CN-121788533-B - Visual identification method and system for defects of underground pipe network

CN121788533BCN 121788533 BCN121788533 BCN 121788533BCN-121788533-B

Abstract

The application relates to the technical field of visual data processing, and discloses a visual identification method and a visual identification system for defects of a subsurface pipe network, wherein the method comprises the steps of presetting the types of the defects of the pipe network, collecting corresponding training image data and expanding, and training two deep learning models based on target detection according to the data; the method comprises the steps of obtaining an underground pipe network image, dividing a part into training data according to a proportion, preprocessing the residual image to realize defect characteristic shift to obtain input data, inputting the input data into a first model to obtain output data and marking the output data on the image, inputting the input data into a second model to obtain verification data according to a preset verification period, calculating a consistency value and a consistency ratio of the two, and adjusting an image extraction proportion according to the negative correlation of the ratio. According to the method, the accuracy and the adaptability of the underground pipe network defect identification are improved through the double-model collaborative verification and dynamic proportion adjustment.

Inventors

WANG YONGFEI
CHEN QIUJIAN
CAI PENG
LIANG JINQI
JIANG ZIWEN
YU TING
TAN SHILIN
PENG YONGQIANG
LAN LIANG
PENG ZHENG
XIAO XIA
HOU LIN

Assignees

湖南省建设工程质量检测中心有限责任公司

Dates

Publication Date: 20260508
Application Date: 20260304

Claims (9)

