Search

CN-120765933-B - Remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment

CN120765933BCN 120765933 BCN120765933 BCN 120765933BCN-120765933-B

Abstract

The invention discloses a remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment, and belongs to the technical field of remote sensing image processing. The method comprises the steps of building an image segmentation model comprising a wavelet-Mamba global feature enhancement module, a fast Fourier detail adjustment unit and a decoding and segmentation prediction module, carrying out remote sensing image segmentation training on the built image segmentation model, and carrying out image segmentation on a target remote sensing image by utilizing the trained image segmentation model. According to the invention, through the wavelet-Mamba global feature enhancement module and the fast Fourier detail adjustment unit, the expression capability of ground feature structures, textures and edge information in the remote sensing image can be effectively improved, and the high-precision segmentation of small targets and fuzzy boundaries in a complex scene is realized. The method is particularly suitable for accurately identifying the targets such as buildings, roads, water bodies and the like under the high-resolution remote sensing images, and has higher practical value and popularization prospect.

Inventors

  • LI GUANQUN

Assignees

  • 耕宇牧星(北京)空间科技有限公司

Dates

Publication Date
20260512
Application Date
20250626

Claims (7)

  1. 1. The remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment is characterized by comprising the following steps of: s1, building an image segmentation model comprising a wavelet-Mamba global feature enhancement module, a fast Fourier detail adjustment unit and a decoding and segmentation prediction module; s2, performing remote sensing image segmentation training on the built image segmentation model; S3, performing image segmentation on the target remote sensing image by using the trained image segmentation model; In a wavelet-Mamba global feature enhancement module, carrying out multi-frequency domain decomposition on input features by combining discrete wavelet transformation, respectively modeling the structure, edge and texture information of an image, and introducing a channel level Mamba mechanism to realize global dependency modeling, wherein the specific process comprises the following steps: Normalizing the input features to obtain normalized features Normalizing features using two-dimensional discrete wavelet transforms The method is decomposed into four frequency domain sub-bands, and the expression is as follows: Wherein, the Is a low frequency-low frequency sub-band comprising overall structure and texture information; characterizing horizontal edges for low-to-high frequency subbands; characterizing vertical edges for high-frequency-low frequency subbands; for high frequency-high frequency sub-bands, characterizing diagonal edges, DWT representing a two-dimensional discrete wavelet transform; The low frequency-low frequency sub-band is enhanced by the convolution layer, the activation function, the normalization layer, the Mamba module and the convolution layer, and the low frequency-high frequency sub-band, the high frequency-low frequency sub-band and the high frequency-high frequency sub-band are enhanced by the convolution layer only; fusing the four subbands and reducing to enhanced features by inverse two-dimensional discrete wavelet transform Will enhance the characteristics And normalizing features And fusing in a residual error connection mode to obtain output characteristics.
  2. 2. The remote sensing image segmentation method based on global feature enhancement and fourier detail adjustment according to claim 1, wherein the Mamba module is configured to perform global dependency modeling on a feature map in a channel dimension, and the key operations include: Normalization, which is used for stabilizing training and preventing gradient from disappearing or exploding; Modeling a state space, namely enhancing the channel information flow by using a state space equation; and the linear mapping and gating mechanism enhances the selective control of the features, so that the differences among classes in the remote sensing image are more obvious.
  3. 3. The remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment according to claim 1, wherein in the image segmentation model, a multi-level wavelet-Mamba global feature enhancement module is stacked, and deep semantic features are extracted after multiple downsampling and enhancement
  4. 4. The remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment according to claim 3, wherein in the fast Fourier detail adjustment unit, a frequency domain amplitude-phase dual-channel enhancement strategy is adopted, dynamic convolution and expansion convolution are introduced to strengthen structures and edges respectively, and feature reconstruction is realized through frequency-space interaction modeling, and the specific process comprises the following steps: Will deep semantic features As input, mapping it to the frequency domain by applying a two-dimensional fast fourier transform to obtain an amplitude spectrum and a phase spectrum; the method comprises the steps of enhancing an amplitude spectrum by applying full-dimensional dynamic convolution, an activation function and a convolution layer, and enhancing a phase spectrum by applying the convolution layer, the activation function and an expansion convolution layer; fusing the enhanced amplitude spectrum and the phase spectrum, and applying inverse fast Fourier transform to restore to a space domain to obtain an enhanced detail characteristic diagram: Applying a depth separable convolution layer to the enhanced detail feature map to obtain detail adjustment features: adjusting features and deep semantic features Adding and fusing to obtain fusion characteristics Fusion features Respectively passing through two branches, wherein the upper branch comprises a linear layer, a convolution layer and a Sigmoid activation function to obtain characteristics The lower branch obtains the characteristics through the linear layer and the ReLU activation function Fusing the two branches, and obtaining the output characteristic of frequency domain detail adjustment through a linear layer
  5. 5. The remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment according to claim 4, wherein in the decoding and segmentation prediction module, the output features of the frequency domain detail adjustment And performing up-sampling operation in a series of transposition convolution or bilinear interpolation and convolution modes, restoring to spatial resolution, and performing semantic information restoration through a decoder to generate a final remote sensing image segmentation result.
  6. 6. The remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment according to claim 1, wherein when the remote sensing image segmentation training is performed on the built image segmentation model, a joint loss function is adopted for training and optimizing, the joint loss function comprises cross entropy loss and boundary perception loss, and the expression is as follows: In the formula, Lambda 1 ,λ 2 represents the weight coefficient; Representing cross entropy loss; Representing a boundary perception loss; A true label representing that the (i, j) th pixel is in class c; h, W respectively representing the height and width of the image, and C representing the channel number; representing a binary edge map; representing a prediction boundary map.
  7. 7. An electronic device comprising a processor and a memory, the memory storing machine executable instructions executable by the processor to implement a remote sensing image segmentation method based on global feature enhancement and fourier detail adjustment as recited in any one of claims 1-6.

