Research On Disturbance Recognition Technology Of Rock Collapse Image Based On Mask R-CNN

Road slope landslides and rock collapses have caused tremendous damage to people’s lives and property.Real-time monitoring of the instability of rocks on slopes is essential.The image processing method is widely used as one of the monitoring methods of rocks on slopes,which can monitor the state of rocks on slopes in real time and efficiently.The monitored scene contains rocks、slopes、people、cars、trees and other objects,where the movement of people、cars and trees will affect the judgment of the state of the rocks on slopes.Traditional image processing methods usually use the motion of objects in the image as the basis for judgment.When objects such as people,cars,and trees move,the rocks on slopes becomes unstable,leading to misjudgment.Therefore,traditional image processing methods cannot accurately determine the unstable state of the rock on the monitoring slope.In view of the above problems,this thesis applies artificial intelligence algorithms to rocks on slopes monitoring scenes to identify the moving targets that distinguish the monitoring images.The Mask R-CNN is used to identify the common easy animals in slope images,Therefore,the probability of misjudgment is reduced,and the accuracy of monitoring the instability of the rocks on slopes is improved.The main work carried out in this thesis is as follows:(1)The inter-frame difference method,background subtraction method and optical flow method are commonly used in traditional monocular image detection methods.They are implemented in engineering.Based on the analysis of the experiment results,it is concluded that the traditional method is insufficient to deal with the complex and variable slope highway environment.Design an algorithm system architecture that combines traditional methods and deep learning techniques to enhance the performance of image warning algorithms.(2)Mask R-CNN is used to identify moving objects in the slope images,thereby reducing the misjudgment of moving interferences on the image monitoring method.This thesis establish the original dataset of tree,people and car samples,combined with deep learning method to automatically mark the easily-moving-targets in the dataset.The sample enlargement method is used to increase the size of image samples of the dataset.(3)By comparing the performance data of various neural networks,Mask R-CNN is selected as the convolutional neural network suitable for engineering.The data obtained from engineering experiments confirm that Mask R-CNN can fit for engineering.At the same time,a parametric pruning residual neural network is proposed,Res Net101 after parameter pruning is used as the basic network of Mask R-CNN,which improves the efficiency of Mask R-CNN operation.After using the Res Net101 after parameter pruning,Mask R-CNN is basically the same as the original one in terms of detection accuracy,and the efficiency of single-frame image processing per second increased by 12.307%.Through engineering experiment verification,this study can show the accurate identification of the easy animals in the rocks on slopes,and improve the accuracy of judgment of the unstable state of the rocks on slopes,which has certain practical significance in engineering.