| The construction of water conservancy projects is an important way to regulate the flow of water and remove the damage,of which dams are important water retaining structures,located in an environment with complex stresses,prone to cracks and other diseases in the long-term service process,if not detected and dealt with in a timely manner,the development of a dam may cause a dam failure,bringing disastrous consequences to people’s property,ecological environment and social economy.In this paper,we study the rapid detection of cracks based on unmanned aerial imagery,which is of great importance for the safe service of dams.At present,there are few methods to detect cracks in embankment dams,mainly based on manual visual inspection,which has problems such as low efficiency,subjectivity and easy to miss inspection.In addition,dams are characterised by long axes,dangerous detection environments and inconspicuous crack characteristics,making it difficult to detect cracks using conventional methods.With the development of UAVs and digital photography technology,the use of UAVs to capture images of the dam surface combined with image processing algorithms for crack detection can improve efficiency.However,previous crack detection algorithms based on image processing have disadvantages such as poor adaptability,low accuracy and slow speed.The more popular deep learning method has achieved significant results in the field of image recognition,with the advantages of strong fitting ability,high accuracy and fast speed,which is expected to make up for the shortcomings of image processing-based crack detection algorithms.However,deep learning requires a large amount of data to drive and is expensive to train.In addition,existing deep learning methods do not perform well on embankment crack data and are not suitable for detecting embankment cracks in UAV image.To address the above problems,this paper optimises the U~2-Net model and combines migration learning to achieve automatic detection and feature extraction of cracks in embankments.The main research elements are as follows:(1)To reduce the training cost of U~2-Net,a new residual block(RSU-ECA-AS)is proposed,on the basis of which the U~2Net-ECA-AS model is constructed for crack detection,and the new model has lower training cost and higher accuracy.Comparing the proposed model with a variety of commonly used deep learning crack detection models,an IOU of 80.45%and an F-measure of 88.88%were achieved,which verified the applicability of the model to dam crack detection.(2)To address the lack of crack detection datasets for embankment scenes,migration learning is performed based on open source building crack datasets,and a fine-tuning strategy using migration of shallow coding layer features is identified through experiments on the U~2Net-ECA-AS model.(3)Study the slice detection and alignment correction method based on UAV images,and realize the model’s crack detection of high-resolution images taken by UAVs at a safe distance away,and obtain good detection results.(4)A feature extraction study was conducted on the results of end-to-end crack detection of UAV images,and a corresponding feature extraction process was proposed.The cracks were divided into linear cracks and non-linear cracks,and the methods for calculating the length and width of cracks in different ways were studied in comparison.The article achieves automatic and non-contact detection and quantification of cracks in embankments,and applies the results to field crack detection in an embankment and a floodwall in Xinzhou District of the Yangtze River,with good results.It provides a new method for the detection of cracks in embankments and can provide inspection data for maintenance and repair. |
PDF Full Text Request