point cloud reconstruction scene is adjusted to be aligned with the model of the outer side surface of the core sample, restoring the size and the coordinate position of each photo according to the ratio of 1:1 by taking the model as a reference. When the point cloud reconstruction scene has a large deviation from the model, deformation and dislocation of a subsequent spliced composite image are easily caused. If the point cloud coordinate obviously deviates from the model, the matching and aligning effect of the photo is not good, the wrong associated point position is obtained through calculation, and the matching and aligning operation needs to be carried out again or the photo needs to be shot again.

The cylinder model is the most common in the non-planar model, and is the one with better effect in realizing panoramic image synthesis. The invention selects Microsoft research institute non-open source software Image Composite Editor (ICE) to splice and synthesize the Image. ICE has obvious advantages in the stitching treatment of the overlapped part of the photos, the splicing of the super-large images and the splicing efficiency. The composite plot corresponds to the cylindrical core sample being spread flat on its outside surface in a planar fashion.

If distortion points exist during photo matching cropping, distortion of the edges of the composite image may result, which is often the case at both edges of the composite image, which should be cropped out when the composite image has a sufficient size. The most effective way in determining the resolution of the composite image is to change the value per unit pixel at the time of the photo cropping process. The optimal resolution of the composite image is determined by taking pixels by an area-array camera, and taking a picture of 1200 ten thousand pixels as an example, the side surface of the core sample contains about 200 effective pixels per centimeter unit, and the target resolution of the composite image should not be less than the value considering the precision of label training and recognition. The resolution of the composite image is preferably 200 to 500 pixels per cm, with reference to the photographing apparatus and the photographing distance used in the present invention.

Compared with the common panoramic synthesis method, the matching alignment, reconstruction, splicing and synthesis method provided by the invention has the advantages that the pretreatment work is added on the operation flow, so that the synthetic image effect is closer to the side surface of the real core sample. The cutting texture of the aggregate particles is consistent with the real texture and is distributed horizontally and parallelly, and the common panoramic splicing method has partial corrugated distortion, besides, the obvious difference is the deformation and dislocation of the aggregate contour. Usually, the side surfaces cannot be completely spread to the same plane by the direct splicing method, and a boundary effect exists during splicing synthesis, so that the shape and distribution of part of aggregate are distorted, and even picture information is lost. Therefore, when synthesizing images on the side surface of the cylindrical core, it is necessary to align the point cloud reconstructed scene with the standard size model, as well as to align the pictures.

The asphalt mixture forms a three-phase system by asphalt mortar, gaps and aggregates, the surface forms of the coarse aggregates and part of the fine aggregates are easy to distinguish in a cut surface image, and the gaps, the asphalt and the asphalt mortar cannot be directly distinguished due to small scale, dark color and the like, so that the image identification method of the asphalt mixture is used for distinguishing the aggregates from other phase systems. When image marking is carried out, the cutting exposed surface of the aggregate particles has a closed and relatively regular outline, the color is easy to identify, and the method is more rapid and convenient compared with the marking gap and the asphalt mortar component, so that the aggregate particles are selected as a marking object.

The purpose of image marking is to use the result of artificial discrimination as a training set for training and learning of a neural network, and image marking software should have the characterization function of the discrimination result and a visual interactive interface. There are many high-quality marking software, such as TeeAnnotting software or Labelme software, etc., the basic marking method includes manual marking and interactive marking, and the interactive marking usually adopts built-in algorithm for distinguishing and is used for rough marking, and then manual modification is carried out. When the target of image recognition is the aggregate in the asphalt mixture, the marked objects are characterized by large quantity, wide distribution, irregular shape, larger difference in scale and the like. Although the characteristics of each aggregate are complex, the interference information of the cut surface image of the asphalt mixture is not much, and the semantic segmentation is very effective for a large amount of aggregate particles, namely, only whether pixel points are aggregate components or not is judged, and the independent segmentation is not needed for different aggregate particles, so that the method is consistent with the target of a ray scanning method such as CT (computed tomography). Another class of example segmentation methods can also distinguish different individuals in the same class, i.e., the contours of each aggregate, based on identifying pixel classes. The semantic segmentation method and the example segmentation method have the advantages that the training and recognition speed is high, the precision is good, and the independent individual segmentation can be realized in the case of the example segmentation method, but higher requirements are provided for a segmentation algorithm. The invention adopts the semantic segmentation idea to mark the aggregate particles, the accuracy and fineness of the marking (namely the contour information of the particles) are crucial to the training effect of the neural network, and in addition, the problem of local ambiguity is effectively solved for the global scale matching of the image. In the asphalt mixture, it is considered that fine aggregates having a particle size of 2.36mm or less form asphalt mortar together with asphalt, and it is difficult to finely distinguish fine particles in the asphalt cement. During marking, the small-granularity or similar-color fine particles which are difficult to distinguish are ignored, so that the marking workload is reduced, and the training precision of the neural network can be improved.

