Automatic semantic labeling method of high resolution remote sensing image

Abstract

The invention discloses an automatic semantic labeling method of a high resolution remote sensing image. In the automatic semantic labeling of the high resolution remote sensing image, the automatic semantic labeling method carries out modeling on multi-scale semantic information of the remote sensing image by using a level semantic model, and achieves the automatic labeling of the high resolution remote sensing image by combining a multi-instance learning method. The automatic semantic labeling method of the high resolution remote sensing image is characterized in that the level semantic model is used for achieving a modeling expression of the prior membership function of surface features; a multi-instance multi-label method is introduced to the semantic labeling of the remote sensing image, and the difficulty of labeling work is reduced; the output results of image labeling are provided in the type of the surface feature grade membership function, and the probability confidence coefficient of the labeling results is automatically provided.

Description

technical field [0001] The invention relates to the field of remote sensing image processing, in particular to a method for semantic labeling or semantic classification of high-resolution remote sensing images. Background technique [0002] Automatic semantic annotation of remote sensing images is the basis of remote sensing image organization and indexing. Especially with the explosive increase of remote sensing data in recent years, the amount of high-resolution remote sensing image data is huge, and it is more difficult to directly search and acquire images. High-resolution remote sensing images The meaning of semantic annotation is more obvious. However, the existing manual direct labeling is time-consuming, laborious, and not feasible, and an efficient labeling method from the image semantic level is urgently needed. In the prior art, there are mainly methods for semantic labeling of high-resolution remote sensing images: multi-instance multi-label learning (Multi-inst...

AI Technical Summary

Problems solved by technology

[0004] However, the traditional multi-instance-multi-label learning method ignores the semantic correlation between examples in the bag of words, and often uses a method of uniformly segmenting images when constructing examples, resulting in a balanced multi-instance-multi-label learning method for image annotation. Treating every example, it is impossible to truly realize the semantic layer annotation of images from the perspective of image targets

This way of dealing with the problem is obviously not in line with the characteristics of remote sensing images: the size of remote sensing images is generally large, reflecting the features and landforms in a large area, while remote sensing images are usually only interesting to users during the interpretation process of remote sensing images. Some of the local areas, therefore, directly extracting the global features of the entire image cannot reflect the target concept of interest, so it is necessary to decompose or segment the entire image

Images