Research On The Detection Of Ground Objects Along The Railway In Remote Sensing Images Based On Deep Learning

The large-scale construction of railways in China not only promotes the rapid development of the national economy but also poses new challenges in terms of safe operation and maintenance.The illegal features along the railway line seriously affect the safe operation of the railway.Timely detection and detection of the features along the railway line has become the primary problem to be solved by the current railway maintenance personnel.The rapidly developing high-resolution remote sensing technology provides a fast,convenient and efficient solution for the timely detection of features along the railway line.Therefore,this thesis studies the feature detection method along railway lines based on the deep learning target detection method and remote sensing image data of a section of railroad lines.The main research contents are as follows.(1)Remote sensing image feature detection method based on improved YOLOv4.Firstly,aiming at the missed detection problem caused by the loss of small target information in remote sensing images along the railway line,the transmission layer of Densely Connected Convolutional Networks(Dense Net)is improved for YOLOv4 to perform feature extraction of ground objects and enhance the detection ability of small ground objects.Then a SE-CSP module is proposed,which improves the detection efficiency by separating the SE attention module in the residual unit.Finally,in view of the difficulty of feature extraction of large-scale railway targets,the attention mechanism module of CBAM(Convolutional Block Attention Module)is improved to enhance the ability to extract large-scale railway targets.The experimental results show that the m AP of the improved method is increased by 2.58%compared with YOLOv4,and the model size is reduced by 8.53% compared with the original YOLOv4.(2)Remote sensing image feature detection method based on lightweight network.The YOLOv4 model is used for feature detection along railroad lines,which has a complex model structure,a large number of parameters,and takes up many resources,making it difficult to meet the efficient and real-time requirements of target detection tasks of mobile devices.Firstly,by combining the lightweight Mobilenetv3 and YOLOv4 models for feature extraction of feature targets,the number of parameters and computation of the feature detection model is reduced to improve the model detection speed.Secondly,an efficient ECA attention module is incorporated at the prediction result side of the network to avoid the impact of the reduced model parameters on the feature detection accuracy,thus improving the feature target detection accuracy.The experimental results show that this method reduces the size of the YOLOv4 model by 81.67%,reduces the amount of computation,and improves detection efficiency.(3)Optimal anchor configuration method based on improved K-means and differential evolution.In the face of large changes in the scale of ground objects in remote sensing images along the railway line,the anchor frame configuration scheme of K-means clustering used in the YOLOv4 model will cause deviations in the anchor frame settings,resulting in deviations in the trained model.This thesis proposes an anchor configuration scheme based on improved K-means and Differential Evolution(DE)to better capture the relationship between target scale and number,and improve the overall performance of the feature detection model by balancing the detection accuracy of all features target classes.The experimental results show that by using the optimal anchor configuration method proposed in this research,the model complexity does not increase,i.e.,no additional consumption is added,but the detection accuracy for each feature class is improved and the detection accuracy of each class of feature targets can be balanced.