Remote sensing image semi-automatic labeling method and device based on deep learning

Abstract

The invention relates to a remote sensing image semi-automatic labeling method and device based on deep learning. The method comprises the following steps: pre-training a full convolutional neural network by using a cross entropy loss function based on a public remote sensing data set; predicting a to-be-labeled remote sensing image by adopting a pre-trained full convolutional neural network, and outputting a category attribute probability; according to the category attribute probability, calculating an uncertainty metric value of a remote sensing image pixel, and setting a threshold value to extract an uncertain pixel; according to the minimum percentage of uncertain pixels in the superpixels segmented by the remote sensing image, screening the superpixels as recommended labeling areas, and carrying out manual labeling on the recommended labeling areas; and combining the manual labelling of the recommended labelling area and the full convolutional neural network prediction result of the remaining area to obtain a final labelling result. A manual labelling person can be free from burdens of heavy manual drawing of accurate boundaries, the manual annotation efficiency is improved, and the labelling workload and subjective assume of manual labelling are reduced.

Description

technical field [0001] The invention belongs to a remote sensing image tagging method in the field of remote sensing images, relates to efficient semi-automatic tagging and recommended tagging strategies, and is mainly used in remote sensing image information extraction, pixel-level tagging data acquisition and other applications. Background technique [0002] In recent years, the technology of automated remote sensing image information extraction has made great progress. Due to the diversity of remote sensing images, a single machine learning model cannot cover them all. That is, a deep learning model that performs well on some datasets may perform poorly on other datasets. For example, farmland has completely different image representations in different regions, different seasons, and different landforms, so it still faces great challenges. [0003] The information extraction of remote sensing images mainly relies on the color and shape of image targets. In the face of la...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the problem of high cost of image annotation acquisition in the process of remote sensing image information extraction, the present invention, based on the idea of ​​uncertainty sampling, proposes a semi-automatic remote sensing image annotation method that uses a combination of deep neural network probability density function and superpixel segmentation

Images