Description

Remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment Technical Field The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment. Background Along with the rapid development of remote sensing technology and the popularization of high-resolution imaging systems, remote sensing image segmentation is a key step for extracting ground feature information, and plays an important role in the fields of urban planning, land utilization monitoring, disaster assessment, agricultural monitoring and the like. The remote sensing image segmentation aims at dividing pixels in the image into different areas with semantic meaning, so that the automatic identification and extraction of ground features such as buildings, water bodies, roads, farmlands and the like are realized. The traditional image segmentation method has significant challenges in the field due to the characteristics of high resolution, complex content, blurred ground feature boundary, severe scale change and the like of the remote sensing image. At present, the mainstream remote sensing image segmentation method is mostly based on a deep Convolutional Neural Network (CNN) or a Transformer architecture, and semantic modeling and pixel classification are completed by means of strong characterization capability. However, CNN is good at extracting local spatial features, but has the defects in capturing large-scale structures, long-distance dependence, ground feature boundary continuity and the like in remote sensing images, while Transformer has global modeling capability, but has high calculation cost and weak space detail depicting capability in remote sensing image processing, so that the practical application effect in high-resolution remote sensing image segmentation is limited. In addition, the existing method is easy to have the problems of boundary blurring, small target omission, texture information deletion and the like when facing fine-grained ground objects (such as narrow roads, rivers, channels, ridges and the like) in the remote sensing image, and particularly in the areas with complex scenes and small inter-class differences, the segmentation precision is difficult to further improve. Therefore, how to combine global semantic modeling and local detail perception of the remote sensing image becomes a key problem of the current remote sensing image segmentation technology in need of breakthrough. Disclosure of Invention In view of the above, the invention provides a remote sensing image segmentation method based on global feature enhancement and Fourier detail adjustment, which aims to integrate a multi-level information enhancement mechanism of a frequency domain and a space domain, and can effectively improve the expression capability of ground object structures, textures and edge information in a remote sensing image through a wavelet-Mamba global feature enhancement module and a fast Fourier detail adjustment unit, thereby realizing high-precision segmentation of small targets and fuzzy boundaries in a complex scene. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: In a first aspect, an embodiment of the present invention provides a remote sensing image segmentation method based on global feature enhancement and fourier detail adjustment, where the method mainly includes the following steps: s1, building an image segmentation model comprising a wavelet-Mamba global feature enhancement module, a fast Fourier detail adjustment unit and a decoding and segmentation prediction module; s2, performing remote sensing image segmentation training on the built image segmentation model; and S3, performing image segmentation on the target remote sensing image by using the trained image segmentation model. Further, in the wavelet-Mamba global feature enhancement module, multi-frequency domain decomposition is performed on input features in combination with discrete wavelet transformation, the structure, edge and texture information of an image are respectively modeled, and a channel level Mamba mechanism is introduced to realize global dependency modeling, and the specific process comprises the following steps: Normalizing the input features to obtain normalized features Normalizing features using two-dimensional discrete wavelet transformsThe method is decomposed into four frequency domain sub-bands, and the expression is as follows: Wherein, the Is a low frequency-low frequency sub-band comprising overall structure and texture information; characterizing horizontal edges for low-to-high frequency subbands; characterizing vertical edges for high-frequency-low frequency subbands; for high frequency-high frequency sub-bands, characterizing diagonal edges, DWT representing a two-dimensional discrete wavelet t