The synthetic image is cut into small-size images for marking and training, the semantic segmentation method is adopted to meet no requirement on the cutting integrity of the images, namely, single aggregate particles can be cut at will, and in order to ensure the marking accuracy, the cutting image contains rich information (including aggregate particles, asphalt mortar and the like of all scales) as much as possible. Take 1024 by 1024 pixel clipping as an example. During marking, coordinate points are picked up on the outer contour of the aggregate particles, adjacent marking points are connected in a straight line, namely, the aggregate particles are regarded as a polygon, and the larger the number of the marking points, the closer the real form of the aggregate cutting surface is. Selecting large-particle mark points in an area with obvious curvature change as fine as possible; the small particles are regular in shape, the contour characteristics can be reflected by using fewer marking points, but the number of the marking points of a single particle is not less than 10. Because the aggregate particles are densely distributed, the contour obtained by adopting the interactive marking algorithm needs to be distinguished and corrected manually, a part of small particles are easy to ignore, and a marking method for picking up coordinates manually can be adopted if necessary.

In order to realize better universality, the artificial neural network can be used for learning asphalt mixtures with different mixing ratios. The data set creating method is characterized in that asphalt mixture images with different mix proportions are randomly selected, a large number of cutting mark images with various mix proportions are included, and the pavement core sample synthetic images of an upper surface layer (T, SMA-13), a middle surface layer (M, AC-20) and a lower surface layer (B, AC-25) are selected as training objects. Wherein SMA-13 represents an asphalt mixture with the maximum nominal grain diameter of 13mm and the SMA is a typical close-packed structure with a discontinuous graded framework; and the AC is a dense suspension dense asphalt mixture. SMA and AC are clearly distinct on the side surface composite image because of the structural differences caused by grading: SMA has a large coarse content and a small fine content, and AC has a uniform coarse-fine ratio.

Compared with the picture classification function realized by cnn (relational Neural networks), fcn (full relational networks) realizes that some kind of objects in the picture are recognized by a semantic segmentation method. After the FCN converts the full-connection layer into the convolutional layer, the convolutional layer can be reversely sampled to the feature map of the last convolutional layer, the feature map is restored to the size of the input image to be output, each pixel is restored from the abstract features one by one and probability classification is carried out, and therefore semantic segmentation is achieved. The segmentation thinking is very suitable for identifying aggregates in the asphalt mixture, so that deep learning can replace ray methods such as CT (computed tomography) in partial application scenes, and gray level representation of aggregate particles is realized.

Compared with a deconvolution layer connected after FCN downsampling, the U-NET adopts an upsampling structure symmetrical to the downsampling, so that a symmetrical U-shaped network is formed. The U-NET reserves the multi-scale features of the images by a concat feature fusion method and a cascade jump connection mode, and simultaneously calculates loss in a weighted cross entropy mode, optimizes the learning efficiency of image boundaries and enables the image boundaries to be excellent in high-resolution medical image recognition. And constructing short connections of each layer from shallow to deep on the basis of long connection sampling of the U-NET, and filling blank areas of the same level, namely the U-NET + +. The abstract features sensed by different levels can be fused in the long connection and can be transmitted in the short connection, so that the information loss in the sampling process in the long connection is reduced to the maximum extent. The U-NET + + filling structure can increase the parameter quantity of the network by a small amount to obtain a more ideal segmentation effect, which is incomparable to the simple widening of the U-NET.

The cutting surface diagram of the asphalt mixture has general similarity with the characteristics of cell research images in the medical field, namely, the segmentation of an aggregate phase system and an asphalt and void phase system is similar to the segmentation of biological cells and voids, and the characteristic forms and the characteristic quantities and the distribution of aggregate particles are relatively close to those of cells and are as complex as those of a cell diagram. The U-NET and the improved U-NET + + for medical image segmentation can meet the requirements of high resolution and segmentation fineness of a synthetic image, so that the U-NET + + with a better segmentation effect is adopted for transplantation to realize asphalt mixture image identification.

