Slope Surface Defect Detection Based On Convolutional Neural Networks And Transfer Learning

It is very useful to construct a certain part of highway engineering,which is slope.Stable and firm slope has a lot of influence on driving safely on highway.How to accurately identify the disease on the outer surface of the slope,so as to carry out timely repairs,avoid causing more serious slope disasters,and ensure the smoothness and safety of the road.So far,the inspection of the outer surface of the slope still relies on manual walk inspections,This is a sufficient workload,consumes manpower,time and money,and is highly subjective,and it is prone to errors and missed inspections.Up to now,many experts have carried out sufficient research on the image recognition of slopes.Compared with cumbersome manual detection methods,the increasingly mature deep learning recognition technology,the civil engineering field has also fully applied its image recognition technology,and it is outstanding in a series of difficult,large-scale and large-scale image recognition and image detection tasks,and gradually become the best intelligent means to detect the overall health of the slope surface CNN laid a solid foundation for it.For the difficulty of obtaining many samples and the obstacles of external markers of slopes,transfer learning methods are used to eliminate the problem of overfitting.Research content and innovationpaper are as follows:（1）For the two CNNs,AlexNet,VGG-16 and Res Net-18,model training was carried out with the collected slope image data sets,and the influence of different network models on the defect recognition results was discussed.（2）The influence of Res Net-18 model with different transfer strategies on the detection results of slope surface defects was investigated.For the overfitting phenomenon caused by insufficient data sets of the slope surface defects,this paper used the Image Net data set^[1]to Obtain the pre-trained network model,and continuously modify the parameters to meet the disease recognition of the slope surface.（3）Aiming at the recognition and learning ability of the three models of AlexNet,VVG-16,and Res Net-18 under the same migration strategy.Complete the pre-training of the above three models in the Image Net image collection,and input the slope surface image set into the above three pre-training models,and all the structural layers are fine-tuned and trained again,all detection accuracy rates show that the recognition effect of Res Net-18combined with transfer learning is higher than other models.