Research On Bridge Crack Detection Algorithm Based On Deep Learnin

The intelligence,information,and digitization of bridge maintenance are gradually becoming additional dimensions to measure the engineering world due to the growing number of bridges.Cracking is a typical structural ailment in bridges made of concrete.Since manual methods are labor-intensive and may pose safety problems,automated methods of crack detection have attracted interest.In this paper,we design a human-computer interaction framework for bridge surface crack detection,which is centered on automatic detection of bridge surface cracks using deep learning-based target detection and semantic segmentation algorithms to provide strong support for the safety assessment of bridge structures.The main research work in this paper is as follows:1.In this thesis,we build a dataset of bridge cracks to train algorithms for crack identification and segmentation.The dataset construction involves image collection,preprocessing,and annotation.To enhance the robustness and generalization ability of the algorithms,effective data augmentation methods such as geometric transformation,optical transformation,and noise addition are utilized.Moreover,an advanced generative adversarial network model is developed to generate realistic bridge crack images,thereby increasing the diversity of the bridge crack dataset.2.The detection algorithm for surface cracks on bridges is applied using a target detection approach.In order to achieve algorithm acceleration to meet the needs of industrial applications,this paper proposes the CR-YOLO algorithm to achieve the trade-off between detection accuracy and speed in response to the problem of lightweight design leading to the sacrifice of algorithm detection accuracy.CR-YOLO is an improved crack detection algorithm based on the YOLO algorithm with improvement points focusing on the network structure,attention mechanism,loss function,and training techniques.The experimental findings demonstrate that the CR-YOLO algorithm’s accuracy,recall,and F1 are,respectively,91.93%,89.58%,and 90.73%.In addition,it uses only 23.8MB of storage space and performs best,averaging 120.4 frames per second.3.To achieve accurate identification of bridge crack boundaries and shapes,this study proposes an improved PSPNet semantic segmentation algorithm for the segmentation task of bridge surface cracks.Initially,a pre-trained lightweight convolutional neural network is employed to extract crack features.Next,to enhance the algorithm’s feature capture capability in the crack region,a dual self-attentive module is introduced to obtain more contextual information.The experimental findings indicate that the improved PSPNet crack segmentation algorithm’s accuracy,recall,and F1 are,respectively,87.84%,86.45%,and87.14%.In addition,it only occupies 10.3 MB of storage space and the FPS can reach 125.2on average,which is better than other baseline algorithms.Finally,this paper deployed and tested the algorithm on edge devices with limited computing power.The test results demonstrate that the edge devices can execute the bridge fracture detection and segmentation method in real-time.