Research On Concrete Meso-Damage Feature Detection Based On Deep Learning

The internal damage and failure modes of concrete show significant discreteness and randomness.Identification of internal damage characteristics is an important branch of concrete mesoscopic research.The traditional concrete meso-structure inspection method mainly uses image pixel processing to identify the damaged structure inside the concrete.Due to the particularity of the internal structure of concrete,the noise generated in CT image will affect the accuracy of identifying holes and cracks in concrete,thus affecting the meso damage research of concrete to a certain extent.Convolutional neural network can effectively eliminate the influence of noise in the image,and then accurately abstract the object features from the image.Therefore,this paper selects the target detection method based on deep learning to study the internal damage characteristics of the concrete CT image.This paper has designed an improved algorithm for concrete CT image detection difficulties,and proposed to use Cascade R-CNN as the main detection network to introduce Feature Pyramid Network(FPN)to construct a feature fusion Cascade R-CNN algorithm and a dual-head Cascade R-CNN detection based on feature sharing model.The main research contents are as follows:(1)Based on Cascade R-CNN,feature pyramid network is introduced to construct feature fusion Cascade R-CNN algorithm,the algorithm is used to identify the damage features of concrete CT image.The feature pyramid model is used to extract high-resolution feature maps with rich semantics,which effectively improves the identification accuracy of small target holes.The experimental results show that the average accuracy of ResNet-101 + FPN + Cascade R-CNN and ResNet-50 + FPN +Cascade R-CNN on the concrete CT image test set reaches 88.27% and 86.34%.Compared with ResNet-101 + Cascade R-CNN and ResNet-50 + Cascade R-CNN,the average accuracy is improved by 3.96% and 4.7%.Moreover,the introduction of FPN increases the parameters of the detection model,so the training and detection time of the model is also slightly improved.(2)On the basis of FPN’s Cascade R-CNN as the main structure,the Double-head structure of feature sharing is introduced.Taking advantage of the spatial sensitivity of fully connected head and the spatial correlation of revolution head,this paper proposes to combine the two forms to perform classification and regression tasks to effectively identify the spatial information of damage features.Secondly,the classification information of the detection network’s each level is fused by the method of feature sharing to improve the accuracy of low IOU(Intersection over Union)threshold(0.5-0.7)ROI(Regions of Interest)detection task.The experimental results show that the average accuracy of the proposed method on the concrete CT image test set is 3.04%higher than the FPN-based Cascade R-CNN,and the average accuracy of the low IOU threshold(0.5-0.7)ROI is improved by 1.49%.According to the actual engineering needs,this paper uses Cascade R-CNN as the main detection network,combined with the FPN architecture to build a feature fusion Cascade R-CNN network model and a double-head Cascade R-CNN network model for fusion feature sharing,through this model,the damage structure of concrete is visualized.The model can detect the mesoscopic damage part from the concrete CT image,and has the advantages of simple operation,low model complexity and convenient engineering application.