And cutting the image by taking 512 pixels by 512 pixels as the cutting size and taking 128 pixels as the translation step length, and using the generated cutting surface image and the corresponding marked image for training the U-NET + + original network. The examination indexes of the training parameters are val _ mean _ iou, iou (intersection over Union), namely the ratio of the overlapping area (Overlap) to the Union area (Union), wherein the overlapping area is the image intersection of the artificial labeling and the learning result, and the Union area is the Union of the two. For the training set, iou is a more strict indicator of the verification accuracy. By comparing the two accuracy rate change conditions of 100 training rounds, the fact that the increasing speed of val _ acc is very high in the initial training period, the val _ mean _ iou is relatively stably improved after being stabilized by about 97%, and the accuracy rate of particle image recognition is ideal after 100 training rounds by about 88%.

The training effect also varies with the mix proportion of the study objects, the learning efficiency of AC-20 and AC-25 is better than that of SMA-13, the single mix proportion image has better effect than that of the mixed image, but the training effect variation is closer as the number of training rounds is increased. In addition, the four training curve trends of the SMA-13, the AC-20, the AC-25 and the mixed image have similarity, the accuracy rate can reach more than 69% when 10 rounds of training are carried out, the accuracy rate can reach 88% -90% after 100 rounds of training, and the training curve with higher accuracy rate and stable rising shows that the training method of the artificial neural network model is effective. When the translation step length of picture cutting is changed, the shorter the step length is, the more the number of generated data set pictures is, the lower the recognition rate at the beginning is, the recognition accuracy rate is very close along with the improvement of the training round, and the recognition accuracy rate can reach more than 88% after 100 rounds.

The high resolution composite image was cropped into 512 x 512 pixel tiles for trained U-NET + + (hereinafter T-U-NET + +). The small image block identification method has some problems in outputting images, namely, the edge matching effect between different image blocks is not ideal, when the cutting step size is the same as the size of the image block, the cut particles positioned at the edge of the image block have abrupt change of gray scales, which is caused by the difference of contour gray scales given to the aggregate particles when two image blocks are respectively identified by T-U-NET + +. In addition, when different image blocks are spliced, a grid effect appears, namely, thin black lines among the image blocks have a great influence on the binarization effect of the gray level image.

In order to solve the edge effect and the grid effect, the invention adopts a Gaussian weight optimization method. When the small blocks are cut, the translation distance of half of the sampling blocks is reduced (i.e. the blocks are translated by 256 pixels), so that the blocks have an overlapping part, and the grid effect is avoided. And using a Gaussian kernel as the weight when each image block region is superposed for the identification result of the image block, taking a new identification result for the part with the background weight smaller than the kernel weight, updating the part into a new kernel weight, and finally forming a complete mosaic identification result. Generally speaking, the gaussian weight optimization is to optimize the matching recognition result by using the weight judgment of the overlapping part, and the gaussian weight optimization enables the matching of the image blocks in the synthesized image to be smoother. When a smaller translation distance is adopted (i.e. the area of the overlapped region is increased), the smoothing effect of the image is slightly improved, but the calculation efficiency is reduced more.

The highest accuracy of the U-NET + + training mode is maintained to be about 90% -91% (epoch is greater than 180, val _ mean _ iou is approximately equal to 90%), and generally, the higher accuracy can be achieved by improving the neural network model structure. However, another key point limiting the accuracy is labeling, the level of iou is directly dependent on the accuracy of the labeled region, and the training for the T-U-NET is the standard for optimizing the model according to whether the iou is promoted or not. The training effect is very slight even the accuracy rate is reduced after a certain turn, which shows that the model self-training effect only achieves the bottleneck. At the moment, a gray scale image of the identification result shows that a small amount of fuzzy areas exist in the edge outline of part of aggregate particles, the characteristic form identification of some finer particles is wrong, the false identification caused by the pollution of the cutting surface is also wrong, and the false identification is some differences which can be identified artificially and is also the most intuitive optimization direction of the T-U-NET.

