.png){kind=icon}
.png){kind=icon}
Return
Semantic segmentation for remote sensing images using multi-layer semantic Markov random field with semantic text information
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2025.105174.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Semantic segmentation is an important task in remote sensing image processing. It assigns semantic labels to each pixel of the image. However, as the resolution of the remote sensing image increases, the boundary of the segmentation result appears to be blurred, and the region appears to be over-smoothing. Deep learning algorithms are unable to finely model textures due to operations such as convolution, pooling, and the introduction of a prior knowledge would increase the complexity of the network structure. The traditional Markov random field method can achieve effective fitting of the spatial structure through probabilistic statistical ideas, but it is highly dependent on the initialization results and does not make full use of the multi-modal information. In this paper, we propose a multi-layer structured object-based Gaussian-Markov random field model that introduces semantic text information. Firstly a multi-layer semantic structure is constructed and the deep learning-based methods are used to generate the initial results of each semantic layer; then the object-based Gaussian-Markov random field was constructed at each semantic layer. Specifically, object-based linear regression equations are constructed using parameters containing semantic text information in the feature field of each layer, and a two- point potential energy function combining statistical information and semantic text information is used in the label field. Given the above process, the multi-layer semantic inference process of the algorithm proposed in this paper is more interpretable. The proposed method is tested on the GID dataset, the LoveDA dataset, and some images from GeoEye. The experimental results show that the introduction of semantic text information and the multi-layer semantic structure can significantly improve the segmentation accuracy compared to other state-ofthe-art methods.
Keywords:
Remote sensing image
Semantic inference
Markov random field
Information transmission
Text information