Novel convolutional neural network-based remote-sensing image road extraction method

Abstract

The present invention provides a novel convolutional neural network-based remote-sensing image road extraction method, and belongs to the field of remote-sensing image treatment. The method comprises the steps of establishing a full convolution neural network, obtaining a penalty weight and constructing a new loss function based on a minimum euclidean distance between a pixel point and a road region, training a full convolution neural network model through the loss function obtained based on the road structure, and extracting a road based on the well trained model. According to the technical scheme of the invention, the geometric structure of a road is embodied in the loss function in the form of the penalty weight. Therefore, not only the global structural characteristics of the road are reserved, but also a corresponding significant coefficient is provided for each pixel point. Based on the method of the invention, an obtained remote-sensing image road detection model is improved in both recall ratio and precision compared with the prior art, so that a complete road structure can be obtained. The road interruption condition is improved and the roof false detection condition is excluded. Therefore, a better road extraction effect is obtained.

Description

technical field [0001] The invention belongs to the field of remote sensing image processing, in particular to a method for realizing road extraction of remote sensing images based on a convolutional neural network model of road structural characteristics. Background technique [0002] Studies have shown that after nearly 20 years of research, road extraction technology from remote sensing images is still immature. Because there are often trees, pedestrians and vehicles in the remote sensing images taken by satellites or UAVs, the occlusion and shadow of the road, and the shape of the road varies widely. These characteristics bring great challenges to remote sensing road extraction. Traditional heuristic road extraction methods from remote sensing images, such as dynamic threshold, morphological processing, template matching, etc., are not ideal. Learning-based remote sensing image road extraction methods, such as support vector machines, clustering, Markov random fields, e...

AI Technical Summary

Problems solved by technology

[0004] Reference document [1] (S.Saito, T.Yamashita, and Y.Aoki, "Multiple object extraction from aerial imagery with convolutional neural networks," Electronic Imaging, vol.2016, no.10, pp.1–9, 2016.) Convolutional neural network is also used to extract roads from remote sensing images, but the output in the form of vectors will lead to the loss of two-dimensional correlation information of roads in remote sensing images, resulting in structural breaks and misjudgment of roofs as roads. The test results

Images