Remote sensing image semantic segmentation method based on region description self-attention mechanism

Abstract

The invention discloses a remote sensing image semantic segmentation method based on a region description self-attention mechanism. The method comprises the steps that a visible light remote sensing image is input into an encoder, advanced semantic features of the visible light remote sensing image are extracted, feature maps of different levels are obtained, global scene extraction and essentialfeature extraction based on a self-attention module are conducted on the basis of the feature maps of different levels, and a scene guiding feature map and a noiseless feature map are correspondinglyobtained; and inputting the scene guide feature map and the noiseless feature map into a decoder, performing up-sampling to return to the size of an original image, and performing pixel-by-pixel classification to obtain a remote sensing image semantic segmentation result. Through the encoder for extracting the semantic features, the self-attention module for increasing the internal relation of theimage and the decoder for mapping the attention-weighted semantic features back to the original space so as to perform pixel-by-pixel classification, the receptive field of the model is improved, themodel can adapt to the scale change of data, and the problem of category imbalance can be solved.

Description

technical field [0001] The invention relates to the technical fields of remote sensing and computer vision, and more specifically relates to the semantic segmentation of remote sensing images based on a self-attention mechanism for region description. Background technique [0002] Applying semantic segmentation technology to the field of remote sensing can achieve the effect of inputting a remote sensing image and outputting the category label of each pixel in the remote sensing image, which is of great help to the understanding of remote sensing images. For example, in terms of land planning, if the land cover types (cities, roads, forests, cultivated land, rivers, etc.) of each pixel on satellite images can be identified, their distribution and occupied area can be clearly understood, which is beneficial to Carry out overall planning; another example is intelligent identification of buildings, not only can quickly find out whether there are illegal buildings, but also can ...

AI Technical Summary

Problems solved by technology

[0005] However, the traditional point-to-point self-attention mechanism increases context information by calculating the relationship between each pixel and other pixels, but it has a disadvantage that the calculation of similarity between points will pay more

Images