Research On Dam Surface Crack Detection Method Based On Weakly-supervised Learning

With the development and progress of intelligent detection technology,automation control technology,and surveying technology,water conservancy hub surveying is becoming increasingly integrated,standardized,and refined.Periodic inspection of the surface of dams to determine whether cracks or other surface defects have occurred plays an important role in the operation and maintenance of infrastructure such as hydroelectric dams.After completing crack detection,determine the changes in crack morphology and location based on the detection results,infer the causes of cracks and defects on the surface of the dam,and provide data support and scientific guidance for risk assessment of the dam structure.This project focuses on the practical situation of difficult to reach,difficult to detect,and high detection costs faced by concrete dams.It focuses on solving the problems of high cost of obtaining pixel level annotations for cracks on the surface of dams and low accuracy of crack detection.A comprehensive analysis is conducted on the entire process from crack data collection to intelligent crack recognition.The specific research content is as follows:In response to the difficulty in obtaining data on dam surface cracks caused by the special environmental location of the dam,a drone combined with close proximity photography technology was used to complete data collection,and the collected data was screened and preprocessed to complete the construction of the dam surface crack dataset,providing reliable data support for subsequent research.To address the issue of high cost in obtaining pixel level annotation data used in training crack detection models.This article achieves crack pseudo annotation acquisition by combining image block based classification network localization with image block based multi class threshold segmentation method.By using a crack classifier based on convolutional neural networks,pixel level pseudo annotated images can be created using only image level annotation.By combining the classifier with gradient activation mapping based on image blocks,rough crack location images are obtained,and then the image blocks are used as input to achieve weakly supervised crack pseudo annotation generation using threshold segmentation algorithms.The generated crack pseudo annotations achieved0.5638 and 0.5839 at the optimal image scale(OIS)and optimal data scale(ODS),indicating good quality.This paper proposes a dam surface crack detection method based on the improved Deep Lab V3+model to address the problem of poor detection performance caused by the low proportion of crack pixels in the surface image of hydropower stations.The model enhances its feature extraction capability by improving the feature extraction network;Add a three-line attention module to enable the model to better pay attention to crack pixels;Dice Loss and Focal Loss are used together as the loss function of the model to reduce the impact of pixel imbalance and improve the detection accuracy of the model.Combining pixel calibration and morphological information extraction to quantify the length and width information of cracks.The improved model has an increase of 2.89%,1.12%,and 3.33%compared to the MIo U,MPA,and F1 Score before the improvement,and the overall parameter count has been reduced to 3014714.This article provides experimental evidence for the proposed method.The experimental results show that the method used in the project can generate high-quality pixel level crack pseudo annotations,achieve more accurate segmentation of surface cracks in concrete dams,and provide accurate quantitative information on crack morphology.This provides important reference data for the risk assessment and subsequent maintenance of the state of the concrete dam surface,and is of great significance for the engineering practice of dam surface defect monitoring of hydropower stations.