| Pavement crack detection is the main work content of pavement apparent damage testing,and automatic detection driven by pavement crack detection algorithm is the most effective means to reduce the pressure of road maintenance.In computer vision,the purpose of the pavement crack detection algorithm is to accurately and stably extract the cracks in the input pavement crack picture.So far,a considerable part of the work has contributed to pavement crack detection.Compared with traditional image processing methods,deep learning methods can automatically learn features,and the robustness of the model is stronger.The current algorithms related to pavement crack detection are devoted to reusing features and adding complex structures to the model to improve detection accuracy,while ignoring the simplicity of the model and the rationality of feature utilization.On the other hand,the existing methods are almost all supervised training on data sets supported by a large number of labels,the generalization ability of the model is poor,and the stability of the model performance is ignored.The dissertation studies the pavement crack detection problem from two perspectives:rational use of features to improve model performance and domain adaptation segmentation.The rationality and effectiveness of the proposed model and algorithm are verified through sufficient experiments.The main research contents and innovations of this paper are as follows:1.This thesis proposes a pavement crack detection method based on attention and feature fusion.In order to solve the problem of poor crack detection accuracy of traditional image processing methods in uneven lighting conditions,interlaced light and shadow,different road materials and road coverings,a trainable,end-to-end pixel-level deep convolutional network is proposed.Improve the accuracy and robustness of crack segmentation.In this paper,a deep convolutional network is used to extract the complex semantic information in the picture at multiple levels,and the attention module is embedded in the appropriate position in the backbone network,and the weights are recalibrated on the channel and space of the feature map,so that the network can pay more attention to cracks.information.In order to efficiently improve the location accuracy of cracks,a feature fusion structure is designed in this paper,and feature maps of different scales are cascaded through side network fusion to retain accurate location information while absorbing deep semantic information.Finally,multiple atrous convolutions with different atrous rates are used in parallel to fuse the crack context information again to improve the continuity and accuracy of the crack.Finally,the performance improvement and lightweight characteristics of the model are verified through experiments.2.This thesis constructs a domain adaptation method for pavement crack detection based on style transfer and pseudo-labels.In order to solve the problem that the performance of the model drops sharply in different data sets,CycleGAN is used to fuse the style features of the target domain data set and the location structure features of the source domain into an intermediate domain data set,optimize the model trained in the source domain,and then use pseudo-label self-training The way to further promote the model to learn the characteristics of the target domain.Experimental results show that this method can significantly improve the performance degradation of the model in the detection of cracks in the target domain. |
