Method And Application Of High-precision And Fast Recognition Of Pavement Crack Image Based On Multi-scale And Machine Learning

In recent years,China’s highway mileage has increased significantly.The highway nondestructive testing and the maintenance management have increasingly become an important research direction.Pavement cracks are one of the most common pavement diseases.Crack recognition and extraction based on pavement images have excellent research value for pavement maintenance.In this study,based on machine learning and digital image processing methods,the crack region recognition and extraction technology in the road surface image were studied.The main research contents are as follows:First,the YOLO v3 model was used to quickly recognize cracks on high-precision road images with full lane width.The YOLO v3 neural network model uses a Darknet-53 neural network with 53 convolutional layers and a multi-scale fusion feature network,which offers the benefit to improve the detection of small-scale objects.The results show that YOLO v3-based lane-scale pavement image crack detection has the advantages of fast recognition processing speed and relatively high recognition accuracy,and can realize real-time pavement crack detection and marking based on road surface information collection vehicles.Secondly,spatial domain filtering and grayscale transformation were applied to suppress the image noise in the road image and improve the contrast of the two-dimensional image of the road surface.The road image based on the characteristics of the line array camera system and the modified top-hat / bottom-hat transformation was proposed and applied,respectively,to eliminate the effect of uneven exposure on the two-dimensional image of the road surface.The normalization of the two-dimensional image of the road surface based on the optimal global threshold was proposed and applied to improve the effect of identifying and extracting crack areas in the road surface image.A series of image preprocessing algorithms and image adaptive normalization algorithms used in this study effectively reduce the diversity of road surface images and improve the processing effect of road surface crack target detection and crack area segmentation using road surface images as processing objects.The subsequent extraction of pavement crack areas laid a good foundation.Thirdly,an adaptive and improved pavement crack extraction method based on region growing was proposed to identify and extract crack areas from preprocessed and normalized pavement images.The crack area extraction method effectively suppresses the influence of disturbing factors such as uneven exposure and high-brightness pavement texture on the complex pavement background on the identification and extraction of pavement crack areas and obtains a complete pavement crack area.It has achieved satisfactory results in pavement images obtained by different pavement information acquisition systems,especially crack images with complex pavement backgrounds,which are difficult to obtain ideal results with classic algorithms,and are very robust.In addition,the concept and method of road surface image normalization have achieved good results in the extraction of road surface crack areas based on digital image processing technology,to achieve the effect of suppressing the influence of road surface high-brightness texture on the extraction of crack areas,and effectively improved the classic algorithm Performance...Finally,the morphological close operator was used to connect the crack regions obtained by image segmentation,and the morphological refinement was used to extract the crack region feature information.At the same time,the design of batch processing architecture of the road image crack recognition and crack region extraction system were completed.