Road Extraction From Aerial Image Based On Deep Neural Network

The road is an important feature of the remote sensing image.Accurate extraction of roads in remote sensing images can play an important role in urban navigation,urban planning,military mapping,and disaster assessment.The use of pattern recognition,image processing and machine learning algorithms to extract the roads of remote sensing images has become an important research direction in the field of remote sensing.Although the road extraction algorithm based on machine learning has great advantages in performance compared to other methods,such algorithms generally face the problem of unevenness between road and background samples in the sample data.The imbalance between the road and the background sample often leads to the classifier's tendency to predict the sample as a background class that occupies the mode after convergence,which makes the road network in the road extraction result map not perfect,and the road is inspected.The completeness is lower.In order to solve the above problems,a road extraction algorithm based on Fully Convolutional Network ensemble learning and a road extraction algorithm based on weakly supervised Deep Convolutional Network are proposed.Based on the road extraction result map,a road centerline extraction method based on Markov random field is proposed.The centerline extraction algorithm further mines the information in the remote sensing image.The road extraction algorithm was tested on the Massachusat road dataset and the algorithm validity was verified.The main work of this paper is summarized as follows:1.A road extraction algorithm based on integrated learning for Fully Convolutional Networks is proposed.Based on the original Fully Convolutional Networks,the loss function is changed to a weighted cross entropy loss function,and different weights are given to positive samples.The value,when the road pixel is wrongly divided into the background,will get a higher penalty weight.Since the weight of the loss function is difficult to determine adaptively,an integration strategy based on spatial consistency is proposed to integrate the results of the Fully Convolutional Networks with different weight loss functions to improve the quality of road extraction.2.A road extraction algorithm for Deep Convolutional Networks based on weak supervision is proposed.This method only uses the data labels of the scene level as the supervision information of the training samples.Using the remote sensing image and the corresponding scene annotation,a road classification network is trained.When theclassification network converges,the road heat map in each remote sensing image is obtained by using the class activation mapping,and the road heat map is used as the supervision information,and the training is used for road extraction semantic segmentation network.The weakly supervised road extraction network compensates for the incompleteness of the heat map as the supervision information by setting the three principles of seed loss,expansion loss and boundary constraint loss.3.A road center line extraction algorithm based on conditional random field is proposed.The method uses the kernel density estimation of road extraction results to obtain the road probability density.The logarithm of the probability is used as the unary of the conditional random field energy function.At the same time,the relationship between adjacent pixels is taken as the binary potential energy;to ensure the continuity of the center line,the global connected region is used as the high-order potential energy.