CN-122023230-A - Wading concrete microcrack intelligent evaluation method and system based on visual depth model

CN122023230ACN 122023230 ACN122023230 ACN 122023230ACN-122023230-A

Abstract

The invention discloses a wading concrete micro-crack intelligent evaluation method and system based on a visual depth model, wherein the method comprises the steps of collecting a concrete structure image through an unmanned aerial vehicle and preprocessing, and segmenting a crack area of the collected concrete structure image through a GANFormerNet segmentation model, wherein the segmentation method comprises the steps of extracting global semantic crack characteristics of the image, optimizing and fusing crack characteristics crossing different scales, modeling a spatial topological relation among crack pixels, segmenting the crack area of the concrete structure image, carrying out pixel-level quantitative analysis on the crack area based on an output segmentation result, and integrating an analysis result into a user interface system.

Inventors

LI YANGTAO
YANG GUO
YANG WENKUN
MA YUNLIN
ZHAO HAITAO
CHEN YANLI
GU HAO
WEI YANG
ZHU MINGYUAN
YU YICHEN
TIAN JUNHAO

Assignees

南京林业大学
河海大学

Dates

Publication Date: 20260512
Application Date: 20251202

Claims (10)

1. A wading concrete microcrack intelligent evaluation method based on a visual depth model is characterized by comprising the following steps of: collecting a concrete structure image through an unmanned aerial vehicle and preprocessing; And segmenting a crack region of the acquired concrete structure image through a GANFormerNet segmentation model, wherein the segmentation method comprises the following steps: Extracting global semantic crack characteristics of an image, optimizing and fusing crack characteristics crossing different scales, modeling a space topological relation among crack pixels, and dividing a crack region of a concrete structure image; And carrying out pixel-level quantitative analysis on the crack region based on the output segmentation result, and integrating the analysis result into a user interface system.
2. The intelligent assessment method for the micro-cracks of the wading concrete according to claim 1, wherein the GANFormerNet segmentation model is subjected to light weight processing through depth separable convolution, and the depth separable convolution disassembles standard convolution operation into depth-wise convolution and point-wise convolution for calculation.
3. The method for intelligently evaluating wading concrete microcracks according to claim 2, wherein in the depth-wise convolution, a convolution kernel size of K is assumed K 1, The convolution number is M, each needs to be H W times multiply add operation.
4. The method for intelligently evaluating wading concrete microcracks according to claim 2, wherein in the point-by-point convolution, a convolution kernel size of 1 is assumed 1 M, the convolution number is N, each is H W times multiply add operation.
5. Wading concrete microcrack intelligent evaluation system based on visual depth model, characterized by comprising: The image acquisition module is used for acquiring a concrete structure image through the unmanned aerial vehicle and preprocessing the concrete structure image; The image segmentation module is used for segmenting a crack region of the acquired concrete structure image through a GANFormerNet segmentation model, wherein the segmentation method comprises the following steps: Extracting global semantic crack characteristics of an image, optimizing and fusing crack characteristics crossing different scales, modeling a space topological relation among crack pixels, and dividing a crack region of a concrete structure image; And the analysis evaluation module is used for carrying out pixel-level quantitative analysis on the crack area based on the output segmentation result and integrating the analysis result into the user interface system.
6. The wading concrete microcrack intelligent assessment system of claim 5 wherein the GANFormerNet segmentation model comprises a visual transducer unit, a multi-scale semantic interaction unit, a graph roll-up attention unit, and a spatial and channel co-attention unit; The GANFormerNet segmentation model adopts depth separable convolution to carry out light weight processing, and a loss function is set to Focal Tversky Loss.
7. The wading concrete microcrack intelligent assessment system of claim 6 wherein the visual transducer unit is configured to extract global semantic features of an image comprising: Dividing an input image into a plurality of image blocks, embedding each image block through linear projection transformation, adding position embedding, and generating an embedding sequence; Inputting the embedded sequence to an encoder, and outputting enhanced semantic features after encoding by a multi-layer encoder; The convolution is optimized by adopting a depth separable convolution, the standard convolution operation is disassembled into two parts of depth-by-depth convolution and point-by-point convolution for calculation, and the output quantity ratio and the parameter quantity ratio are calculated.
8. The wading concrete microcrack intelligent assessment system of claim 6 wherein the multi-scale semantic interaction unit is configured to solve semantic information interaction problems between multi-scale features, comprising: Regularizing a plurality of feature images with different scales to the same channel dimension, and distributing the feature images into a boundary box with the same color through downsampling operation; Positioning related space perception image blocks from a plurality of continuous feature maps, reserving the space corresponding relation of the most related image blocks, and aggregating the space related image blocks on different scales by a scale fusion method to generate a mark; and inverting the obtained enhancement sequence into image patches according to the splicing sequence by using proportion segmentation, merging all patches with the same scale into enhancement feature mapping by patch inversion, fusing with CNN feature pyramid output, and further feeding into a decoder stage through jump connection.
9. The wading concrete microcrack intelligent assessment system of claim 6 wherein the graph convolution attention unit is configured to model the spatial topological relationship between the pixels of the crack, comprising: performing linear transformation on the nodes of each input node characteristic, and then calculating the importance coefficient of the first node to the second node through a shared attention mechanism; and aggregating the characteristics of adjacent nodes through multi-layer graph convolution, strengthening the difference between a crack area and a background area, and connecting a plurality of characteristics to obtain output characteristics.
10. The wading concrete microcrack intelligent assessment system of claim 6 wherein the channel collaborative attention unit consists of shareable multi-semantic spatial attention and progressive channel self-attention; Dividing the features into a plurality of sub-features with the same size, and respectively carrying out one-dimensional convolution with different convolution kernel sizes on each sub-feature to extract multi-space semantic information; The progressive channel self-attention is carried out along the channel dimension by utilizing the multi-channel crack semantic information extracted by the sharable multi-semantic space attention, the sub-features obtained by decomposing the sharable multi-semantic space attention are fused with the features processed by PCSA, and the perception capability of the model on the crack semantic information is deepened.

