CN-121999382-A - Remote sensing image scene classification method and system based on curved surface-frequency domain information graph

CN121999382ACN 121999382 ACN121999382 ACN 121999382ACN-121999382-A

Abstract

The application discloses a remote sensing image scene classification method and a remote sensing image scene classification system based on a curved surface-frequency domain information graph. The method comprises the steps of obtaining a remote sensing image feature map, carrying out feature enhancement, converting the remote sensing image feature map into a high-dimensional manifold curved surface map structure, carrying out feature extraction on the high-dimensional manifold curved surface map structure to obtain map structural features, and classifying remote sensing image scenes according to the map structural features. The application breaks through the limitation that the traditional feature modeling method only carries out local aggregation in Euclidean space, carries out frequency domain spectrum optimization by innovatively introducing Fourier transform, and combines curved surface characterization association to construct a dynamic scene topological graph, thereby effectively extracting the internal topology features which are not influenced by the rotation angle of the scene, ensuring that the model can ignore the direction change and concentrating on the core semantic structure of the scene.

Inventors

LI GUANQUN

Assignees

耕宇牧星(北京)空间科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (9)

1. A remote sensing image scene classification method based on curved surface-frequency domain information graph is characterized in that, Acquiring a remote sensing image feature map, and converting the feature map into a high-dimensional manifold curved surface map structure after feature enhancement; Extracting features of the high-dimensional manifold curved surface graph structure to obtain graph structural features, and classifying remote sensing image scenes according to the graph structural features, wherein the conversion into the high-dimensional manifold curved surface graph structure comprises the following steps: mapping each spatial position in the enhanced feature map to a map node, wherein the initial feature vector of each node is a multidimensional feature vector of the corresponding position in the enhanced feature map; and determining a local neighborhood set for each node according to the manifold geodesic distance, and constructing an inter-node edge set based on the local neighborhood set.
2. The method of claim 1, wherein the feature enhancements include edge and texture feature enhancements and key feature enhancements in the frequency domain; the enhanced feature map is obtained by adding an edge and texture feature enhancement map and a key feature enhancement map.
3. The method of claim 2, wherein edge and texture feature enhancement is performed on the remote sensing image feature map using a gradient histogram HOG operator.
4. The method of claim 2, wherein the key feature enhancement is performed on the remote sensing image feature map using a dual activation mechanism, the dual activation mechanism comprising a parallel gating activation function and a Softmax function.
5. The method of claim 1, wherein calculating manifold geodesic distances including magnitude differences and direction differences in space comprises: in the formula, Representing the manifold geodesic distance between node i and node j, And In order for the coefficient of balance to be present, For the multi-dimensional feature vector of node i, Is a multidimensional feature vector for node j.
6. The method of claim 1, wherein classifying the remote sensing image scene according to the graph structural feature comprises: Carrying out multi-layer feature extraction on the remote sensing image, and respectively obtaining the image structure features corresponding to each layer of feature images; and integrating the structural features of each layer of image from shallow to deep based on the exponential moving average, and classifying the remote sensing image scenes according to the integrated features.
7. The method for classifying remote sensing image scenes according to claim 1, wherein the focus loss is used for training the process of classifying the remote sensing image scenes based on the curved-frequency domain information map.
8. A remote sensing image scene classification system based on a curved surface-frequency domain information graph, which is characterized by being used for realizing the remote sensing image scene classification method based on the curved surface-frequency domain information graph as claimed in any one of claims 1-7, comprising: the cascade feature extraction module is used for carrying out multi-layer feature extraction on the remote sensing image; Each level of feature extraction module is respectively connected with a feature optimization module, each feature optimization module comprises a curve-frequency correlation mapping unit and a multi-layer perceptron, each curve-frequency correlation mapping unit is used for carrying out feature enhancement on a remote sensing image feature map and converting the enhanced features into a high-dimensional manifold curve map structure, feature extraction is carried out on the high-dimensional manifold curve map structure to obtain map structural features, and each multi-layer perceptron is used for classifying remote sensing image scenes according to the map structural features.
9. The remote sensing image scene classification system of claim 8, wherein each multi-layer perceptron is connected in sequence for integrating each layer of image structural features from shallow to deep based on an exponential moving average.

Description

Remote sensing image scene classification method and system based on curved surface-frequency domain information graph Technical Field The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image scene classification method and system based on a curved surface-frequency domain information graph. Background The remote sensing image has wide application prospect in the fields of national defense safety, environment monitoring, urban planning and the like, and the scene classification technology is used as a core means for understanding the semantic content of the high-resolution remote sensing image, and is always the research focus of the remote sensing field. The current mainstream remote sensing image scene classification method is mainly based on Convolutional Neural Network (CNN) extraction features, and global feature representation of images is obtained by stacking convolutional layers and pooling layers, so that scene classification judgment is further carried out. However, most of the existing scene classification methods are designed based on priori knowledge of a natural image dataset (such as ImageNet), feature disturbance caused by the random scene direction in a remote sensing image and frequency domain noise introduced by the complexity of ground feature textures are ignored, and therefore a significant bottleneck exists in the process of extracting a global representation of a scene with rotation invariance of a model. In the deep learning model, standard convolution operations and common global pooling are based primarily on regular euclidean grid structures. This structure, while having some robustness to translation, lacks extreme rotational invariance. In remote sensing images, typical scenes (e.g. "ports", "industrial areas", "overpasses") generally exhibit an arbitrary angular arrangement. When a scene rotates, its spatial distribution on the feature map varies strongly and non-linearly. Due to the lack of the internal modeling capability of the rotation geometric transformation, when the traditional convolution network extracts scene features, different rotation forms of the same scene are often identified as feature vectors with huge differences, so that the problem of large intra-class differences is caused, and meanwhile, complex background textures and high-frequency noise (such as sea surface waves and cloud fog shielding) also easily interfere with the capture of a model on a core ground object structure, so that the problem of feature confusion with high inter-class similarity is caused. In particular, feature aggregation in euclidean space is difficult to overcome characterization dislocation caused by geometric rotation, and stable and topologically consistent scene descriptors cannot be constructed. Therefore, how to purify noise in the scene feature extraction process, and overcome the limitation of the grid structure, is a problem that needs to be solved by those skilled in the art. Disclosure of Invention Aiming at the problems of rotation sensitivity and background noise interference in the classification of the remote sensing image scenes, the application provides a remote sensing image scene classification method and a remote sensing image scene classification system based on a curved surface-frequency domain information graph, which overcome or at least partially solve the problems. In order to achieve the above purpose, the present invention adopts the following technical scheme: In a first aspect, an embodiment of the present invention provides a remote sensing image scene classification method based on a curved surface-frequency domain information graph, including the steps of: Acquiring a remote sensing image feature map, and converting the feature map into a high-dimensional manifold curved surface map structure after feature enhancement; Extracting features of the high-dimensional manifold curved surface graph structure to obtain graph structural features, and classifying remote sensing image scenes according to the graph structural features, wherein the conversion into the high-dimensional manifold curved surface graph structure comprises the following steps: Mapping each spatial position in the enhanced feature map to a map node, wherein the initial feature vector of each node is a multidimensional feature vector of the corresponding position in the enhanced feature map; and determining a local neighborhood set for each node according to the manifold geodesic distance, and constructing an inter-node edge set based on the local neighborhood set. Preferably, the feature enhancement includes edge and texture feature enhancement and key feature enhancement in the frequency domain; the enhanced feature map is obtained by adding an edge and texture feature enhancement map and a key feature enhancement map. Preferably, a gradient Histogram (HOG) operator is adopted to enhance the edge and texture characteristi