Research On Crack Detection Of Concrete Pavement Based On Attention Mechanism And Multi-Features Fusion

With the rapid development of road construction in my country,road maintenance and repair work has also become very onerous.During the use of concrete,its performance will be affected by the surrounding environment and driving load,etc.,and gradually appear different degrees of damage.Among them,cracks are a relatively common and more harmful road disease,and timely and effective detection of these cracks is of great significance to the later maintenance and maintenance of the road.The crack detection technology based on digital image processing overcomes the shortcomings and deficiencies of traditional manual detection and instrument detection,and provides an important reference for the daily maintenance of roads.However,most of the traditional concrete pavement crack detection technologies based on digital image processing can only be effective in specific scenes.Once the detection scene changes,its detection accuracy will drop more.In recent years,deep learning technology has achieved remarkable results in the field of image segmentation,and has been widely used in traffic and industrial production scenarios.At the same time,it also provides a basis for automated concrete pavement crack detection.Aiming at the shortcomings of traditional pavement crack detection algorithms,the main research work of this thesis is as follows:1.Aiming at the problems of inaccuracy and weak robustness of concrete pavement crack detection by traditional digital image processing technology,the current situation of existing crack detection algorithms at home and abroad is summarized.According to the existing crack detection methods and the characteristics of crack images.A detection model based on the improved residual network Res2Net101 as the backbone network was proposed,which can improve the feature extraction ability of the network.In the decoding stage,the prediction results of different convolutions are feature fused,and the detection of cracks in different scales is realized while ensuring that high-level semantic information is not lost,and the accuracy of crack detection is improved.2.Aiming at the situation that small cracks are easy to be missed,a cascade and parallel mode dilated convolution is used in the central part of the network.Using dilated convolution can not only increase the receptive field of feature points,but also keep the resolution of the feature map from decreasing.Combining the dilated convolutions with different dilated rates will have different receptive fields,and obtained multi-scale information,which can be adapted to crack detection of different scales.3.In this thesis,the attention mechanism is introduced into the network model,so that the neural network learns to pay attention,and at the same time better analyzes the context information of the model without increasing the amount of model parameters and calculations.This makes the feature representation of each convolution stage more compact,improves the anti-noise ability of the network,reduces the false detection rate,and improves the accuracy and robustness of crack detection.4.A concrete pavement crack detection system is designed and implemented in this thesis,processes the input crack image,and finally outputs the crack prediction map to achieve convenient and simple use goals and meet the requirements of fast and accurate crack detection.Finally,the research work is summarized and the future prospect is pointed out,and the future work is prospected.