1. A visual identification method for defects of a subsurface pipe network is characterized by comprising the following steps: Acquiring a preset pipe network defect type, acquiring corresponding training image data based on the pipe network defect type, expanding the training image data, and training a first deep learning model and a second deep learning model which are preset and based on target detection according to the expanded training image data; The method comprises the steps of obtaining an underground pipe network image, extracting one part of the underground pipe network image as training data according to a preset extraction ratio, preprocessing the other part of the underground pipe network image to obtain input data, and shifting defect characteristics in the underground pipe network image by preprocessing; Inputting the input data into a first deep learning model to obtain output data, and marking the output data on an underground pipe network image; Based on a preset verification period, inputting the input data into a second deep learning model to obtain verification data, and calculating a consistency value of the verification data and the output data; Dividing an acquired underground pipe network image into a dark area, a bright area and a transition area according to illumination intensity, adopting self-adaptive histogram equalization to improve the gray dynamic range of the dark area, adopting contrast-limited histogram equalization to the bright area to avoid overexposure, and keeping the gray gradient smooth in the transition area; extracting defect characteristics by adopting an illumination invariant characteristic extraction method based on the preprocessed underground pipe network image, wherein the illumination invariant characteristic extraction method comprises the steps of extracting defect frequency domain amplitude characteristics by adopting Fourier transform and wavelet transform based on a frequency domain method, and extracting defect color component characteristics by adopting HSV color space and Lab color space based on a color space method, wherein the defect characteristics comprise gray invariant moment, color moment and texture characteristics; loading image Net data set pre-training weights by a pre-training method, fine-tuning the model by transfer learning, generating a defect sample under simulated complex illumination by combining an countermeasure generation network, and optimizing the training process of the deep learning model; Collecting multi-mode data at the same visual angle as a visible light image of an underground pipe network, wherein the multi-mode data comprises an infrared image and a depth image, and extracting depth characteristics and temperature characteristics after registering the multi-mode data; the extracted defect features and the multi-mode features are subjected to feature level fusion, specifically, feature graphs are spliced, weighted and summed through a convolution layer, and fusion features are obtained; And inputting the fusion characteristics into an optimized deep learning model to finish the defect identification of the underground pipe network, and outputting the defect type, the identification confidence and the approximate region of the defect in the image.
2. The visual identification method for defects of underground pipe network according to claim 1, wherein the step of collecting the corresponding training image data based on the type of the defects of the pipe network further comprises the following sub-steps: The training image data has a type attribute, a size attribute and an illumination attribute; the method comprises the steps of generating training image data of corresponding illumination attributes based on a preset first illumination condition and a preset second illumination condition, wherein the same training image data adopts the first illumination condition and the second illumination condition in a time sharing mode, the duty ratio of the first illumination condition is a first illumination duty ratio, the duty ratio of the second illumination condition is a second illumination duty ratio, and the sum value of the illumination duty ratios is one; Selecting a preset first size amplitude and a preset second size amplitude as data capacity of training image data of corresponding size attributes according to the generated content conditions, wherein the duty ratio of the first size amplitude is selected as a first size duty ratio, the duty ratio of the second size amplitude is selected as a second size duty ratio, and the sum of the size duty ratios is one; generating training image data with corresponding type attribute by adopting a preset first type channel and a preset second type channel based on the generated content condition and the data capacity, wherein the duty ratio of the first type channel is a first type duty ratio, the duty ratio of the second type channel is a second type duty ratio, and the sum value of the type duty ratios is one; the first illumination duty cycle, the first size duty cycle, and the first type duty cycle are randomly adjusted within a preset duty cycle range.
3. The visual identification method for defects of underground pipe network according to claim 1, wherein the step of expanding training image data further comprises the following sub-steps: Acquiring a first expansion instruction, responding to the first expansion instruction to rotate the training image data, extracting an expansion range value and an expansion amplitude value in the first expansion instruction, adjusting the quantity of the rotating training image data according to the positive correlation of the expansion range value, and adjusting the rotation angle of the training image data according to the positive correlation of the expansion amplitude value; Obtaining a second expansion instruction, scaling the training image data in response to the second expansion instruction, extracting an expansion range value and an expansion amplitude value in the second expansion instruction, adjusting the quantity of the scaled training image data according to the positive correlation of the expansion range value, and adjusting the scaling multiple of the training image data according to the positive correlation of the expansion amplitude value; And acquiring a third expansion instruction, turning over the training image data in response to the third expansion instruction, extracting an expansion range value and an expansion amplitude value in the third expansion instruction, regulating the quantity of the turning-over training image data according to the positive correlation of the expansion range value, and selecting and regulating the turning-over direction of the training image data according to the expansion amplitude value.
4. The visual identification method of defects of underground pipe network according to claim 1 or 2, wherein the method further comprises the steps of: Acquiring illumination types in an underground pipe network image based on a preset illumination extraction algorithm, and calculating the number of the illumination types; If the number of the illumination types is smaller than the preset illumination reference number, calculating an illumination type ratio according to the number of the illumination types and the illumination reference number, and negatively and correspondingly adjusting the classification number of the illumination types in the illumination attribute according to the illumination type ratio; the pretreatment step further comprises the following sub-steps: preprocessing an underground pipe network image by using an illumination compensation algorithm, wherein the illumination compensation algorithm comprises histogram equalization or adaptive histogram equalization; and training the target detection by adding a feature extractor and a classifier with illumination robustness, and applying the trained feature extractor and classifier to the first deep learning model and the second deep learning model.
