A segnet remote sensing image semantic segmentation method combined with random walk

Abstract

The invention relates to a SegNet remote sensing image semantic segmentation method combined with random walk, which is divided into a SegNet initial segmentation step and a random walk optimization segmentation step, the SegNet initial segmentation step is to output the initial semantic segmentation image and category strength information through SegNet; random In the walk optimization segmentation step, first select the random walk seed area, calculate the classification significance index of different categories according to the classification intensity information output by SegNet, and set the threshold to select different types of seed areas; secondly, according to the original image gradient and The classification strength information of SegNet is used to calculate the undirected edge weight; the third step is to start from the seed area and combine the undirected edge weight to perform a random walk on the entire initial segmented image, and finally obtain the optimization on the entire image Split results. The invention walks randomly on the whole image, realizes prediction error and control, greatly reduces edge burrs and patchy classification errors, and completes high-precision semantic segmentation of remote sensing images.

Description

technical field [0001] The invention relates to a SegNet (Random-Walk-SegNet) remote sensing image semantic segmentation method combined with random walk, which belongs to the field of information technology. Background technique [0002] In recent years, remote sensing technology has developed rapidly, and remote sensing image processing technology is increasingly used in disaster analysis, urban monitoring, and resource management. Remote sensing image change detection is one of the key technologies. It can detect what changes have occurred in a specific area within a certain period of time and the degree of change based on images of different periods. Semantic segmentation is a core issue in remote sensing image change detection. , through semantic segmentation, the object category information to which each pixel in the image belongs can be obtained, and on this basis, the change information between the two images can be obtained by comparison. [0003] Image semantic se...

AI Technical Summary

Problems solved by technology

Although the various improved methods mentioned above have improved the accuracy of semantic segmentation as a whole, there are still large errors in the recognition and positioning of the edge of the ground object, the edge segmentation is not smooth enough, and there are many patch noises

For other high-precision segmentation algorithms, the number of neural network layers is large, the error control model is integrated into the training network, and the model structure is complex, such as Gated Convolutional Neural Network for Semantic Segmentation in High-ResolutionImages[J](Wang H, Wang Y, Zhang Q, et al. Remote Sensing, 2017, 9(5): 446.) proposed the information entropy control model ECM based on ResNet-101, which can effectively control the segmentation error, but there will be comparisons in scenarios where the amount of training is insufficient. Obvious errors, which cannot be avoided by using error control algorithms integrated in the network model

[0007] So far, no deep learning-based semantic segmentation algorithm combines SegNet with traditional random walk algorithms

Images