Description

Wading concrete microcrack intelligent evaluation method and system based on visual depth model Technical Field The invention relates to the field of intelligent evaluation methods, in particular to a wading concrete microcrack intelligent evaluation method and system based on a visual depth model. Background In the current concrete structure inspection technology, particularly for monitoring surface cracks of hydraulic structures (such as dams, gates, embankments, etc.), the expansion thereof may cause serious deterioration of dam strength, stability, durability, etc. Traditional methods rely on a slit meter and manual visual inspection, which, although convenient to operate, are inefficient, have limited accuracy, and are difficult to implement in real time. In recent years, rapid development of unmanned aerial vehicle technology provides new possibilities for crack detection, enabling rapid acquisition of high resolution images of structural surfaces. However, the crack image acquired by the unmanned aerial vehicle has the characteristics of complex background, changeable crack morphology and scale, blurred edges and the like, and the conventional crack detection system provided with the deep learning model is difficult to achieve an ideal effect when extracting the crack characteristics and fusing multi-scale information. The system is easy to have the problems of insufficient feature extraction, high model complexity and difficulty in accurately dividing and quantitatively analyzing the cracks, and seriously influences the precision and reliability of crack detection. Furthermore, conventional deep learning models perform poorly in attempting to address class imbalance problems (e.g., significant proportion differences that exist between cracks and background pixels). The proportion of the crack in the image is small, so that the deep learning model tends to ignore the crack in the training process, and the crack cannot be effectively identified and quantified. Although many studies have proposed improved methods, these problems remain. The traditional loss function is easy to bias to the dominant category under the condition of category proportion bias, so that the accurate segmentation of the difficult category with small occupied crack is greatly influenced. Disclosure of Invention The invention aims to provide a wading concrete micro-crack intelligent evaluation method based on a visual depth model, which solves the problems of inaccurate concrete crack segmentation, high model redundancy and difficulty in quantitative analysis in a complex environment in the prior art. According to the technical scheme, the invention provides an intelligent evaluation method for the micro-cracks of the wading concrete based on a visual depth model, which comprises the following steps: collecting a concrete structure image through an unmanned aerial vehicle and preprocessing; And segmenting a crack region of the acquired concrete structure image through a GANFormerNet segmentation model, wherein the segmentation method comprises the following steps: Extracting global semantic crack characteristics of an image, optimizing and fusing crack characteristics crossing different scales, modeling a space topological relation among crack pixels, and dividing a crack region of a concrete structure image; And carrying out pixel-level quantitative analysis on the crack region based on the output segmentation result, and integrating the analysis result into a user interface system. The invention further provides a wading concrete microcrack intelligent evaluation system based on a visual depth model, which comprises the following steps: The image acquisition module is used for acquiring a concrete structure image through the unmanned aerial vehicle and preprocessing the concrete structure image; The image segmentation module is used for segmenting a crack region of the acquired concrete structure image through a GANFormerNet segmentation model, wherein the segmentation method comprises the following steps: Extracting global semantic crack characteristics of an image, optimizing and fusing crack characteristics crossing different scales, modeling a space topological relation among crack pixels, and dividing a crack region of a concrete structure image; And the analysis evaluation module is used for carrying out pixel-level quantitative analysis on the crack area based on the output segmentation result and integrating the analysis result into the user interface system. The method has the advantages that compared with the prior art, the method has the advantages that 1, multi-scale crack characteristics are fully extracted and fused through multi-module cooperative work, the segmentation precision of complex crack forms is remarkably improved, 2, the perception capability of a crack space topological structure and key characteristics is enhanced through drawing a picture convolution attention and space channel cooperative attentio