CN-122023341-A - Dyke global hidden danger detection method based on radar image interpretation and vision large model

CN122023341ACN 122023341 ACN122023341 ACN 122023341ACN-122023341-A

Abstract

The invention discloses a dam global hidden danger detection method based on radar image interpretation and a vision large model, which comprises the following steps of collecting dam data to obtain three-dimensional radar imaging data, a vision image and corresponding dam GPS coordinates, denoising the three-dimensional radar imaging data to obtain a standardized radar feature map, obtaining surface disease position labeling information based on the vision image to form a standardized multimode feature map, inputting the standardized multimode feature map into a radar feature enhancement Attention module RFE-Attention to output an enhancement feature map, inputting the enhancement feature map into a special detection framework CS-VIT to obtain a multi-category disease detection result, executing cross-domain self-adaptive training optimization on the CS-VIT to obtain an updated multi-category disease detection result, and generating a dam three-dimensional hidden danger distribution map. The method improves the detection precision and the generalization robustness of the global hidden danger of the dykes and dams.

Inventors

SHEN GUODONG
XU SHAOBIN
ZHAO KANGWEN
Sun Ruopei
CAI JIANFENG
NIU XIN
WANG BEI
MING PAN
NIE WENHUA
HUANG MIAO
LI HAO
ZHANG XIANGWEI
YANG YIXIN

Assignees

安徽省交通规划设计研究总院股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (8)

