CN-122024090-A - Construction method, equipment and medium of iterative generalized remote sensing recognition model

CN122024090ACN 122024090 ACN122024090 ACN 122024090ACN-122024090-A

Abstract

The application discloses a construction method, equipment and medium of an iterative generalized remote sensing recognition model, which comprises the steps of extracting features of an initial remote sensing image sample, constructing a semantic topological graph, constructing an initial sample library by combining importance of the sample and similarity among the samples, optimizing correlation between features extracted by a target sensing head and a background sensing head of a base model by adopting an orthogonal constraint loss function and based on the initial sample library, monitoring the new scene remote sensing image sample by utilizing the optimized base model, calculating prediction entropy and feature drift distance of each sample, constructing a domain adaptation data set training adaptation module, projecting parameter variation trained by the adaptation module to an orthogonal compensation space of a first feature subspace determined based on the optimized base model and the initial sample library, and constructing the iterative generalized remote sensing recognition model by parameter fusion. The application realizes the high-efficiency adaptation to the new scene, does not disturb the learned old knowledge, and solves the problem that the old knowledge is forgotten.

Inventors

ZHANG ZHIXIN
LIU ZHE
WEI RENWEI
ZHANG JUN
PI XINYU

Assignees

长江水利委员会网络与信息中心

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. The construction method of the iterative generalized remote sensing recognition model is characterized by comprising the following steps of: extracting features of an initial remote sensing image sample, constructing a semantic topological graph, and constructing an initial sample library by combining the importance of the sample and the similarity among samples; adopting an orthogonal constraint loss function, and optimizing the correlation between the features extracted by the target perception head and the background perception head of the base model based on the initial sample library; Monitoring new scene remote sensing image samples by using the optimized base model, calculating the prediction entropy and the characteristic drift distance of each sample, and establishing a domain adaptation data set training adaptation module; Projecting the parameter variation trained by the adaptation module to an orthogonal complement space of a first characteristic subspace determined based on the optimized base model and the initial sample library, and constructing an iterative generalized remote sensing recognition model through parameter fusion.
2. The method for constructing an iterative generalized remote sensing recognition model according to claim 1, wherein the steps of extracting features from an initial remote sensing image sample, constructing a semantic topological graph, and constructing an initial sample library by combining sample importance and sample-to-sample similarity, further comprise: extracting high-dimensional feature vectors of the initial remote sensing image samples by using a visual basic model, and constructing a neighbor graph by taking cosine similarity among the high-dimensional feature vectors as a measure; Calculating importance scores of the high-dimensional feature vectors corresponding to each sample in the neighbor graphs; And performing iterative sampling based on the importance scores and the similarity among the samples, screening the samples meeting the preset conditions, and adding the samples into an initial sample library.
3. The method for constructing an iterative generalized remote sensing recognition model according to claim 1, wherein the employing an orthogonal constraint loss function and optimizing correlation between features extracted by a target perception head and a background perception head of a base model based on the initial sample library further comprises: Constructing a base model comprising a backbone network, a target perception head and a background perception head; Extracting a target feature vector and a background feature vector through the target perception head and the background perception head respectively; The correlation between the target feature vector and the background feature vector is constrained by using the sum of the absolute values of the inner products of the target feature vector and the background feature vector of each sample as a loss value by adopting an orthogonal constraint loss function.
4. The method for constructing an iterative generalized remote sensing recognition model according to claim 3, wherein the extracting the target feature vector and the background feature vector by the target sensing head and the background sensing head respectively further comprises: carrying out maximum pooling and average pooling on a feature map output by a backbone network through the target perception head, generating a space weight map through a convolution layer after the pooling results are spliced, multiplying the space weight map with the feature map element by element, and obtaining a target feature vector through global average pooling and multi-layer perception mapping; And convoluting the feature images output by the backbone network through the background perception head, pooling the feature images output by the convolution by adopting global standard deviation pooling and global average pooling, and splicing the pooling results to be used as background feature vectors.
5. The method for constructing an iterative generalized remote sensing recognition model according to claim 1, wherein the method for constructing an adaptive domain data set training adaptation module by monitoring new scene remote sensing image samples with the optimized base model and calculating a prediction entropy and a feature drift distance of each sample further comprises: determining model prediction category probability of the new scene remote sensing image samples by using the optimized base model, and calculating prediction entropy of each sample; calculating nearest Euclidean distance between the feature vector of the new scene remote sensing image sample and the feature vector of all samples in the initial sample library by using the optimized base model to serve as a feature drift distance; and screening samples with prediction entropy and characteristic drift distance exceeding preset thresholds to form a domain adaptation data set to train the adaptation module.
6. The method for constructing an iterative generalized remote sensing recognition model according to claim 5, wherein said domain adaptation data set trains an adaptation module, further comprising: Constructing an adaptation module comprising a dimension reduction matrix and a dimension increase matrix based on the domain adaptation data set; calculating a drift covariance matrix of the eigenvectors of all samples in the domain adaptation data set relative to the mean vector of the initial sample library; Singular value decomposition is carried out on the drift covariance matrix, a left singular vector corresponding to a plurality of maximum singular values is selected as an initialization value of a dimension-reduction matrix, and the dimension-increase matrix is initialized to zero; And training the low-rank adaptation module by using the domain adaptation data set, wherein a weighted loss function is adopted in the training process, and the cross entropy loss of each sample is multiplied by the prediction entropy of the sample to be used as a weighted loss value.
7. The method for constructing an iterative generalized remote sensing recognition model according to claim 1, wherein the projecting the parameter variation trained by the adaptation module into an orthogonal complement space based on the optimized base model and the first feature subspace determined by the initial sample library, and constructing the iterative generalized remote sensing recognition model by parameter fusion, further comprises: calculating a feature matrix based on feature representation of all samples in an initial sample library by using the optimized base model to perform singular value decomposition, and selecting left singular vectors corresponding to a plurality of maximum singular values to form a first feature subspace; Calculating parameter variation after training of the adaptation module, and projecting the parameter variation to an orthogonal complement space of a first characteristic subspace to obtain a safe increment parameter orthogonal to the first characteristic subspace; and fusing the safety increment parameters with the original parameters of the base model to construct an iterative generalized remote sensing identification model.
8. The device for constructing the iterative generalized remote sensing recognition model is characterized by comprising the following components: the sample library module is configured to perform feature extraction on an initial remote sensing image sample, construct a semantic topological graph and construct an initial sample library by combining sample importance and sample-to-sample similarity; a correlation module configured to employ an orthogonal constraint loss function and optimize correlation between features extracted by a target perception head and a background perception head of a base model based on the initial sample library; The monitoring training module is configured to monitor the new scene remote sensing image samples by utilizing the optimized base model, and calculate the prediction entropy and the characteristic drift distance of each sample to establish a domain adaptation data set training adaptation module; and the fusion construction module is configured to project the parameter variation trained by the adaptation module to an orthogonal complement space of the first characteristic subspace determined based on the optimized base model and the initial sample library, and construct an iterative generalized remote sensing identification model through parameter fusion.
9. The device for constructing the iterative generalized remote sensing recognition model is characterized by comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit is caused to execute the steps of the method for constructing the iterative generalized remote sensing recognition model according to any one of claims 1-7.
10. A storage medium storing a computer program executable by a construction device of an iterative generalized remote sensing recognition model, the computer program causing the construction device of the iterative generalized remote sensing recognition model to perform the steps of the construction method of the iterative generalized remote sensing recognition model according to any one of claims 1 to 7 when the computer program is run on the construction device of the iterative generalized remote sensing recognition model.

