Research On Crack Detection Of Concrete Pavement Based On Multi-scale Feature Fusion

With the development of public transportation infrastructure,China has become one of the countries with the longest road mileage in the world,and the road maintenance mileage is increasing.Pavements cracks are the early symptom of road surface diseases.Periodic inspection and timely maintenance of cracks can avoid the deterioration of road diseases and extend the service life of roads.The traditional digital image processingbased pavement crack detection methods realize the automatic inspection,replace manual inspection,and improve the efficiency of road maintenance.However,those methods only use a certain type of crack features for recognition,which tends to be disturbed road noise,and can’t provide the accurate position and shape of cracks for evaluating the degree of crack damage.With the continuous improvement of deep learning and computing power,the deep learning-based pavement crack detection methods have been proposed.Compared with these traditional digital image processing-based methods,deep-learning-based methods can extract more robust crack features and are of strong robustness.Pavement crack detection algorithms based on deep learning are researched on this thesis.An improved multi-scale convolution feature fusion network structure is designed to achieve robust crack extraction in complex road scenes.Attention mechanism is introduced into Pool Net structure to improve the representation ability of crack context information.A data augmentation method is used to amplify training samples and improve generalization ability of model.The weighted cross entropy function is improved to solve the problem of unbalanced number of crack and no-crack samples.The main works of the thesis are listed as following:1.Aiming at the problem that deep convolution feature is continuously “diluted”while deep convolution feature is fused with shallow convolution features layer by layer in the U-shaped multi-scale feature fusion structure,a novel multi-scale feature fusion module is proposed in this thesis,and a network model for pavement crack detection is constructed.In this network,Deep Lab is used as the backbone,which yields four convolution feature maps of different scales.A novel multi-scale feature fusion module is proposed,in which the deepest feature is introduced respectively into three shallow features in order to make full use of the deepest feature with rich semantic information.Besides,deep supervised learning is adopted and three side networks are constructed,to alleviate the problem of gradient disappearance or explosion and improve the performance of crack detection.2.The network structure of Pool Net,which has the perfect performance in salient object detection task,is studied in this thesis.Attention mechanism is combined into the feature integration module and the atrous spatial pyramid pooling module of Pool Net,which enhances the representation ability of crack context information.Aiming at the problem of unbalanced number of crack and no-crack samples,caused that crack pixels are far less than no-crack pixels in a road image,a novel weighted binary cross entropy loss function is proposed to enlarge the contribution of crack samples to the total loss.The negative logarithm of the percentage of no-crack pixel number is the weighting coefficient of the crack term,and the negative logarithm of the percentage of crack pixel number is the weighting coefficient of the no-crack term.3.Aiming at the problem that current pavement crack datasets lack of training sample and cause that the trained model is easy to overfit,the characteristics of pavement crack geometry are analyzed in this thesis,and an image data augmentation method is designed,which organically combines affine transformation and non-rigid transformation to amplify training samples and improve the generalization of the model.Besides,the prototype system for pavement image detection is designed and developed in this thesis,which can segment cracks from the input pavement image,and outputs and saves the gray image of cracks,so as to achieve the use goal of fast and simple.