Research On Application Of Lightweight Deep Learning Model For Pavement Crack Detection

In recent years,the scale of China’s highway construction has gradually increased,and highway transportation has developed at a high speed.However,the road damage caused by it has seriously affected people’s travel safety.Therefore,the detection and repair of road diseases has been highly regarded by relevant departments.Attention,however,pavement cracks are the earliest direct manifestation of pavement diseases,so detecting pavement cracks is of great significance.The traditional manual method for detecting cracks has high labor density,high economic cost,and low detection accuracy,which can no longer meet the current detection requirements.With the advancement of computer technology,the automatic crack detection technology based on image processing has been deeply studied.However,due to the complex road surface environment,the method of detecting cracks based on digital images contains more noise and the detection effect is not good.Accurate and fast crack image processing methods are particularly important.In order to further improve the existing pavement crack detection method,this paper takes deep learning methods as the main line and studies a variety of deep learning models to segment crack images.First,starting from the cause,types,evaluation criteria and required crack data of pavement cracks,a set of schemes for collecting crack images was designed.A large number of crack images were taken on different road sections,and then data augmentation and sliding cropping were used Increasing the number of crack images,the crack image data set in this paper is constructed.Secondly,the U-Net model and the hierarchical feature learning network model are used for crack image segmentation.The experimental results show that the overall detection results of the two models are good,but the crack images in the complex segmentation environment contain noise,and the parameters of the two models are large The reasoning speed is slow when testing pictures,which is not conducive to the application of actual engineering.Then,this paper improves a neural network architecture automatic search model,and proposes a lightweight mobile terminal neural architecture search full convolution network model to detect cracks.By analyzing the conventional convolution operation and the deep separable convolution group operation used in this paper,it is found that the calculation amount of the model proposed in this paper is reduced by dozens of times compared with other models.In addition,the experimental results show that the model can efficiently test crack images on common computer equipment,and the detection effect is better than other network models.After that,this paper uses a thinning algorithm to extract the crack skeleton in the segmented image,calculates the actual crack length,area,average width and maximum width value and obtains good test results.Finally,in combination with practical application scenarios,this paper uses Py Qt tool to develop an intelligent pavement crack detection software with functions of dividing crack images and videos,calculating actual physical values of cracks,and analyzing historical data of cracks.In summary,the lightweight deep learning model proposed in this paper has a high accuracy and practicability,which provides a new research idea for the development of pavement crack detection technology in the future.