Research On Damage Characteristic Recognition Of Concrete CT Images Based On Deep Convolution Neural Network

Concrete is a special complex material with natural defects.The detection of internal damage characteristics is an important aspect of damage evolution mechanism of concrete meso-structure.In traditional methods,the damage characteristics are largely interfered by noise,which affects the identification effect of porosity and crack.The convolutional neural network has a certain degree of non-deformation for noise and deformation.It automatically learns useful features according to different task requirements,helps the algorithm to complete the detection and recognition,and has a stronger ability to express targets.Therefore,the improved convolutional neural network is adopted to detect the porosity and crack.The candidate region proposal represented by Faster R-CNN and the regression represented by YOLO are studied.The main contents include:(1)Based on the Faster R-CNN,the deep residual network ResNet-101 and ResNet-50 as the main framework.Feature Pyramid Network(FPN)and ROI Align are introduced into Faster R-CNN network.Using FPN is to generate high-quality pyramid feature maps.ROI Align is to solve the regional mismatch caused by quantization operation in ROI Pooling.Experiments show that the improved ResNet-101+FPN+ROI Align and ResNet-50+FPN+ROI Align are reaches 87.08% and 81.36%,which increase 4.74 and 3.12 respectively.But the detection time of single picture is slower than the original algorithm.(2)K-means and Batch Renormalizationare are introduced into YOLOv3.Using k-means to obtain appropriate Anchor Box and compare the different numbers of Anchor Box,and Batch Renormalization to solve the over-fitting due to many layers and computation.Experiments show that YOLOv3 eliminates the candidate region extraction and turns the detection into regression.It can predict the position and category of multiple bounding boxes at one time,increasing the detection speed and reaching an average of 0.0672s per time,achieving end-to-end detection.However,no region extraction will make the target positioning inaccurate,declining in the detection accuracy of YOLOv3.Based on Faster R-CNN and YOLOv3,combined with the new theoretical achievements,this paper proposes the improved Faster R-CNN and improved YOLOv3 for concrete damage characteristics detection.On the basis of theoretical research and engineering application background,it provides a certain technical support for the study of concrete damage mechanism.