Development Of Intelligent Detection And Evaluation System For Steel Bridge Defect Based On Deep Learning

Steel structures are widely used in bridges because of its high strength,light weight,meeting the requirements of long-span,complexity and aesthetics.However,in the process of use,it is prone to coating deterioration,corrosion,weld cracking and other defects due to long-term natural erosion and human action,which affects the beauty of the bridge,reduces the structural strength of the steel bridge,and timely defect detection is of great significance.The traditional steel bridge defect detection mainly depends on manual detection and nondestructive detection,which is time-consuming,labor-intensive,high operation cost and strong subjectivity.Therefore,it is necessary to find an intelligent detection technology to carry out the detection task of steel bridge,so as to reduce the work difficulty,reduce the work cost and realize efficient and accurate defect detection.The development of computer vision and image processing technology provides power for intelligent defect detection,especially the emergence of deep learning,which can extract input data layer by layer and realize a good mapping from bottom signal to high-level semantics.In this paper,the steel bridge defect data set is established based on the deep learning technology.According to the characteristics of the data set,the steel bridge defect intelligent identification algorithm and segmentation quantization algorithm are developed,and a set of defect identification and evaluation system is compiled,which not only reduces the use and learning cost,but also improves the test efficiency,in order to realize the intellectualization of steel bridge defect detection.The main contents of this thesis are as follows:(1)Convolution neural network and transfer learning are used to intelligently identify steel bridge defects.Feature extraction and fine-tuning are used for training,and compared with the new training model.The single epoch time of feature extraction model is 47.2s,which is stable after about 100 times of training;The single epoch time of the fine-tuning model is 119.2s,which is stable after about 25 times of training,and the accuracy of 89.39% and 97.88% are obtained respectively.The transfer learning greatly shortens the training time and improves the operation efficiency.The fine-tuning model reduces over fitting and has the highest recognition accuracy.The fine-tuning model is selected as the defect recognition model to realize the rapid and accurate identification of steel bridge defects.(2)Aiming at the three local defect forms of steel bridge: coating deterioration,corrosion and weld cracking,an improved solov2 instance segmentation algorithm is proposed.The AP value obtained by the improved algorithm is 7.559% higher than tensormask algorithm,6.774% higher than blendmack algorithm and 5.192% higher than the unmodified solov2 algorithm.The detection ability of the model for small targets such as weld cracking is improved,and the segmentation effect is good;Comparing the actual value of defect size measured by image processing technology with the quantitative value of defect size output by the model,it is found that the relative error of defect area is concentrated within 10%,showing good detection and quantitative performance.(3)Aiming at the cumbersome configuration of the code running environment,a web version user interface is developed.After comparing several development schemes based on Python web framework,Django is used for web interface development,which realizes the functions of detection,evaluation and result export of specified pictures and videos,and avoids a series of cumbersome processes such as environment configuration,dependency library installation and code reading before users use the algorithm.