High-robustness deep road extraction method based on label probability sequence

Abstract

A high-robustness deep road extraction method based on a label probability sequence provides a new SDL framework, a road extraction task of a label probability sequence learning DCNN is introduced, the label probability sequence is obtained by front-end learning DCNN, and abundant real label information can be provided to correct error labels and steadily learn DCNN; based on a label probability sequence, adaptive label correction is provided to improve the quality of a noise data set; meanwhile, a noise correction loss function based on label probability sequence uncertainty is developed, so that DCNN is learned steadily by using an improved data set; empirical evaluation is carried out on two challenging noise road data sets, i.e., a uniform road and a confluence road, so as to verify that the method learns a DCNN model steadily.

Description

technical field [0001] The invention relates to the technical field of intelligent processing of remote sensing images, in particular to a highly robust depth road extraction method based on a label probability sequence. Background technique [0002] Extracting roads from remote sensing images has attracted extensive attention from academia and industry due to its important role in many applications such as intelligent transportation systems and urban planning. Supervised deep convolutional neural networks (DCNNs) are widely adopted methods in this task, while their superior performance is highly dependent on a sufficient number of pixel-labeled samples. Recently, adopting crowdsourced geographic data, such as Open Street Map (OSM) and GPS, to train DCNNs has been considered as an effective way to overcome this problem. However, these crowdsourced datasets are usually noisy and include wrong labels. Some road objects are incorrectly labeled with crowdsourced geographic dat...

AI Technical Summary

Problems solved by technology

However, studies have shown that this powerful feature extraction capability can also lead DCNNs to learn or memorize complete random data distributions

According to the conclusion of the study, when the labels in the training data set are corrupted, DCNN is likely to overfit this noisy data set, which leads to significant performance degradation problems

Images