Road Segmentation Of Visible Light Remote Sensing Image

Automatic recognition and segmentation of roads in remote sensing images have very important application values in different application fields such as GIS data update,land management,urban planning and military strikes.With the continuous progress of scientific and technological means and the flourishing development of machine learning and artificial intelligence technologies,how to use deep learning algorithms to solve the difficulties of road segmentation in remote sensing images,improve the accuracy of road segmentation in remote sensing images and simplify the difficulty in the training process of neural networks has become a key research direction for scholars from various countries.Based on the deep learning theory and the characteristics of visible remote sensing images,this research is inspired by the structure of Deep Residual and U-Net neural network,combined with the means of data augmentation and redefined loss function,and proposes a method for road segmentation of remote sensing images.A visualization system for automatic road segmentation of remote sensing images is also developed.The details of the research are as follows:(1)Data augmentation method for remote sensing road images.Since there are few satellite images resources available for neural network training,this paper performs data augmentation of images in the dataset by image morphological transformation(random flip,zoom,etc.)and image color transformation in HSV space to ensure that the neural network has sufficient dataset for training.The good results of the experiments show that the data augmentation means is an important tool to drive the experiments effectively when the dataset resources are not sufficient.(2)A U-net network model based on residual structure is proposed.The neural network structure adopted in this paper uses residual units instead of ordinary neural units as the basic blocks to build a deep residual U-Net network structure,which combines the advantages of U-Net network structure and residual network structure,reduces the parameter design of the neural network and simplifies the training process of the neural network,meanwhile,the jump connections within the residual units,and the jump connections between the high and low layers of the network can be used to facilitate the propagation of information without destroying the semantic information.The experimental results obtain a 3% performance improvement over U-net.(3)A weighted cross-entropy loss function based on the road structure is proposed.The ordinary loss function penalizes each pixel with the same intensity for the occurrence of error.However,combined with the road structure consideration,we feel that the impact of deviations occurring in pixels close to the road network should be greater than that of deviations occurring in pixels far from the road network.Different weights are generated by calculating the minimum Euclidean distance from each pixel to the road area,and the calculated weights are applied to the cross-entropy loss function,which can effectively improve the extracted road structure information.Experimental proofs show that the performance of the model trained with this loss function supervision is improved by at least 3%.(4)An application system for road segmentation of remote sensing images is designed and completed.Based on C/S architecture,users can import images in batch for recognition.