Semantic Segmentation Method of Remote Sensing Image Based on Multi-scale Decoding Network

Abstract

The invention discloses a remote sensing image semantic segmentation method based on a multi-scale decoding network, which includes the following steps: randomly cutting the high-resolution remote sensing image used for training and its corresponding label image into small images, and the network structure is divided into coding And the two parts of multi-scale decoding, the resolution of the encoded information is doubled through the anti-pooling path and the deconvolution path, and the result of the hole convolution is channel-connected, and the feature image is restored to the original by deconvolution upsampling Size, and then input the output label map into the PPB module for multi-scale aggregation processing, and finally use the cross entropy as the loss function to update the network parameters through stochastic gradient descent; input the small images that are sequentially cut into the test pictures into the neural network to predict its The corresponding label map, and then stitch the label map into the original size. The above technical solution reduces the complexity of the network while improving the segmentation accuracy of the model, and saves training time.

Description

technical field [0001] The invention relates to the technical field of machine vision, in particular to a method for semantic segmentation of remote sensing images based on a multi-scale decoding network. Background technique [0002] Semantic segmentation is an important issue of general concern in areas such as unmanned driving, medical image analysis, and geographic information systems. Semantic segmentation is to allow the computer to segment according to the content of the image. Segmentation is to segment different objects in the image from the pixel level, label each pixel in the original image, and classify it into different labels. The accuracy of segmentation is It contains the understanding of the information in the image. Remote sensing images have the characteristics of complex imaging, high picture pixels, and large amount of information. Therefore, how to use artificial intelligence technology to quickly and accurately extract useful information from remote s...

AI Technical Summary

Problems solved by technology

Compared with FCN and SegNet, U-Net has a more symmetrical encoding and decoding structure, and the skip connection from encoding to decoding part helps the recovery of position information, but also makes the network structure complex and requires more training time

The above network structure often uses pooling to increase the receptive field, but pooling will reduce the spatial resolution when increasing the receptive field.

Although dilated convolutions are used to expand the receptive field and avoid loss of resolution, and the use of convolutions with different dilated rates can capture information of different scales, but dilated convolutions use sparse sampling to lose local information, making long-distance information lack of relevance

In semantic segmentation, a large receptive field can provide more global information, but ignore local information

Images