A remote sensing image semi-automatic labeling method based on deep learning

Abstract

The invention provides a semi-automatic remote sensing image annotation method based on deep learning, and the method comprises the steps of 1) carrying out the manual annotation of a target in a region of interest of a remote sensing image, adding the annotated image into a sample library, and training a prediction model; 2) carrying out target labeling on the original image of the selected areaby utilizing the prediction model to obtain a target area of the prediction image, carrying out regularization processing on the target area, adding the regularized prediction image into a sample library, and training the prediction model; 3) continuously repeating the step 2) to optimize the target annotation and the prediction model training at the same time to obtain an optimal prediction model, and 4) using the optimal prediction model to perform target annotation on the remote sensing image. According to the method, the annotation and the learning are carried out at the same time, the fewer manual intervention is needed along with increase of learning characteristics, and the annotation efficiency is greatly improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence image recognition, in particular to a semi-automatic tagging method for remote sensing images based on deep learning. Background technique [0002] Image annotation is the most cumbersome part of artificial intelligence image processing, and it requires a lot of manpower, material and financial resources for pure manual annotation. The images processed by common annotation software are relatively small, and none of them are aimed at remote sensing image processing. The amount of remote sensing image data is relatively large, and the commonly used image labeling method needs to divide the large image and label it one by one, which is extremely inefficient. It is cumbersome to use remote sensing professional software for vector labeling. It needs to be vectorized first, then assign values ​​to the label elements, and finally convert the vector data to raster data. [0003] Automati...

AI Technical Summary

Problems solved by technology

[0005] To sum up, the conventional prediction method is to label a large number of data sets, spend a long time training the model, and then predict the test image, the image labeling is relatively slow, and the training efficiency is relatively low

Images