CN-121999437-A - Method and system for checking hidden danger of dykes and dams

CN121999437ACN 121999437 ACN121999437 ACN 121999437ACN-121999437-A

Abstract

The invention relates to the technical field of dam safety detection and discloses a dam hidden danger investigation method and a system, wherein the invention processes the vision dam image by utilizing a target detection model through acquiring the vision dam image acquired in a multi-mode, outputs at least one boundary frame information with hidden danger targets, then cuts the boundary frame to obtain a local image with the hidden danger targets, processes the local image by utilizing a semantic segmentation model, outputs a pixel-level segmentation mask corresponding to the hidden danger targets, based on the pixel level dividing mask, the quantitative geometric parameters of hidden danger targets are calculated, objective data support is provided for risk research and judgment of subsequent dyke hidden danger, subjective judgment errors of manual detection are avoided, automatic check of dyke hidden danger is realized, manpower and cost of manual inspection are greatly reduced, automatic collection, detection and quantitative analysis of dyke images are realized, dyke hidden danger can be found in time, correction is performed, and occurrence risk of dyke safety accidents is reduced.

Inventors

HU JINGWEN
WU ZHIHONG
WANG QIXIANG
DONG WENLONG
FU TIAN

Assignees

山东省海洋预报减灾中心

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (11)

