Bridge Crack Identification Based On Deep Learning And Conditional Random Field

Cracks on the surface of reinforced concrete bridges are one of the main diseases affecting the safety performance of bridges.The detection of cracks is a very important link in the appearance inspection of reinforced concrete bridges.At present,manual detection methods are widely used.Traditional detection methods based on manual detection have problems such as insufficient accuracy,low efficiency,and poor security.Therefore,an automatic bridge crack detection method based on machine vision and image recognition is urgently needed to replace the current manual detection method,so as to efficiently and accurately evaluate the health status of reinforced concrete bridges.In recent years,the continuous improvement of computer hardware performance and the rapid development of artificial intelligence have provided new ideas for the detection of bridge surface cracks.Based on the U-net model structure,this dissertation builds a reinforced concrete bridge crack recognition model based on full convolutional neural network,extracts the crack characteristics by FCN,and then uses the conditional random field as the back-end optimization module to modify the classification results of FCN.The specific work is as follows:1.A dataset of surface cracks in reinforced concrete bridges was established.Annotate800 original pictures of surface cracks of reinforced concrete bridges collected in the actual engineering environment are annotated.The original pictures and labels are cropped.The data is expended by means of rotation,mirroring,blurring,noise,etc.The reinforced concrete bridge surface crack dataset includes 3750 pieces of crack original image and 3750 pieces of label image,which are 128 pix × 128 pix.2.Extraction of crack characteristics on concrete surface.Based on the U-net model structure and the FCN model,a revised fully convolutional neural network model including the feature contraction stage and the feature expansion stage is built.The feature fusion method is used to enhance the model’s recognition ability.The highest accuracy of the model training verification is 98.77 %.3.Correct the fracture segmentation results of FCN.Based on the concrete crack image,a fully connected conditional random field model is established.The crack segmentation result of the front-end classifier Re-FCN model is used as input,and the CRF model is used as the back-end optimization module to optimize the segmentation result of the front-end classifier.4.Under laboratory conditions,the Re-FCN model + CRF model identification method is used to identify the concrete crack width,and the final width error can be controlled at5.64%.This dissertation uses the Re-FCN network model to extract pixel-level features from the surface cracks of reinforced concrete bridges,and optimizes the extraction results by CRF model.