Road extraction method based on fully convolutional network ensemble learning

Abstract

The invention provides a road extraction method based on fully convolutional network ensemble learning to obtain a road extraction result with coherent roads, complete road network structure and a high recall ratio. The method comprises steps of: firstly, selecting a sample set from an existing remote sensing database and dividing the sample set into a test sample and a training sample; secondly,training a fully convolutional neural network with cross entropy as a loss function, then changing the positive sample penalty weight of the loss function, optimizing the positive sample penalty weight on the basis of the network model parameters of a previous network to obtain the fully convolutional neural network of the loss function with different penalty weights; extracting the roads by usingthe trained fully convolutional neural network to obtain road extraction results of different fully convolutional networks; and finally, according to the principle of spatial consistency, integratingthe different extraction results according to the spatial consistency-based integration strategy, and outputting the final result map. The method can improve the recall ratio of the road extraction result and has good robustness.

Description

Technical field [0001] The invention belongs to the technical field of image processing, and relates to a road extraction method, and further relates to a road extraction method based on full convolutional neural network integrated learning, which can be used to obtain a road network structure of an optical remote sensing image. Background technique [0002] Road extraction from remote sensing images is designed to replace tedious manual tasks and use road extraction methods to obtain the semantic segmentation result map of the road and background input to the remote sensing image. The current road extraction methods can be roughly divided into three categories. The first category is knowledge-based road extraction methods, such as: threshold method, template matching method; the second category is morphology-based methods, such as: edge detection method, watershed algorithm, split growth method; these two types of methods road extraction The effect is not ideal, and the overall...

AI Technical Summary

Problems solved by technology

Although this method makes full use of the excellent characteristics of the fully convolutional neural network, it can extract discriminative features and extract the road results of these features, but when using this model for road extraction, there are certain deficiencies: first, the When the method is used for road extraction, it does not take into account the unbalanced distribution of roads and backgrounds in remote sensing images. The fully convolutional neural network uses the cross-entropy loss function with the same penalty weight for positive and negative samples; secondly, using the fully convolutional neural network When the network extracts roads, it does not take into account the spatial consistency of remote sensing images

However, since the penalty weight of the loss function largely depends on the road characteristics of the training samples, and the spatial consistency of remote sensing images is not considered when extracting roads, the test samples that are quite different from the training samples are The improvement of road extraction recall rate is limited, and the robustness is poor

Images