1. The method for checking hidden danger of the dam is characterized by comprising the following steps: acquiring a multi-mode acquired visual dyke image; processing the visual dyke image by using a target detection model, and outputting at least one boundary frame information of a hidden danger target; cutting the boundary frame to obtain a local image where the hidden danger target is located; Processing the local image by using a semantic segmentation model, and outputting a pixel level segmentation mask corresponding to the hidden danger target; And calculating the quantization geometric parameters of the hidden danger targets based on the pixel-level segmentation mask.
2. The method of claim 1, wherein the object detection model includes at least a backbone network, a neck network, and a detection head, the processing the visual dyke image with the object detection model outputs at least one bounding box information of a hidden object, comprising: Extracting features of the visual dyke images by using a backbone network to obtain a multi-scale feature map; Performing feature fusion on the multi-scale feature images by using a neck network to obtain multi-scale fusion feature images; and detecting hidden danger on the multi-scale fusion feature map by using a detection head, and outputting at least one bounding box information and class label with hidden danger targets.
3. The method of claim 2, wherein prior to hidden danger detection of the multi-scale fusion signature using a detection head, the method further comprises: Pooling operation is carried out on the fusion feature images of the current scale by pooling check of different sizes, so that a multi-scale pooling feature image is obtained; Performing attention weight calculation on the pooled feature map of the current scale to obtain a channel attention weight; Multiplying the channel attention weight with the pooled feature map of the current scale element by element in the channel dimension to obtain a pooled enhancement feature map of the current scale; fusing the fusion characteristic map of the current scale with the corresponding multi-scale pooling enhancement characteristic map to obtain a fusion enhancement characteristic map; updating the fusion enhancement feature map into the fusion enhancement feature map, and inputting the fusion enhancement feature map into a detection head for hidden danger detection.
4. The method of claim 1, wherein the acquiring the multi-modality acquired visual dyke images comprises: acquiring an original vision dyke image acquired in a multi-mode manner; adjusting a contrast ratio, a brightness offset and a haze ratio based on the brightness of the original visual dyke image; Performing environment inhibition enhancement on the visual dyke image based on the adjusted contrast ratio, brightness offset and haze ratio to obtain a first enhanced image; Acquiring different ocean background images; performing image fusion on the different ocean background images and the first enhanced image based on the pre-marked hidden danger areas to obtain a second enhanced image; And amplifying the pixel ratio of the hidden danger area in the second enhanced image to obtain the processed visual dyke image.
5. The method of claim 1, wherein the semantic segmentation model includes an encoder and a decoder, the decoder includes a detail branch and a semantic branch, the processing the local image using the semantic segmentation model outputs a pixel level division mask corresponding to a hidden danger target, comprising: performing layer-by-layer downsampling on the local image by using an encoder to obtain a shallow characteristic map and a deep characteristic map; Performing up-sampling processing on the shallow feature map by utilizing a detail branch of a decoder, and multiplying the up-sampled shallow feature map by a preset edge detection mask element by element to obtain an edge detail feature map; performing up-sampling processing on the deep feature map by utilizing semantic branches of a decoder to obtain a semantic feature map; based on the edge detail feature map and the semantic feature map, attention weights are calculated, and fusion attention weights are obtained; Based on the fused attention weight, fusing the semantic feature map and the edge detail feature map to obtain a final segmentation feature map; and carrying out binarization processing on the final segmentation feature map to obtain a pixel level segmentation mask corresponding to the hidden danger target.
6. The method of claim 1, wherein the calculating the quantization geometry of the hidden danger target based on the pixel-level segmentation mask comprises: when the hidden danger target is a crack, performing skeleton extraction on the crack segmentation mask to obtain a crack skeleton with single pixel width; calculating a crack length based on the first pixel total amount and the image resolution of the crack skeleton; Performing distance transformation on the crack segmentation masks to obtain the shortest distance from each crack mask pixel to the background; calculating the average value of the shortest distances corresponding to all the crack mask pixels to be used as the crack width; And when the hidden danger target is a dyke peeling or cavity, calculating the area of the hidden danger target based on the total second pixel amount and the actual single pixel area of the hidden danger target.
7. The method according to claim 1, wherein the method further comprises: Acquiring a double-time-phase remote sensing image pair, wherein the double-time-phase remote sensing image pair comprises a reference image and a monitoring image; performing standardized processing on the reference image and the monitoring image to obtain a first reference image and a second monitoring image corresponding to the current image block, wherein the standardized processing at least comprises coordinate system registration and fixed-size cutting images to obtain a plurality of image blocks; respectively carrying out feature extraction on the first reference image and the second monitoring image through a twin feature extraction network to respectively obtain a multi-scale first feature map and a multi-scale second feature map; Weighting and fusing the first feature map and the second feature map of the current scale to obtain a cross-layer feature map of the current scale; performing multi-type difference calculation on the first feature map and the second feature map of the current scale, and performing channel splicing on the difference feature maps subjected to all-type difference calculation to obtain the difference feature map of the current scale; performing channel splicing on the cross-layer feature map of each scale and the difference feature map of the corresponding scale to obtain a spliced feature map; carrying out fusion processing on the spliced feature images of all scales to obtain a global difference feature image; Upsampling the global difference feature map so that the size of the global difference feature map is the image size; the method comprises the steps of carrying out fusion processing on an up-sampled global difference feature map and a low-level cross-layer feature map to obtain an original change map, wherein the size of the original change map is consistent with the image size, the low-level comprises a first scale and a second scale, the first scale is the maximum size, and the second scale is an adjacent size smaller than the maximum size; Performing binarization processing on the original change map to obtain a binary change map; Splicing the binary change graphs corresponding to all the image blocks and performing edge restoration processing to obtain a spliced binary change graph; Multiplying a preset seawall vector boundary generation mask by the spliced binary change map to obtain a seawall risk change map; and marking the connected domains in the seawall risk change map, and calculating the geometric quantification parameter and risk type of each connected domain.
8. The method according to claim 1, wherein the method further comprises: Acquiring sea wall monitoring video frame data; Abnormal target identification is carried out on the video frame data by utilizing a target detection model; Extracting time sequence characteristics of an abnormal target by using a time sequence behavior analysis model, performing behavior recognition on the time sequence characteristics of the abnormal target, and judging behavior types of the abnormal target, wherein the behavior types are classified into abnormal behaviors and normal behaviors.
9. The method of claim 8, wherein the object detection model comprises at least a backbone network, a neck network, and a detection head, the neck network being integrated with a small object feature enhancement layer, the using the object detection model to perform outlier object recognition on the video frame data comprising: extracting features of the video frame data by using a backbone network, and outputting a feature map of a first scale; Performing convolution operation on the feature map of the first scale by using a small target feature enhancement layer to obtain an enhanced small target detail feature map; fusing the small target detail feature map with a feature map of a third scale output by a neck network to obtain a fused feature map, wherein the third scale is a size between a minimum size and a maximum size; And carrying out abnormal target identification on the fusion characteristic map by using a detection head.
10. The method of claim 8, wherein the time series behavior analysis model is an improved time convolution network, wherein a multi-layer residual block stacking structure is added to the time series behavior analysis model, wherein each residual block comprises a cavity convolution layer, the expansion rate of the cavity convolution in the residual block is positively correlated with the level of the residual block, and the time series receptive field of the time series behavior analysis model is exponentially enlarged; And before the output layer of the time convolution network, a time sequence attention module is added and used for converting the extracted time sequence characteristics into attention weights through a multi-layer perceptron, multiplying the time sequence characteristics by the attention weights element by element and carrying out self-adaptive weighting on the time sequence characteristics output by the model to obtain the time sequence characteristics subjected to attention weighting.
11. A dam hidden trouble investigation system is characterized in that the system comprises a data perception module and an intelligent analysis module, wherein, The data perception module is used for acquiring the vision dyke images acquired in multiple modes; The intelligent analysis module is used for processing the visual dyke image by utilizing the object detection model, outputting at least one boundary frame information with hidden danger objects, cutting the boundary frame to obtain a local image with the hidden danger objects, processing the local image by utilizing the semantic segmentation model, outputting a pixel level segmentation mask corresponding to the hidden danger objects, and calculating the quantization geometric parameters of the hidden danger objects based on the pixel level segmentation mask.

