Remote sensing image segmentation method based on multitask semi-convolution

Abstract

The invention discloses a remote sensing image segmentation method based on multi-task semi-convolution. The method comprises the following steps: step 1, carrying out the preprocessing of an originalremote sensing image IO, and obtaining a remote sensing image I1 after the interference factors in the image are removed; step 2, constructing a multi-task segmentation network, performing boundary prediction and segmentation prediction tasks on the remote sensing image and adjusting the structure of the multi-task segmentation network to adapt to a specific application scene; and step 3, addingthe semi-convolution into the multi-task segmentation network so as to further improve the effect of the multi-task segmentation network. According to the method, the purpose of boundary refinement isachieved through targeted extraction of boundary information by multi-task reuse features and semi-convolution. Benefited from the boundary refinement of the method, the remote sensing image segmentation method provided by the invention remarkably improves the overall segmentation accuracy, improves the segmentation accuracy by 0.9% in comparison with the existing optimal method in a public dataset test and reduces 7.9% of detail errors in the optimal method.

Description

technical field [0001] The invention belongs to the fields of deep learning, computer-aided design and remote sensing image processing, relates to using a deep neural network to understand image semantics, classifying and segmenting images, and proposes a multi-task-based semi-convolution remote sensing image semantic segmentation method. Background technique [0002] Satellites such as NOAA, MODIS, Landsat TM, etc. for Earth observation produce an increasing number of images of different types, resolutions, spectral resolutions and temporal resolutions. These remote sensing images are widely used in land surveying and mapping, agricultural research, environmental research, urban area division and other fields, and also have many applications in the analysis and processing of the relationship between natural resources and human activities. Therefore, it is necessary to segment each remote sensing image. Remote sensing image segmentation is the process of processing remote s...

AI Technical Summary

Problems solved by technology

[0004] Due to the high complexity of the remote sensing image background, the semantic segmentation method directly applied to the remote sensing image can roughly segment the semantic modules, but there are blurred boundaries and classification errors due to the lack of spatial position relationship reasoning.

Images