Therefore, subsequent training can be targeted to let the artificial neural network learn where improvement is needed. The invention uses a secondary marking method to reversely import the T-U-NET + + identification result into image marking software to obtain the contour after the particle segmentation, finely modifies the identification contour mark according to the manually distinguished correct contour, and expands a data set with the modified marking picture for T-U-NET + + training. The secondary marking method has the advantages that according to the thought of correcting model errors, the learning efficiency is improved through different marks with the same content in the data set, and the upper limit of the identification accuracy rate is also improved. After 100 rounds of training and 200 rounds of training, the secondary marking method is adopted respectively, and the training is continued after the data set is expanded. Compared with the direct training of 300 rounds, the secondary marking method achieves 90.6% accuracy of the 300 rounds of the common method already in 160 rounds, the accuracy is improved from 90.6% to 91.8% after the 300 rounds of training, and the effect of accelerating and improving the training is obvious.

The synthetic image on the side surface of the core sample is identified and segmented by utilizing T-U-NET + +, the identification results take AC-20, AC-25 and SUP-20 as examples, the outline shape of the coarse particles is very close to the real situation, and the adhesion with the fine particles rarely occurs; most of the finer particles are closer to the manually distinguished outline, but the problem of fuzzy edges exists; the distribution of the gaps and the asphalt mortar is basically not different from the real situation. The SUP-20 mixture ratio is not used for T-U-NET + + training, but the SUP-20 mixture ratio and the AC belong to a suspension-compact structure, the adopted coarse aggregates are limestone with the same structure, and the identification result is ideal. And the identification and segmentation of the SMA-13 are not ideal, the proportion of coarse particles is large, the outline is easy to blur, and the SMA-13 is adhered to some fine particles. The method is mainly caused by two reasons, the framework-dense structure of the SMA causes the aggregate with smaller grain diameter to be lost, the coarse grains are directly contacted with each other, and the difficulty of identifying the artificial neural network is increased when the aggregate and the AC image of the suspension-dense structure are identified and segmented together; on the other hand, the upper SMA-13 layer is usually made of basalt aggregate, and the color of particles of a cut surface of the core sample is dark and is different from that of bright and whitish limestone.

After the data set is expanded and the training is carried out on the SMA mixture ratio, the identification accuracy of the SMA-13 on the upper layer is greatly improved. The training effect of various mixture structures or different aggregate lithologies on the model is different, the high accuracy of the mix proportion with poor recognition effect can be achieved by expanding a data set and strengthening training, and the special model can be retrained only on the mix proportion. Compared with T-U-NET + +, which is subjected to mixed training at different mix proportion, the mix proportion recognition effect is obviously improved by adopting single mix proportion training, but the universality is greatly reduced. For example, the same T-U-NET + + can be used for AC and SUP identification, and it is clearly more appropriate for SMA to use a dedicated T-U-NET + +.

Comparing the identification results, the cleanliness of the side surface of the core sample is an important factor influencing the identification effect. Dirt and abrasion increase the difficulty of image recognition, and can cause blurring of the outline of the particles or adhesion between particles. If a lot of soil and dust are attached to the bottom of the core sample and the core sample is used for shooting a composite image without cleaning, the particle segmentation effect of the bottom is poor, and the identification accuracy is ideal if the particles are relatively clean in an adverse view. The particles have color difference, so that the influence on the identification result is small in practice; the transverse cutting lines left on the surfaces of the particles during core drilling basically have no influence on identification and segmentation, and the aggregate particles are complete in shape, which are favorable results of model training. The illumination condition influences the brightness of the synthesized picture when the image is shot, but the recognition effect is not obviously changed. The image reconstruction and splicing method does not need to fix the machine position and the shooting angle, the technical requirements on image shooting are not high in essence, and only clear pictures need to be shot after focusing.

The T-U-NET + + identification result is a gray level image of aggregate particle segmentation, and the final contour of the aggregate needs to be judged and determined through binarization processing. The image is subjected to binarization calculation and processing by adopting an Otsu method (OTSU), a minimal Cut algorithm (Minimum Cut), a target tracking algorithm (Mean shift) and an ISODATA method respectively, and the threshold division rule shows that for the image of the Cut surface of the asphalt mixture, the segmentation threshold values calculated by the Otsu method and the ISODATA method are very close, while the segmentation threshold value calculated by the minimal Cut algorithm is slightly higher, and the target tracking algorithm calculates a very high threshold value. From the aspect of segmentation effect, the reduction of the surface characteristics of aggregate particles by the Otsu method and the ISODATA method is close to the real condition, the treatment of the minimum segmentation algorithm on the edges of the aggregate particles is slightly conservative, but the difference between the minimum segmentation algorithm and the minimum segmentation algorithm is not obvious, the outline of coarse particles in the target tracking algorithm is narrowed, the size of fine particles is obviously smaller than the real size, some small cavities are formed inside the aggregate particles, and the binarization effect is not ideal. Therefore, the Otsu method, the ISODATA method and the minimal cutting algorithm are preferably used as the binarization algorithm after T-U-NET + + identification, the high identification accuracy of the T-U-NET + + is benefited, the proportion of the gray fuzzy part in the aggregate particle contour is reduced, and the image binarization difficulty is reduced.

