Research On Crack Detection Of Highway Pavement Based On High-resolution Network Model

Deep convolutional neural networks are widely used in computer vision,especially in face recognition,autonomous driving and medical imaging.As one of the main highway diseases,pavement cracks have a very significant impact on the evaluation of pavement conditions.The traditional manual detection methods are inefficient and low in accuracy,and cannot meet actual needs.Such pain points can be well solved by combining pavement crack detection with deep convolutional neural network.The image classification network can accurately distinguish the types of pavement cracks and interference factors;the image segmentation network can measure the actual area ofpavement cracks and describe the size of pavement cracks.Boundary,to achieve high-precision,high-efficiency detection.At present,there have been many achievements in the research on pavement crack detection at home and abroad,but there are still many problems in the actual operation process,such as difficult data collection,low detection accuracy,poor robustness of network models,etc.These problems make it difficult to achieve the detection effect.It is expected that in order to meet the current requirements for efficient and accurate pavement crack detection,it is necessary to design and optimize a new pavement crack detection network.In this paper,the characteristics of pavement cracks and the current common problems are analyzed and studied,the current network algorithm is optimized,and an improved pavement crack image classification network and pavement crack image segmentation network are designed.Identification and detection,the following are the main research contents of this paper:(1)Backbone network selection.When selecting the backbone network,referring to the narrow and slender characteristics of pavement cracks,a high-resolution network(HRNet)is selected as the backbone network.The high-resolution sub-network and the low-resolution subnetwork adopt a cross-parallel method to ensure that the network can always be in a highresolution state,so as to obtain more characteristics of pavement cracks.As the backbone network,HRNet ensures the accuracy of the classification network and segmentation network,and lays the foundation for the subsequent network model optimization.(2)Pavement crack image classification network.The cracks and interference factors that often appear in the pavement are classified,and a high-resolution image classification network combined with multi-scale feature hybrid hole convolution is proposed based on their characteristics.Among them,the multi-scale feature comes from the SSD network,which is characterized by using the feature map of small receptive field to detect small objects,and the feature map of large receptive field to detect large objects,which can help to detect and identify cracks of different sizes;Under the premise of further improving the network receptive field,it ensures that the local information of crack images is not lost,and solves the problem of difficulty in crack feature extraction.(3)Image segmentation network of pavement cracks.Aiming at the problems of blurred edge segmentation and low accuracy in road crack image segmentation,HRNet is improved.The original bilinear interpolation upsampling of the model is replaced by dense upsampling convolution(DUC),and the downsampling convolution is replaced by passthrough layer to optimize the extraction of crack edge information.At the same time,a feature fusion method combined with the attention mechanism is proposed.By squeezing and stimulating the attention mechanism module(SE-block)combined with a step-by-step upsampling structure,the weight ratio of the feature fusion is demarcated,and the details of the feature map in the decoding process are solved.Missing question.The research results show that the classification and segmentation network based on the high-resolution network has a great improvement in the detection effect of the pavement crack image compared with the traditional convolutional network,solves the existing problems and difficulties,and has certain application value.