Image Filtering And Segmentation Of Asphalt Pavement Cracks Based On Convolutional Neural Network

As one of the basic public facilities,the normal operation and maintenance of the highway system plays a vital role in the stability and development of the society.The pavement crack is the most common,easy to form and more harmful in pavement diseases.The timely discovery and treatment of the pavement crack plays a key role in the pavement maintenance work.Now,only using the traditional image processing method for pavement crack recognition,already cannot satisfy the needs of the transportation system.The rapid development of deep learning brings new research ideas and directions for pavement crack recognition,and using the deep learning method can obtain more accurate prediction results and identification accuracy.Focusing on asphalt pavement images,this thesis studies more efficient methods for filtering and segmentation of asphalt pavement crack images based on convolutional neural networks.The main research work of this thesis is as follows:(1)On the basis of analyzing the image characteristic of pavement cracks,the pavement image feature extraction method based on deep convolution feature is studied.Firstly,the self-collected dataset is preprocessed,and all the datasets are partitioned and enhanced.Then,by comparing different deep convolution feature networks,the method of pavement crack feature extraction based on convolutional neural network is studied.The studied feature network is applied to the pavement crack filtering and segmentation,and the comparison experiments were done.(2)A hierarchical crack filtering method for asphalt pavement based on convolutional neural network is proposed.Firstly,the training dataset is used to train three different classification networks to obtain different classification models.After that,the softmax layer is fine-tuned with the fine-tuning dataset to achieve the purpose of all crack images being recalled.Finally,on the basis of comparing the similarities and differences of non-crack images filtered by different convolutional neural networks,a three-level hierarchical pavement crack image filtering model is proposed.The experimental results show that the three-level hierarchical filtering model can achieve high true negative rate and accuracy when all crack images are completely recalled.(3)A pavement crack image segmentation method based on DAU-Net network model is proposed.Firstly,the different feature networks are combined with U-Net segmentation network,and the feature network with the best segmentation effect is selected by training and testing the road crack images.Then,due to the limitations of the existing semantic segmentation network if it is directly used to segment pavement cracks,this thesis introduces DAC module and ACM module between encoding and decoding to improve the U-Net segmentation network.The experimental results show that the proposed DAU-Net segmentation network based on the dilated convolution and attention mechanism has a better mean intersection over union value for segmentation of pavement crack images.