Description

Construction method, equipment and medium of iterative generalized remote sensing recognition model Technical Field The application relates to the technical field of remote sensing models, in particular to a construction method, equipment and medium of an iterative generalized remote sensing identification model. Background The remote sensing image is widely applied to various scene analysis, however, due to the fact that the satellite sensor model, the earth surface type, the atmospheric conditions, the seasonal variation and other complex factors are obviously different, the remote sensing image naturally has multi-source isomerism and obvious inter-domain difference. When the intelligent recognition model is trained in a certain area or under a certain condition and then is transferred to a new area or under a new condition, the performance of the model is often drastically reduced, which is always a key problem for restricting the large-scale application of the remote sensing intelligent recognition technology. The current mainstream remote sensing recognition model construction method comprises the steps of firstly constructing a training sample set in a manual screening or random sampling mode, then training a deep neural network model based on the sample set, and finally deploying and applying the trained model to a new scene. With the development of the general visual basic model, some methods begin to attempt to extract features by using a prediction training model to improve generalization capability, and some methods also adopt full model fine tuning or a simple incremental learning strategy to update the model aiming at a new scene. However, the existing methods still have some drawbacks. The traditional manual screening or random sampling cannot ensure the representativeness, diversity and coverage of training samples, is easy to cause insufficient model generalization, the conventional model training mode fuses and codes target features and background features, the model identification performance is seriously disturbed when the background environment is obviously changed, when the model is faced with a new scene, the whole model fine tuning cost is high, forgetting is easy to generate, simple incremental learning is difficult to quickly adapt to domain drift, and in addition, the existing method lacks an effective mechanism for safely fusing new knowledge into an old model, and continuous iterative evolution of the model is difficult to realize. Disclosure of Invention Aiming at least one defect or improvement requirement of the prior art, the invention provides a construction method, equipment and medium of an iterative generalized remote sensing recognition model, which are used for solving the problem that domain drift caused by multi-source heterogeneous images cannot be solved when the remote sensing intelligent recognition model is constructed in the prior art, so that the generalization capability of the model is declined in the process of cross-scene migration, and the problem that old knowledge is forgotten is difficult to absorb while new domain knowledge is difficult to effectively avoid. In order to achieve the above object, according to a first aspect of the present invention, there is provided a method for constructing an iterative generalized remote sensing recognition model, including: extracting features of an initial remote sensing image sample, constructing a semantic topological graph, and constructing an initial sample library by combining the importance of the sample and the similarity among samples; Adopting an orthogonal constraint loss function, and optimizing the correlation between the characteristics extracted by the target perception head and the background perception head of the base model based on an initial sample library; Monitoring new scene remote sensing image samples by using the optimized base model, calculating the prediction entropy and the characteristic drift distance of each sample, and establishing a domain adaptation data set training adaptation module; Projecting the parameter variation trained by the adaptation module to an orthogonal complement space of a first characteristic subspace determined based on the optimized base model and the initial sample library, and constructing an iterative generalized remote sensing recognition model through parameter fusion. In one possible implementation manner, feature extraction is performed on an initial remote sensing image sample, a semantic topological graph is constructed, and an initial sample library is constructed by combining importance of the sample and similarity among samples, and the method further comprises: extracting high-dimensional feature vectors of the initial remote sensing image samples by using a visual basic model, and constructing a neighbor graph by taking cosine similarity among the high-dimensional feature vectors as a measure; calculating importance scores of the high-dimensional