The Research Of Bridge Crack Automatic Identification Method Based On Improved Mask R-CNN Model

This paper proposes a bridge crack automatic identification method based on an improved Mask R-CNN model,and modularizes and systematizes the method for extensive application.The main tasks of this article were as follows:(1)The bridge crack dataset was built and the Mask R-CNN model was trained based on this dataset.Images of cracks in steel and concrete bridges were collected,the number and diversity of the data were enhanced using image broadening techniques,and the images were annotated using Labelme annotation software and then converted into files that could be used for training.The Mask R-CNN model was trained based on the established crack dataset and continuously optimised through parameter adjustment to obtain the optimal model weights to complete the classification,detection and segmentation of cracks.(2)Improvement strategies for the Mask R-CNN model were proposed.In order to address the issues of inaccurate object detection boxes,rough segmentation masks,and interrupted masks when dealing with small cracks in the Mask R-CNN model for crack detection,the attention module(CBAM)and path aggregation module(PAFPN)were combined with the feature extraction network to improve the focus on target region features and the comprehensive utilization of multi-scale features.Furthermore,a cascade multi-threshold detector was added to improve the detwasection accuracy under high thresholds.Subsequently,the models before and after improvement were compared and analyzed through experiments.The results showed that the improved Mask R-CNN model achieved m APs of 84.1% and 77.6% for crack detection and segmentation,which were 13.5%and 12.6% higher than those before the improvement,respectively.(3)The quantification of the geometric parameters of bridge cracks was completed.Firstly,the problem of broken crack masks was solved with morphological closure,and a complete crack mask was obtained.Secondly,Open CV related functions were used to perform operations such as connected domain labeling and skeleton extraction on the crack mask image to achieve pixel-based quantification.Finally,the transformation value between actual size and pixel size was determined based on the camera’s parameters and distance information at the time of shooting,and the length and width of the crack was measured.The experimental results show that the error was about 10%,which proves that the quantization method has practical application significance in engineering.(4)The design and packaging of an automated bridge crack identification system were realised.The Py Qt library in Python was used to design the GUI interface of the bridge crack automated detection system,which allows users to import,detect,quantify,and save images.The system and related installation packages were packaged into an executable EXE file using Pyinstaller,which allows users to efficiently and quickly obtain the results of bridge crack detection and quantification results,realizing the automation and intelligence of bridge crack detection.