Cracks and voids within the aggregate particles inevitably exist in the binarized image, which are mostly caused by mud dust or abrasion on the particle surface by comparison with the original image, and this part is regarded as a non-particle component by the binarization process. For small amounts of data, direct methods of manual discrimination and manual population may be employed, and for large amounts of data, software-assisted population functions, such as the Fillholes algorithm of ImageJ, may be employed. The cracks and voids in the image are partially filled, reducing the true morphology of the aggregate particle surface. But part of the asphalt mixture gaps may be filled by mistake because the particles are adhered in a part of the area, the gaps are enclosed among the particles, and the area needs to be selected as a non-filling object through manual framing.

In the image binarization and filling processes, adhesion among aggregate particles is inevitably caused, and the result of particle analysis is influenced, so that particle segmentation is necessary. The particle analysis is used as a judgment basis for the aggregate morphology characteristics, and the adhesion segmentation among some coarse particles is particularly important. The invention adopts an Adjustable dividing line (Adjustable watercut) plug-in ImageJ to realize particle segmentation, and is a method for judging whether a particle object adopts segmentation operation or not according to an Euclidean Distance Map (EDM). The condition controlling the segmentation is the tolerance value of the euclidean distance, which determines the difference between the smaller value of the largest inscribed circle and the radius between the "neck" inscribed circles between the particles, the conglutinated parts between the aggregate particles are the "necks" to be segmented, macroscopically narrow regions with abrupt width changes, and thus it is logical to use the adjustable dividing line method.

Analyzing the particle segmentation results under different tolerances, and adopting ImageJ default tolerance value of 0.5 to cause a large amount of error cutting on the aggregates with strong edges and corners and the needle-shaped aggregates; after the tolerance value is increased to 10, the cutting effect is good, and the aggregate parting line is clear and accurate; but when the tolerance value is increased by 25, the adhesion of some aggregates is not separated. The Euclidean distance algorithm is characterized by an inscribed circle, the closer the aggregate form is to the circle, the better the segmentation effect is, and the tolerance value should be selected with flat particles not cut off as a lower limit, which is as close to a real state as possible. This also indicates the necessity of the filling operation in the present invention, that is, cracks and voids are liable to cause a dividing malfunction. A large number of image segmentation works show that the tolerance of the aggregate particles is preferably 5-20, and if the segmentation effect of the whole image is not ideal, the whole image can be segmented independently by selecting a part in a frame mode.

The recognition result can be used for morphological feature analysis of the particles after being processed by binarization, filling, particle segmentation and the like. The two-dimensional morphological characteristic indexes of the aggregate mainly comprise convexity, roundness, axiality and the like, and the three-dimensional morphological characteristic indexes mainly comprise needle slice degree, sphericity, morphological factors and the like. The application range of aggregate data analysis is very wide, and by taking the length-to-fineness ratio (needle sheet shape) analysis and roundness analysis of aggregate particles as examples, the length-to-fineness ratio of the aggregate particles determines the mechanical characteristics of the aggregate and can be used for evaluating the needle sheet shape content and the crushing resistance of the aggregate; the roundness value represents the similarity degree of the aggregate plane contour and a circle, indirectly represents the obvious degree of the aggregate corner angle, and has a close relation with the mechanical property of the asphalt mixture.

Through statistical image recognition and image processing of the upper, middle and lower layers, the obtained needle sheet-shaped index analysis shows that the average slenderness ratio of the coarse aggregate is in an overall increasing trend along with the increase of the particle size, the slenderness ratio within the range of 13.2mm-26.5mm is 1.30-1.38, the slenderness ratio within the range of 4.75mm-13.2mm is 1.36-1.44, and the slenderness ratio within the range of 2.36mm-4.75mm is 1.28-1.34. The calculation and analysis show that the needle flake content of the coarse aggregate used in the highway section is only 3.2 percent at most, and is positioned on the lower layer, and the needle flake content is 2.7 percent at least, and is positioned on the upper layer.

