Method for detecting moisture damage on asphalt pavement based on adaptive selection of GPR image grayscale

Patent metadata

Field	Value
Publication number	US-11828840-B2
Application number	US-202017761207-A
Country	US
Kind code	B2
Filing date	Dec 16, 2020
Priority date	Nov 1, 2019
Publication date	Nov 28, 2023
Grant date	Nov 28, 2023

Abstract

A method for detecting a moisture damage on an asphalt pavement based on adaptive selection of a penetrating radar (GPR) image grayscale includes the following steps: step 1: obtaining a moisture damage GPR image dataset through asphalt pavement investigation by using a ground GPR, where a GPR image with an appropriate plot scale is selected according to an adaptive GPR image selection method; step 2: adjusting image resolution, specifically, scaling a resolution of an initial GPR image dataset of a damage directly to 224×224 to obtain a BD dataset; step 3: inputting the dataset into a recognition model, specifically, inputting the BD dataset obtained in step 2 into the recognition model, performing operation by the recognition model, and performing step 4; and step 4: outputting a moisture damage result. The new method truly realizes automatic and intelligent target detection based on the GPR.

First claim

What is claimed is:

1. A method for detecting a moisture damage on an asphalt pavement based on adaptive selection of a ground penetrating radar (GPR) image grayscale, comprising the following steps:
step 1: obtaining a moisture damage GPR image dataset through asphalt pavement investigation by using a GPR, specifically comprising:
step S11: performing the asphalt pavement investigation and data collection by using the GPR: performing on-site data collection on the asphalt pavement by using a GPR system, and during the on-site data collection, determining a damage region on the asphalt pavement, wherein mud-pumping or whitening or stripping occurs in the damage region; and
step S12: obtaining an initial GPR image dataset of a moisture damage: after preprocessing GPR data corresponding to the damage region, selecting a GPR image with an appropriate plot scale, intercepting the GPR image according to a length of 5 m to 6 m, constructing an initial GPR image dataset of the moisture damage, an initial GPR image dataset of a bridge joint, and an initial GPR image dataset of a normal asphalt pavement, and labeling respective features of the moisture damage, the bridge joint, and the normal asphalt pavement;

step 2: adjusting an image resolution:
defining the initial GPR image dataset of the damage as an ID dataset, scaling the ID dataset directly to 224×224 to obtain a first scaled dataset, and defining the first scaled dataset as a BD dataset; and
scaling a resolution of the initial GPR image dataset of the damage directly to 224×224 to obtain the BD dataset;

step 3: inputting dataset into a recognition model, comprising:
inputting the BD dataset obtained in step 2 into the recognition model, performing operation by the recognition model, and then performing step 4, wherein
an input image resolution of the recognition model is 224×224, and an output image resolution is 224×224; and

the recognition model is a mixed deep learning model, and the mixed deep learning model is composed of ResNet50 for feature extraction and YOLO V2 framework for target detection;

step 4: outputting a moisture damage result:
performing a post-processing on an output result of the recognition model in step 3, wherein the post-processing comprises:
step S41: determining a quantity of candidate boxes BBoxes in an image in the output result, and performing step S42 when the quantity of the candidate boxes BBoxes is greater than 1, or directly outputting a result when the quantity of candidate boxes BBoxes is less than or equal to 1;
step S42: determining whether the candidate boxes BBoxes overlap, and performing step S43 when the candidate boxes BBoxes overlap, or directly outputting the result when the candidate boxes BBoxes do not overlap;
step S43: determining whether label names corresponding to overlapped candidate boxes BBoxes are identical, wherein when the label names corresponding to the overlapped candidate boxes BBoxes are identical, a label name corresponding to a combined candidate box BBox maintains unchanged; when the label names corresponding to the overlapped candidate boxes BBoxes are different, two types of label names respectively corresponding to the moisture damage and the bridge joint simultaneously exist, and an output label name is Joint;
step S44: combining the overlapped candidate boxes BBoxes by taking minimum values of x and y and maximum values of w and h of the overlapped candidate boxes BBoxes to obtain the combined candidate box BBox, wherein coordinates of the combined candidate box BBox are [x min , y min , w max , h max ]; and
step S45: outputting the result, wherein in the output result of the recognition model, the output image resolution is adjusted to the image resolution of the initial GPR image dataset of the damage, and the output result is an image with a label name of a target and a position (x, y, h) of a candidate box BBox corresponding to the target;

wherein, the GPR image with the appropriate plot scale is selected according to an adaptive GPR image selection method; and
the adaptive GPR image selection method adaptively selects a suitable GPR image based on a plot scale value of the GPR image, and comprises the followings steps:
step (1): reading preprocessed GPR data:
after preprocessing the GPR data, randomly generating GPR images with different plot scales within a set plot scale range, and constructing an initial random GPR image dataset, wherein the initial random GPR image dataset comprises N images;

step (2): adjusting the image resolution:
defining the initial random GPR image dataset as an RID dataset, scaling the RID dataset directly to 224×224 to obtain a second scaled dataset, and defining the second scaled dataset as an RBD dataset; and
scaling a resolution of the initial GPR image dataset of the moisture damage directly to 224×224 to obtain the RBD dataset;

step (3): inputting dataset into the recognition model:
inputting the RBD dataset obtained in step (2) into the recognition model, performing operation by the recognition model, and then performing step (4), wherein
the recognition model is identical to the recognition model in step 3;

step (4): outputting the moisture damage result:
performing the post-processing on an output result of the recognition model in step (3), wherein the post-processing is identical to the post-processing in step 4, and the output result is the image with the label name of the target and the position (x, y, w, h) of the candidate box BBox corresponding to the target;

step (5): determining, by using the initial random GPR image dataset, whether a detection target exists:
step S51: converting the output result in step (4) into a matrix A i corresponding to pixels on an image, wherein A i is defined as follows:

A
i

[

m
,
n

]

=

{

1
,

x
i

≤
m
≤

x
i

+

W
i

⁢

and
⁢

y
i

≤
n
≤

y
i

+

h
i

Assignees

• Changan Univ
• Univ Chang An

Inventors

• Zhang Jun
• Tao Jun

Classifications

• G01S13/885 Primary

  for ground probing (prospecting or detecting using electromagnetic waves G01V3/12) · CPC title

• G01N27/221

  by investigating the dielectric properties (using microwaves G01N22/00; measuring loss factors or dielectric constants per se G01R27/26) · CPC title

• G01S7/417

  involving the use of neural networks · CPC title

• G01S13/89

  for mapping or imaging · CPC title

• G01V3/12

  operating with electromagnetic waves {(operating with millimetre waves G01V8/005)} · CPC title