5. The visual identification method of defects of underground pipe network according to claim 1 or 2, wherein the method further comprises the steps of: acquiring definition types of the underground pipe network image based on a preset feature extraction algorithm, and calculating the number of the definition types; If the number of the definition types is smaller than the preset definition reference number, calculating a definition type ratio according to the number of the definition types and the definition reference number, and inversely related adjusting the classification number of the size types in the size attribute according to the definition type ratio; the pretreatment step further comprises the following sub-steps: preprocessing the image by using a noise suppression algorithm, wherein the noise suppression algorithm comprises median filtering or Gaussian filtering; And training the target detection by adding a feature extractor and a classifier with noise robustness, and applying the trained feature extractor and classifier to the first deep learning model and the second deep learning model.
6. A visual identification method for defects of underground pipe network according to claim 1 or 3, further comprising the steps of: Collecting defect images of underground pipe networks, wherein the defect images cover the scenes of ageing, abrasion or corrosion of the pipeline, and comprise underground pipe network images of different pipe diameters, laying environments and service stages, and constructing an initial training data set based on the defect images; performing expansion processing on the initial training data set by adopting a data enhancement algorithm to obtain a target training data set, wherein the data enhancement algorithm comprises image transformation and noise addition, the image transformation comprises rotation, overturning, scaling, clipping and brightness adjustment, and the noise addition comprises Gaussian noise addition and salt and pepper noise addition; Performing iterative training on a deep learning model for defect identification based on the target training data set, optimizing the model training process by adopting a regularization algorithm and integrated learning in the training process, wherein the regularization algorithm comprises L1 regularization, L2 regularization and Dropout regularization, the integrated learning comprises random forest integration, gradient lifting tree integration and multi-model voting integration, the optimization operation specifically comprises the steps of introducing regularization terms to inhibit overfitting in the model back propagation process, and integrating output results of a plurality of basic models through integrated learning; presetting a mapping table of pipeline parameter combinations, algorithm thresholds and identification accuracy, wherein the mapping table is pre-stored with optimal threshold intervals and accuracy control ranges corresponding to different pipeline types, pipeline materials and pipeline service life combinations; obtaining pipeline parameters of an actual detection scene, wherein the pipeline parameters comprise pipeline types, pipeline materials and pipeline service life, inquiring the mapping table according to the pipeline parameters, and determining a target algorithm threshold value and target recognition accuracy; and adjusting corresponding parameters of the deep learning model according to the target algorithm threshold and the target recognition accuracy through a model parameter configuration interface to complete model adaptation.
7. The visual identification method for defects of underground pipe network according to claim 1, further comprising the steps of: Acquiring an underground pipe network image containing complex overlapping defects; Performing multi-scale decomposition and feature enhancement processing on the underground pipe network image by adopting a multi-scale analysis algorithm, and sequentially extracting defect distinguishing features of the image under preset small scale, medium scale and large scale, wherein the defect distinguishing features comprise defect edge contour features, gray level distribution features and texture features; Taking a preset neural network and a derivative model thereof as a deep learning model, wherein a feature fusion module is arranged in the deep learning model; Inputting the defect distinguishing characteristics under the small scale, the medium scale and the large scale into the deep learning model, and carrying out weighted fusion on the defect distinguishing characteristics of different scales by the characteristic fusion module to obtain multi-scale fusion characteristics; In the process of classifying the multi-scale fusion characteristics by the deep learning model, integrating migration learning and defect identification related knowledge in other fields, wherein the defect identification related knowledge in other fields comprises metal structure corrosion defect identification knowledge and nonmetal pipeline abrasion defect identification knowledge; and through the synergistic effect of the deep learning model and the multi-scale analysis algorithm, the identification of the complex shape overlapping defects and the defects with different scales of the underground pipe network is completed, and the defect types and the region coordinates of the corresponding defects in the image are output.
8. The visual identification method of defects of underground pipe network according to claim 1, wherein the defect positioning process comprises the following steps: Obtaining an underground pipe network image to be subjected to defect positioning, carrying out denoising pretreatment on the underground pipe network image, and removing salt and pepper noise by adopting median filtering and Gaussian noise by adopting Gaussian filtering; Processing the preprocessed underground pipe network image by adopting a preset neural network algorithm, wherein the neural network algorithm comprises FasterR-CNN, YOLOv5 and SSD, and extracting defect candidate characteristics containing definite position information through a region proposal module or an anchor frame mechanism of the network, wherein the position information comprises pixel coordinates and a region aspect ratio; Optimizing and screening the candidate defect characteristics by utilizing context information of the underground pipeline defects, wherein the context information comprises a gray level difference threshold value of adjacent pixels or connectivity judgment conditions of adjacent areas, and the method comprises the specific steps of calculating gray level differences between the candidate defect areas and the adjacent pixels, and reserving the characteristics that the gray level differences are larger than a preset threshold value; The auxiliary method is introduced to co-locate, and comprises an image segmentation method and a target detection method, wherein the image segmentation method and the target detection method are adopted to segment a pipeline region where the defect is located by adopting a semantic segmentation algorithm, so as to eliminate background interference; presetting a defect positioning strategy library, wherein the strategy library comprises positioning strategies and threshold intervals corresponding to different scenes and different defect types; Matching a target positioning strategy and a target threshold value from the strategy library according to the actual detection scene and the defect type, and adjusting an identification strategy and a threshold value corresponding to defect positioning through a model parameter configuration interface; and determining a final boundary frame of the defect based on the optimized defect characteristics, an auxiliary method output result and an adjusted strategy threshold value, and outputting accurate pixel coordinates and a positioning deviation range of the defect in the image.
9. A visual identification system for defects of a underground pipe network, comprising a processor, wherein the steps of the visual identification method for defects of the underground pipe network as claimed in any one of claims 1 to 8 are executed in the processor.