1. The method for detecting the global hidden danger of the dykes and dams based on radar image interpretation and a visual large model is characterized by comprising the following steps of: acquiring dam data by using a mobile inspection platform carrying a three-dimensional ground penetrating radar to obtain three-dimensional radar imaging data, a visual image and corresponding dam GPS coordinates; Denoising the three-dimensional radar imaging data, and carrying out coordinate registration with the dam GPS coordinates to obtain a standardized radar feature map; obtaining surface disease position labeling information based on a visual image, and performing spatial alignment with a standardized radar feature map to form a standardized multimode feature map; Inputting the standardized multimode feature map into a radar feature enhancement Attention module RFE-Attention, performing feature enhancement processing through a feature decomposition unit, a double-branch weight generator and a weight fusion unit, and outputting an enhancement feature map; Inputting the enhanced feature map into a class exclusive detection architecture CS-VIT, and carrying out feature extraction through a shared feature extraction layer, a class exclusive training branch and a cross-branch fusion layer to obtain a multi-class disease detection result; performing cross-domain self-adaptive training tuning on the CS-VIT, and detecting and updating the multi-category disease detection result based on the tuned CS-VIT to obtain an updated multi-category disease detection result; And correlating the updated multi-category disease detection result with the dam GPS coordinates and the surface layer disease position labeling information to generate a dam three-dimensional hidden danger distribution map.
2. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the obtaining of the three-dimensional radar imaging data, the visual image and the corresponding dyke GPS coordinates specifically comprises: Based on the three-dimensional radar data acquisition module, a mobile inspection platform carrying the three-dimensional ground penetrating radar is provided, the mobile inspection platform moves along a dam inspection path and performs data acquisition in the moving process, and a visual image and a dam GPS coordinate corresponding to an acquisition position are synchronously acquired in the acquisition process; And the three-dimensional ground penetrating radar is in a working state of 200MHz to 500MHz in the central frequency, three-dimensional detection is carried out on the depth range of 0m to 5m of the dam body, three-dimensional radar imaging data are obtained, the three-dimensional radar imaging data output the distribution of the space scatterers of the dam body in a point cloud mode, and the radar profile imaging result along the inspection path is output in a gray profile graph.
3. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the obtaining of the standardized radar feature map specifically comprises: Performing wavelet threshold denoising processing on the three-dimensional radar imaging data to eliminate electromagnetic interference, performing wavelet decomposition and thresholding processing on coefficients obtained by the decomposition, performing inverse wavelet reconstruction on the coefficients subjected to the thresholding processing, and outputting the denoised three-dimensional radar imaging data; Performing coordinate registration processing on the denoised three-dimensional radar imaging data, determining the corresponding relation between each acquisition position in the denoised three-dimensional radar imaging data and the dam GPS coordinates, binding, and outputting the bound radar data; And carrying out standardization processing on the bound radar data, unifying the bound radar data into a preset coordinate reference and a preset data format, and outputting a standardized result as a standardized radar feature map.
4. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the forming of the standardized multimode feature map specifically comprises: Marking the surface layer diseases on the visual image, wherein the marking comprises marking positions, marking lengths and marking widths of the surface layer diseases, and outputting marking information of the surface layer diseases; Establishing a corresponding relation between a visual image and a standardized radar feature map at the same acquisition position according to the dam GPS coordinates; Mapping the surface disease position marking information into the space range of the standardized radar feature map based on the corresponding relation, obtaining the mapped surface disease position marking information, fusing the mapped surface disease position marking information with the standardized radar feature map, and outputting the standardized multimode feature map.
5. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the output of the enhanced feature map specifically comprises Inputting a standardized multimode characteristic diagram into a radar characteristic enhanced Attention module RFE-Attention, wherein the RFE-Attention comprises a characteristic decomposition unit, a double-branch weight generator and a weight fusion unit, and extracting radar channels, namely reflection intensity data, from the standardized multimode characteristic diagram as data to be processed; Performing feature decomposition on the reflection intensity data by a feature decomposition unit, performing sliding calculation on the reflection intensity data by adopting a three-by-three window, taking the variance of reflection values in the window as a local scattering feature, and taking the phase difference of the reflection waves along the depth direction as a global layering feature; the method comprises the steps that a dual-branch weight generator respectively performs weight mapping processing on local scattering features and global layered features, the local scattering features are mapped through a Sigmoid function to output local Attention weights, the global layered features are mapped through a Softmax function to output layered Attention weights, and the local Attention weights and the layered Attention weights are multiplied to obtain an RFE-Attention weight map; The RFE-Attention weight map is multiplied by the normalized multimode feature map pixel by a weight fusion unit, and an enhanced feature map is output.
6. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the obtaining of the multi-category disease detection result specifically comprises: Inputting the enhanced feature map into a class exclusive detection architecture CS-VIT, wherein the CS-VIT is a large visual model and comprises a shared feature extraction layer, class exclusive training branches arranged according to the number of disease categories and a cross-branch fusion layer, and the class exclusive training branches comprise a VIT encoder, a presence judgment head and a target frame regression head; sequentially performing convolution processing and pooling processing on the enhanced feature map in the shared feature extraction layer, and compressing the feature channels into two hundred fifty-six channels through four-layer convolution and pooling operation to obtain the shared feature map; Executing block processing on the shared feature map in each type of exclusive training branch, dividing the shared feature map into blocks according to a preset block rule, and forming a block sequence, so that the block sequence comprises sixty four blocks; Coding the block sequence through six layers of convectors in a VIT coder to obtain class-specific characteristics corresponding to the current disease class; Sequentially executing full connection processing and Sigmoid activation processing on the special class characteristics in the presence judgment head to obtain a judgment result of whether the current disease class exists or not, and obtaining detection confidence corresponding to the judgment result; continuously performing four times of full connection processing on the special class characteristics in the target frame regression head to obtain regression results for determining the left upper corner position and the right lower corner position of the target frame, and forming branch detection results of the current disease class; And in the cross-branch fusion layer, performing weighted voting fusion processing on branch detection results output by various proprietary training branches, weighting according to the detection confidence level, selecting an optimal result, and outputting a detection frame and a class label of all classes of diseases as multi-class disease detection results.
7. The method for detecting global hidden danger of a dyke based on radar image interpretation and visual large model according to claim 1, wherein the obtaining of the updated multi-category disease detection result specifically comprises: Reading multi-category disease detection results, determining three-dimensional radar imaging data with disease category and target frame marking information as marked radar data, otherwise determining unmarked radar data, dividing the unmarked radar data into different regional data sets according to regional sources, and forming a cross-domain adaptive training optimized input data set; Executing cross-domain data enhancement processing, calling a domain adaptation network to transfer and align radar imaging distribution of different regional data sets, and outputting aligned unlabeled radar data as cross-domain enhancement data; Executing radar physical enhancement sample generation processing, calling a radar physical enhancement operator to enhance and generate marked radar data and cross-domain enhancement data, adjusting reflection intensity attenuation in propagation difference simulation processing, superposing region specific noise, and outputting region self-adaptive enhancement samples; performing dynamic weight loss training preparation processing, setting training loss weights corresponding to the current diseases aiming at each type of diseases, and introducing constant items for avoiding zero training loss weights; Performing dynamic weight loss training processing, counting sample frequency of each disease category in real time in the training process, counting detection confidence coefficient of each disease category in real time, updating loss weight corresponding to each disease category, weighting classification loss and regression loss of CS-VIT by adopting the updated loss weight, performing model training updating at the same time, and inputting labeling radar data and region self-adaptive enhancement samples as training data to obtain a CS-VIT model with staged optimization; Performing semi-supervised pseudo tag iterative training processing, detecting cross-domain enhanced data by using a CS-VIT model with phased tuning, screening detection results according to detection confidence, selecting detection results with the detection confidence not lower than 0.85 as pseudo tags, mixing labeling radar data and the pseudo tag data to form an iterative training set, performing iterative training update on the CS-VIT model, and updating the pseudo tags after each iterative training to obtain a tuned CS-VIT model; And performing detection updating on the multi-category disease detection result by using the optimized CS-VIT model, performing cross-branch fusion processing, and outputting the updated multi-category disease detection result.
8. The method for detecting global hidden danger of a dam based on radar image interpretation and visual large model according to claim 1, wherein the generating of the dam three-dimensional hidden danger distribution map specifically comprises Performing coordinate association processing on the updated multi-category disease detection results, binding the space position corresponding to each disease detection result with the dam GPS coordinates, and writing the space position into a data record corresponding to the dam GPS coordinates to form a disease detection record set with coordinates; Synchronously executing coordinate association processing on the surface disease position marking information to form a surface disease marking record set with coordinates; Performing three-dimensional summarization processing based on disease detection records with coordinates and surface disease position marking information with coordinates, converging disease categories, target frames and detection confidence corresponding to the same dyke GPS coordinates, converging surface disease marking positions, marking lengths and marking widths corresponding to the same dyke GPS coordinates, and performing three-dimensional distribution organization on the converged surface disease information and deep disease information according to dyke space positions to obtain a three-dimensional hidden danger information set; And executing CT result output processing on the three-dimensional hidden danger information set, marking the marking position, the marking length and the marking width corresponding to the surface layer diseases by combining the visual image, writing the marking depth, the marking volume and the radar reflection characteristic corresponding to the deep layer diseases into the corresponding space positions, judging whether the surface layer cracks are communicated with the deep layer leakage channels through multi-mode data association and writing the surface layer cracks into the corresponding space positions, and writing the position, the depth and the risk grade information into the corresponding space positions together to generate the dam body three-dimensional hidden danger distribution map.

