Research On Highway Surface Defect Detection Method Based On Deep Learning

The highway is the link of various social activities,and contributes to the economic development of cities and countries.Regular inspection and maintenance of the highway is of great significance.However,traditional manual defect detection methods are not suitable for highway detection,due to the consuming time,limited accuracy and the difficulty in ensuring the safety of inspectors.With the rise of sensor technology and vision theory,defect detection algorithms based on deep learning have achieved fruitful results in many fields,such as strip steel and wood.However,there are many challenges in the field of highway surface defect detection.The open environment of the highway and the dynamic collection of images can cause strong interference,such as oil,leaves and noise in the image.Secondly,the difference between highway defects and other blocky industrial defects is that their structures are mostly irregular and linear,and because the causes of defects are random,their structure is changeable.In addition,the texture and color of partial inlaid patches and other defects are similar to the asphalt background.Therefore,how to effectively capture the attribute characteristics of defects,enhance the regional characteristics of different types of features,and ensure the detection performance of the integrity of the defects is always a challenge in the field of automatic highway surface defect detection.From the perspective of further improving the accuracy and robustness of highway surface defect detection,the highway surface defect detection method based on deep learning is studied in this paper,and the pavement defects studied mainly include inlaid patches,cracks,potholes and so on.Firstly,the highway surface defect datasets are constructed.Then,a semantics-guided and dual memory selection network for anomaly detection and a rectangular convolution pyramid and edge enhancement network for salient object detection are respectively proposed for image-level and pixel-level defect detection tasks.The main research contents of the thesis are as follows:(1)Constructing datasetsIn order to improve the evaluation dataset in the field of pavement defect detection,this paper constructs highway surface inlaid patch anomaly detection and segmentation datasets,and completes the image-level and pixel-level defect labeling.The constructed datasets contain samples with complex backgrounds or diverse topological structures,which can better reflect the actual road conditions and pose certain challenges to the automatic detection algorithm of road defects.(2)Research on anomaly detection algorithm based on semantics-guided and dual memory selection networkIn order to improve the efficiency of traditional detection methods(both defect images and non-defect images are detected at pixel level),an anomaly detection algorithm based on semantics-guided and dual memory selection network is proposed to filter images without defects.Firstly,global high-level semantic information is used to guide low-level features to learn information that are useful for anomaly detection tasks,and then merged with high-level features to make up for the lack of detailed information loss.Secondly,the initial latent vector is utilized as a query to adaptively search for similar units from two memory matrices for reconstruction,which can encourage the model to learn distinguishing features and limit the model’s ability to reconstruct defective samples.Experimental results show that the proposed anomaly detection algorithm can effectively select defective samples,surpassing other similar models.(3)Research on salient object detection algorithm based on rectangular convolution pyramid and edge enhancement networkAiming at the problems of high false detection rate of road defects in complex background and the fuzzy edge segmentation in low contrast,a salient object detection algorithm based on rectangular convolution pyramid and edge enhancement network is proposed.Rectangular convolution groups are used to extract the global and multi-scale comparison information from multiple convolutional layers,effectively enhancing the generalization ability of the model without adding additional calculations.Multiple convolution kernels larger than the upsampling rate are utilized to gradually fuse feature maps of different scales,which can bridge the gap between features of different levels,and expand the reception range of the network.A large number of experimental results prove that the salient object detection algorithm proposed in this paper can sharpen the edges of defects while ensuring the integrity of defects.