Detection Method And Realization Of Self-attention For Dam Surface Defects

During the long-term service of the hydropower station dam,various defects are easily formed.At present,it mainly relies on manual inspection for maintenance.However,the manual inspection has problems such as long cycle,high risk and strong subjectivity.At the same time,the apparent image of the dam has the characteristics of complex background,coexistence of multiple types of defects,and extremely low proportion of defective pixels.From the perspective of improving the accuracy and practicability of the defect detection algorithm,this paper studies the self-attention detection method of the apparent defect of the dam body according to the current situation of the inspection of the apparent defect of the dam body and the characteristics of the image.The specific research work of this paper is as follows:(1)In view of the lack of reliable and effective defect datasets in water conservancy and hydropower scenarios,two high-quality dam apparent defect datasets were established,including an image-level annotated dam apparent defect identification dataset and a pixel dataset.The segmentation dataset of apparent cracks of dam body marked with the highest level can provide scene data support for subsequent research on dam body defect detection.(2)In order to quickly and objectively identify the types of apparent defects in the dam body,carry out research on the self-attention identification method of the apparent defects in the dam body.This method completely adopts the self-attention mechanism to construct the defect recognition network,which has stronger ability to capture long-distance global information and higher defect recognition accuracy.The network performs self-attention calculations at different scales through a dual-branch structure,extracts multi-scale defect features,and enhances the global semantic expression ability;a self-attention hybrid fusion module is designed to share the semantic features of the two branches,effectively dealing with the scale of the dam’s apparent defect image.Big changes,diverse forms,etc.The macro-accuracy rate of this method can reach 98.87%,and the defect identification effect is better than other mainstream defect identification methods.Using Tensor RT to deploy the dam apparent defect identification model to the workstation,the defect identification speed is increased by 22%.(3)In order to finely detect the apparent cracks of the dam body,research on the selfattention segmentation method of the apparent cracks of the dam body is carried out.The method uses the improved self-attention module as the basic component to build an encoding-decoding structure,the encoder extracts semantic features,and the decoder restores the feature resolution;the scaled self-attention module uses the crack features extracted by the encoder to generate a self-attention mask,Disambiguate between encoder and decoder feature sets to improve crack segmentation performance.The mixed loss function is used to balance the background loss and the crack loss,alleviate the extreme imbalance between the positive and negative samples of the apparent cracks of the dam body,and improve the segmentation index of the apparent cracks of the dam body.The average precision of this method is 78.41%,and the average intersection ratio is 67.21%.The crack segmentation effect is better than other mainstream segmentation networks.In summary,the research on the self-attention detection method of dam apparent defects carried out in this paper provides a technical reference for the safety and intelligent detection of hydropower stations,and has certain practical value.