Railway Extraction From High Resolution Remote Sensing Image Based On Convolutional Network By Key Points Location

With the increasing mileage of railway operation in China,how to effectively monitor the railway and its surrounding environment has become a difficult problem,especially in mountainous and desert areas in Western China.The development of remote sensing image provides a solution to this problem.The acquisition of satellite remote sensing image has the advantages of low cost,not limited by region,large image area and so on.this problem can be solved if the railway and its surrounding areas can be monitored directly in remote sensing images.Highly accurate extraction of railway lines and central lines is the basis to solve this problem.This paper solves the problem of simultaneous extraction of railway line and central line.In recent years,convolutional neural networks have been widely used in image processing.Compared with traditional methods,convolution neural network can automatically extract image features and improve the accuracy of classification after training a large number of samples.So this paper chooses convolution neural network to extract image features and classify railway and background.In order to achieve a highly automated extraction of railway area and its central line simultaneously.The main research work of this paper is as follows:(1)Because there is no dataset for railway key point positioning,in order to realize the training and testing of network model.This paper establishes railway dataset and railway key point positioning dataset based on No.2 high-resolution remote sensing satellite image.In this paper,the key points are marked at the central points on both sides of the railway line,and the midpoints of the two points are on the central line of the railway.So the model can extract the central line of the railway while extracting the railway area.(2)On the basis of the dataset,The convolution neural networks for target detection and key point location is established.The whole network model consists of two modules,one is image detection module,the other is key point location module.The image detection module first detects roughly(128*128 railway area)and then accurately(36*36 railway area).The positioning module locates and corrects points for the 36 *36 railway area detected by the detection module.In the experiment,it is found that there are too many networks,which increases the detection time.In this paper,the structure of the network model is transformed.Using multi-task convolution layer to merge the two networks.At the expense of a little recall rate,This transformation improves the accuracy rate and greatly reduces the network running time.(3)The data used in the experiment are remote sensing images of No.2 high resolution remote sensing satellite,and the performance of the network is tested with actual data.The experimental results show that the key point location method can extract the railway area.The network model established in this paper is available.The railway line and the central line are successfully extracted from The images of No.2 high resolution remote sensing satellite.Moreover,the position of the point labeling is also accurate.