Description

Method and system for checking hidden danger of dykes and dams Technical Field The invention relates to the technical field of seawall safety detection, in particular to a method and a system for checking hidden danger of a dam. Background The seawall is used as a core water conservancy protection project for resisting storm surge, seawater backflow and coast erosion in coastal areas, is a 'life line' for guaranteeing life and property safety of coastal people and supporting economic and social development of areas, has long been mainly dependent on traditional modes such as manual inspection, regular field inspection and the like, and has a plurality of defects which are difficult to overcome, including low efficiency, the seawall is distributed along a long coastal zone, the topography is complex, manual hiking inspection is time-consuming and labor-consuming, full-coverage inspection of full-dike sections and high frequency is difficult to realize, the omission rate is high, the manual visual recognition capability is limited, omission and response lag are easy to occur, the manual inspection mode cannot be monitored in real time, and the optimal position opportunity is missed after dangerous situations occur. Disclosure of Invention The invention provides a method and a system for checking hidden dangers of a dyke, which are used for solving the problems that manual inspection is time-consuming and labor-consuming, full-coverage check of a full dyke section is difficult to realize and the omission rate is high. The invention provides a dam hidden danger investigation method, which comprises the steps of obtaining a multi-mode collected visual dam image, processing the visual dam image by utilizing a target detection model, outputting at least one boundary frame information of a hidden danger target, cutting the boundary frame to obtain a local image of the hidden danger target, processing the local image by utilizing a semantic segmentation model, outputting a pixel-level segmentation mask corresponding to the hidden danger target, and calculating the quantization geometric parameters of the hidden danger target based on the pixel-level segmentation mask. According to the dam hidden danger investigation method, the visual dam images acquired in multiple modes are acquired, the target detection model is utilized to process the visual dam images, at least one boundary frame information with hidden danger targets is output, then the boundary frames are cut to obtain local images with hidden danger targets, the semantic segmentation model is utilized to process the local images, the pixel-level segmentation mask corresponding to the hidden danger targets is output, the quantitative geometric parameters of the hidden danger targets are calculated based on the pixel-level segmentation mask, objective data support is provided for risk study and judgment of subsequent dam hidden danger, subjective judgment errors of manual detection are avoided, automatic investigation of the dam hidden danger is realized, manpower and cost of manual inspection are greatly reduced, automatic acquisition, detection and quantitative analysis of the dam images are realized, scale and normal monitoring of the dam hidden danger is supported, the dam hidden danger can be found in time and corrected, and the occurrence risk of a dam safety accident is reduced. In an optional implementation manner, the target detection model at least comprises a backbone network, a neck network and a detection head, wherein the target detection model is used for processing the visual dyke image and outputting at least one bounding box information with hidden danger targets, and the method comprises the steps of extracting features of the visual dyke image by using the backbone network to obtain a multi-scale feature map; and detecting hidden danger of the multi-scale fusion feature map by using a detection head, and outputting at least one boundary frame information and category label with hidden danger targets. According to the invention, the multi-scale characteristics are processed through the backbone network, the neck network and the detection diagram, the detection requirements of the ocean dykes and dams on multiple sizes and multiple distances are accurately adapted, and the detection precision and efficiency of the ocean dykes and dams are doubly improved. In an alternative embodiment, before hidden danger detection is performed on the multi-scale fusion feature map by using a detection head, the method further comprises the steps of carrying out pooling operation on the current-scale fusion feature map by pooling cores of different sizes to obtain the multi-scale pooling feature map, carrying out attention weight calculation on the current-scale pooling feature map to obtain a channel attention weight, carrying out channel dimension element-by-element multiplication on the channel attention weight and the current-scale pooling feature map to