| Pavement distresses are the critical factors affecting road safety,and periodic inspection and maintenance play an important role in maintaining asphalt pavement performance.In recent years,with the rapid development of computer science and technology,it has laid a solid foundation for its wide application in many fields.Because of its intelligent characteristics,deep learning technology provides a trustworthy strategy in the field of infrastructure distress detection in roadway engineering.The deep learning model represented by Deep Convolution Neural Network(CNN)has achieved a leap-forward development in crack image recognition,crack object detection and crack semantic segmentation,and has shown great potential.However,there are still some shortcomings in multi-source data fusion and model efficiency.To solve the above problems,this paper aims to intelligently interpreting the structural information in the image of asphalt pavement distress based on deep learning.We focus on crack image recognition at coarse scale to crack object detection and crack semantic segmentation at fine scale.In this paper,computer vision and deep learning technology are employed as the main technical means,and improvements and innovations are made in dataset construction,model development,model comparison and analysis,etc.Firstly,this paper briefly reviews the composition of CNN,and introduces the principle of CNN’s back propagation,the optimization algorithm of the network,the types of loss functions,evaluation metrics,the causes of model over-fitting and the corresponding solutions.Considering the deployment of mobile terminals in the future,this paper determines the lightweight network architecture as the encoder network of the deep learning model.Secondly,surface images of a section of asphalt pavement are collected based on a multifunctional road inspection vehicle,and then the collected road images are cleaned to generate a road distress image data set LHRD.Due to the low contrast of the collected road images,the global adaptive image enhancement technology is adopted to improve the contrast of the images.Based on the depth wise separable convolution and Shuffle module,a lightweight network model was developed and designed.The model with the best evaluation metrics is selected as the distress image classification network,that is,the PCCNet model.The effects of weight penalty coefficient and optimizer on the prediction performance of the model are discussed,and the prediction performance is compared with that of the mainstream deep learning model at present.In this paper,the influence of image augmentation,i.e.image source,on the prediction performance of the training model is discussed,and the generalization performance of the model is analyzed.The lightweight network model is competent for pavement distress image classification,and has achieved good prediction performance.In addition,the prediction performance of deep learning model also has incomparable advantages over machine learning.The weight coefficient can improve the prediction performance of the training model by as much as 16.48%(m AP).Optimizer SGDM has the best prediction performance,and its m AP reaches 83.14%.The prediction performance of the training model depends on the sample size,the image quality in the samples and the image source with different acquisition equipment.Thirdly,with the help of the open tool Label Img,the location of the LHRD dataset is marked,providing data-driven for deep learning.To locate the distress location,a lightweight model for asphalt pavement distress detection is developed on the basis of the current mainstream YOLO model.From the aspects of activation function,loss function and optimizer,this paper explores the influence of hyper-parameters on the prediction performance of the training model,and compares different object detection models to verify the feasibility and reliability of the model proposed in this paper.Based on images with different illumination intensities,the generalization performance of this training model is verified.YOLO v5 s model is suitable for asphalt pavement distress detection,which can well balance detection accuracy and inference time.The activation function and loss function have the same influence on the prediction performance of training models.The training model can produce objective prediction results for distress images suitable for light intensity,but the detection results for strong and weak light images cannot meet the requirements.Then,since the contrast between the crack pixels and the surrounding pixels is low and discontinuous,to effectively extract the information of pavement crack,a fast non-local means denoising method is proposed,which eliminates Gaussian isolated noise while retaining target pixels.Based on the second derivative of Gaussian two-dimensional function,two directional filters are added to generate a group of basic filters,which are used to calculate the maximum response of a certain position in different directions and generate a crack saliency map.The local order energy is used to extract the crack shape from the crack saliency map,and the binary image of the crack is generated.The location and direction of the crack are located by the mathematical morphology operation method.Noise reduction has a significant impact on the segmentation results of asphalt pavement cracks,and the segmentation performance of the model is improved by 12.19% after noise reduction.The proposed LOF model relieves the local minimum of LBF model and achieves the best prediction performance.Compared with the LBF model and CV model,m Io U of LOF model increased by 6.46% and 12.27%,respectively.In addition,this method can provide a method reference for semantic annotation of pavement images.Finally,in view of the fact that the local order energy method cannot meet the requirements of fast and efficient segmentation,and to realize intelligent segmentation of asphalt pavement cracks,a semantic segmentation model of pavement cracks based on deep convolution neural network was proposed in this article.In this model,depth wise separable convolution is employed to reduce the model parameters,and generalized Dice loss is employed to replace the traditional cross entropy loss function,so as to achieve high-precision segmentation performance.PCSNet model trains on an open image datasets,and compares the segmentation performance with other deep learning models and traditional image processing techniques on the test set.The results show that lightweight CNN with generalized Dice are helpful to improve the forecasting performance of the model.The PCSNet model proposed in this paper has achieved the state-of-the-art segmentation performance in the test set,and the m Io U has reached 78.93%.The use of quantitative measurement at pixel level further proves the effectiveness and robustness of PCSNet in pavement crack detection.In addition,compared with traditional image processing techniques,the deep learning model is robust and generalized in semantic segmentation. |
PDF Full Text Request