The shooting and identification method of the side surface of the core sample is different from the shooting and identification method of the bulk aggregate, the distribution of the aggregate in the spatial position is real, and the method can be used for quickly judging the material distribution of the asphalt mixture and judging the defect position. The invention also provides a data analysis method suitable for the image synthesis and identification, which takes the uniformity analysis and weak point analysis of the asphalt mixture aggregate distribution as an example. The distribution uniformity of the aggregate reflects the distribution uniformity along the depth direction through area weighting according to the barycentric coordinates of two-dimensional aggregate particles, namely whether the segregation phenomenon exists or not, and the method can be used for quickly detecting the engineering quality of the mixing, paving and rolling process; the weak point analysis indirectly reflects the segregation non-uniformity phenomenon of the mixture by judging the distribution of gaps and asphalt mortar, and can also be used for evaluating the engineering quality. Calculating the weighted value of the depth and the area of the coarse aggregate with each grade of particle size according to the formula 1, and recording the value as an aggregate distribution index D_agg. In the formula, A is the aggregate particle area, Y is the aggregate ordinate, and n is the aggregate quantity of the grade particle size.

Equation 1:

ideal D_aggThe calculations should be located near the horizontal centerline of the core sample along the depth, indicating that segregation of the aggregate, generally floating or sinking, does not occur. The aggregate distribution of the upper, middle and lower layers is respectively listed for analysis, wherein the aggregate of 2.36mm-9.5mm is generally uniformly distributed, the aggregate of 9.5mm-26.5mm has a certain deviation, and the distribution is more obvious on the lower layer.

For mixes, either too large voids or a large concentration of asphalt cement can have a detrimental effect on the mechanical and durability properties of the mix. When the coarse aggregate is lost and serious cavities exist, the mechanical property and the durability of the pavement are extremely unfavorable. When the ratio of the white portion in the region is higher than a certain ratio (60% in this example) in the particle analysis, the region is determined to be a weak point. The core sample state corresponds to a real pavement state, the pavement core sample can be used as a rapid surface detection method, a small sample can be sampled and analyzed, and a large sample can be used for detecting and evaluating the pavement engineering quality.

The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions and substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.

Claims

1. A method for rapidly detecting the state of an asphalt pavement is characterized by comprising the following steps:

step one, drilling a core on a road to obtain a cylindrical road test piece;

2. The method for rapidly inspecting the condition of an asphalt pavement according to claim 1, wherein in the second step, the aggregate particles having a particle size larger than the predetermined standard are aggregate particles having a particle size larger than 2.36 mm.

3. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the second step, the area-array camera is a smart phone or a digital camera.

4. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the third step, the artificial neural network preferably adopts a U-NET network and a U-NET + + network.

5. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the second step, the Image synthesis software is Image Composite Editor software; the image marking software is TeeAnnotting software or Labelme software.

6. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the second step, the resolution of the synthetic image is 200-500 pixels per centimeter.

7. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the second step, when shooting is carried out around one circle of the cylindrical road test piece, the shooting rotation angle difference is 20-30 degrees, and 12-18 pictures are shot in total.

8. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein the Gaussian weight optimization is to reduce the translation distance of half of the sampling image blocks when the synthetic image is cut into small image blocks for training and recognition, so that the small image blocks have overlapping parts to avoid the grid effect; and using a Gaussian kernel as the weight when each image block region is superposed on the identification result of the small image block, taking the identification result of a new small image block for the part with the background weight smaller than the kernel weight, updating the part into a new kernel weight, and finally forming a complete identification result image.

9. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein the secondary marking method comprises the steps of reversely importing the recognition result image into image marking software to obtain the segmented outline of the aggregate particles, slightly modifying the recognition outline mark according to the manually distinguished correct outline, and expanding a data set by using the modified marked image for training and recognition of an artificial neural network.

10. The method for rapidly detecting the state of the asphalt pavement according to claim 1, wherein in the fourth step, the binarization calculation and processing are performed by using the Otsu method, the ISODATA method or the minimum cut algorithm; the filling is manually judged and manually filled or assisted by ImageJ software Fillholes algorithm; the particle segmentation is realized by adopting an Adjustable washed plug-in ImageJ software.