Research On Crack Detection Algorithm With Hybrid Dilated And Global Convolution Network

With the continuous development of highway traffic construction,the problem of pavement disease also follows.Crack is a common and harmful disease on the pavement.It is also the source of other diseases.Therefore,timely detection and repair of cracks are important to extend the life of the pavement and ensure the safety of the pavement structure.The crack detection algorithms based on digital image processing are mostly based on hand-made features and have low accuracy and poor robustness.In recent years,deep learning has been widely used in various fields of computer vision,and researchers have been proposed a series of crack detection algorithms based on deep learning.These algorithms can not only realize automatic crack detection,but also improve the accuracy and robustness of crack detection.Thus,crack detection algorithms based on deep learning have very important theoretical significance and practical value.Aiming at the problems of anisotropy of crack structure and complexity of pavement environment,a pavement crack detection algorithm based on hybrid dilated convolution and global convolution is proposed in this thesis.The algorithm divides the crack detection into three parts: feature extraction,feature processing,and feature fusion.The proposed algorithm can better adapt to cracks with different topological structures and complex pavement environments.At the same time,it can improve the accuracy of crack detection.The main research work of this thesis includes:1.Aiming at the problem of local and long-distance feature loss caused by using dilated convolution in the feature extraction stage,a new dilation rate strategy named hybrid dilated convolution is proposed in this thesis.The dilation rate strategy follows the saw-tooth wave-like form: several convolutional layers are divided into one group and the dilation rate of each group increases from small to large,and the next group repeats the same pattern.The convolution with a small dilation rate can extract local feature information,while the convolution with a large dilation rate can extract long-distance feature information.In this way,the top convolutional layer can extract crack features from a larger range of pixels,ensuring that the network can obtain global feature information.2.Aiming at the problem of focusing on localization excessively which leads to low classification accuracy in the feature processing stage,a new feature processing module is proposed in this thesis,which is named global convolution.The global convolution uses a fully convolutional layer with a large convolution kernel to improve classification accuracy while maintaining localization.The fully convolutional layer uses fully convolution without any fully connected layers or global pooling layers to maintain crack localization information.The large convolution kernel can make the feature map and the per-pixel classifier closely connected,which can enhance the ability to deal with the transformation and improve classification accuracy.3.Aiming at the problem of high-level feature with low-resolution and low-level with multi-noise in the feature fusion stage,a new feature fusion strategy named multi-scale feature fusion is proposed in this thesis.The multi-scale feature fusion module firstly uses the convolution with kernel size 1×1 to process the side-output feature maps,then up-samples the feature maps to the size of the input image,finally fuses the feature maps and gets the final prediction.The feature fusion module integrates multi-scale and multi-level feature maps to make the final prediction closer to the ground truth.4.To improve the anti-noise ability of the crack detection network,a spatial-channel attention mechanism module is introduced in the feature extraction stage to enhance crack features,suppress non-crack features,and reduce the influence of noise.In addition,when training the network model,the data augmentation algorithm is used to expand the crack datasets,which can improve the robustness of the network model.Finally,a pavement crack detection system is designed and developed in this thesis,and the research work is summarized.