Description

Dyke global hidden danger detection method based on radar image interpretation and vision large model Technical Field The invention relates to the technical field of intelligent inspection and hidden danger detection of dykes and dams, in particular to a dykes and dams global hidden danger detection method based on radar image interpretation and a visual large model. Background The dam is used as a core infrastructure of flood control and water supply engineering, early identification of surface and deep hidden trouble of the dam is of great significance to engineering safety, the existing inspection means mainly comprises manual inspection combined with ground penetrating radar and drilling verification and an automatic detection system facing a single mode, the manual inspection relies on personnel experience to observe surface layer anomalies such as dam face cracks, the deep detection is carried out by matching with a handheld ground penetrating radar, and the inspection is carried out by the drilling means. Along with the development of a target detection model and a visual large model, the prior art starts to try to use a general detection network for identifying radar images and engineering structural diseases, but has obvious defects in a dam three-dimensional radar imaging scene, three-dimensional radar imaging data comprise physical characteristics such as reflection intensity, phase difference, layered structure and the like, the data distribution and natural image difference are obvious, the general convolution characteristic extraction and standard attention mechanism are difficult to highlight the difference characteristics of diseases and backgrounds, so that the characteristic extraction suitability is insufficient and missed detection is generated, on the other hand, dam disease labeling is usually finished by combining radar interpretation with field verification by a professional, the cost is high and the period is long, labeling samples are scarce and the category distribution is unbalanced, the general multi-category sharing training frame is easy to generate category bias, the detection capability of small sample categories is insufficient, and full category coverage is difficult to realize. Therefore, how to provide a method for detecting global hidden danger of a dyke based on radar image interpretation and a large visual model is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a dam universe hidden danger detection method based on radar image interpretation and a visual large model, which fuses three-dimensional ground penetrating radar imaging and visual annotation, adopts RFE-Attention to enhance disease characteristics, combines CS-VIT type exclusive detection and cross-domain pseudo tag tuning, realizes small annotation, cross-region dam hidden danger positioning and risk classification, and effectively improves detection precision and generalization robustness. The method for detecting the global hidden danger of the dykes based on the radar image interpretation and the visual large model comprises the following steps: acquiring dam data by using a mobile inspection platform carrying a three-dimensional ground penetrating radar to obtain three-dimensional radar imaging data, a visual image and corresponding dam GPS coordinates; Denoising the three-dimensional radar imaging data, and carrying out coordinate registration with the dam GPS coordinates to obtain a standardized radar feature map; obtaining surface disease position labeling information based on a visual image, and performing spatial alignment with a standardized radar feature map to form a standardized multimode feature map; Inputting the standardized multimode feature map into a radar feature enhancement Attention module RFE-Attention, performing feature enhancement processing through a feature decomposition unit, a double-branch weight generator and a weight fusion unit, and outputting an enhancement feature map; Inputting the enhanced feature map into a class exclusive detection architecture CS-VIT, and carrying out feature extraction through a shared feature extraction layer, a class exclusive training branch and a cross-branch fusion layer to obtain a multi-class disease detection result; performing cross-domain self-adaptive training tuning on the CS-VIT, and detecting and updating the multi-category disease detection result based on the tuned CS-VIT to obtain an updated multi-category disease detection result; And correlating the updated multi-category disease detection result with the dam GPS coordinates and the surface layer disease position labeling information to generate a dam three-dimensional hidden danger distribution map. Optionally, the obtaining the three-dimensional radar imaging data, the visual image and the corresponding dam GPS coordinates specifically includes: Based on the three-dimensional radar data acquisition module, a mobile inspec