Description

Visual identification method and system for defects of underground pipe network Technical Field The application relates to the technical field of visual data processing, in particular to a visual identification method and system for defects of a underground pipe network. Background The urban underground pipe network is used as a core carrier of key infrastructure such as water supply, water discharge, fuel gas, heat and the like, and directly relates to civil security and urban normal operation. With the increase of the service life of the pipe network, the pipe is influenced by factors such as soil corrosion, geological settlement, external force construction disturbance and the like, and the pipe is easy to crack, break, deform, block and the like, and if repair is not found in time, safety accidents such as medium leakage, pavement collapse, environmental pollution and the like can be caused. The scale of the urban pipe network is continuously enlarged, the traditional detection means are difficult to cover complex pipe network scenes, the requirements on the efficient and accurate pipeline defect detection technology are urgent, and the method becomes a key requirement point for guaranteeing the operation and maintenance safety of the pipe network. In the existing pipeline defect detection method, the manual interpretation mode based on image acquisition is most commonly applied. The method generally comprises the steps of shooting through equipment such as a pipeline detection robot, a pipeline periscope and the like, obtaining inner wall images or video data by penetrating into the pipeline, then deriving the collected images one by a professional detector, manually observing, identifying and marking the images according to an industry defect judgment standard, determining the type, position coordinates and severity level of the defects, and finally forming a detection report. Although the method can realize defect detection to a certain extent, the method plays an important role in middle-and small-scale pipe network detection, and the whole detection flow is excessively dependent on manual operation. The existing pipeline detection method taking manual interpretation as a core still cannot fully meet the requirements of high-efficiency and accurate detection in actual production. On one hand, the length of a single pipe section in a city can reach hundreds of meters, the number of images acquired by a single pipe section is often hundreds or even thousands of frames, a large number of professionals are needed for manual frame-by-frame interpretation, the detection period is long, especially in a large-scale pipe network detection scene, the human resource allocation is tension, the use efficiency is extremely low, the detection work progress is severely restricted, on the other hand, the manual interpretation is easily influenced by subjective factors, the experience level, the working concentration degree, the judgment standard difference and the like of the detection personnel can cause deviation of the judgment result of the same defect, fine defects are easily omitted, non-defect characteristics can be misjudged as defects, defect information distortion is caused, and adverse effects are brought to the follow-up pipe network repair decision and the safety operation and maintenance. Disclosure of Invention In order to realize automatic detection of pipeline defects, the application provides a visual identification method and a visual identification system for defects of a subsurface pipe network. In a first aspect, the application provides a visual identification method for defects of a underground pipe network, which adopts the following technical scheme: A visual identification method for defects of underground pipe network comprises the following steps: Acquiring a preset pipe network defect type, acquiring corresponding training image data based on the pipe network defect type, expanding the training image data, and training a first deep learning model and a second deep learning model which are preset and based on target detection according to the expanded training image data; the method comprises the steps of obtaining an underground pipe network image, extracting one part of the underground pipe network image as training data according to a preset extraction ratio, preprocessing the other part of the underground pipe network image to obtain input data, and shifting defect characteristics in the underground pipe network image by preprocessing; Inputting the input data into a first deep learning model to obtain output data, and marking the output data on an underground pipe network image; Based on a preset verification period, inputting the input data into a second deep learning model to obtain verification data, calculating a consistency value of the verification data and the output data, calculating a consistency ratio according to the consistency value and a preset consistency reference va