Remote sensing target retrieval method based on generalized likelihood ratio metric learning

Abstract

The invention relates to the technical field of remote sensing target retrieval, in particular to a remote sensing target retrieval method, device and equipment based on generalized likelihood ratio metric learning and a computer storage medium. According to the remote sensing target retrieval method, hypothesis distribution is introduced in the training process of a remote sensing target retrieval model, a generalized likelihood ratio corresponding to statistical distribution parameter estimation is defined as a metric space, a loss function is constructed based on the metric space, and the remote sensing target retrieval model is trained; statistical information of global data is introduced in the training stage, overfitting of the neural network is avoided, the feature expression ability of the neural network is improved, and the corresponding target can be retrieved in the remote sensing image based on the content of the given target image with high precision.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing target retrieval, and in particular to a remote sensing target retrieval method, device, equipment and computer storage medium based on generalized likelihood ratio metric learning. Background Art

[0002] With the development of remote sensing, the number of available remote sensing images has grown rapidly. Therefore, the need to automatically retrieve relevant object images from remote sensing databases based on a given query object slice has become increasingly urgent. In recent years, content-based remote sensing image retrieval (CBRSIR) has been developed. However, most of the existing CBRSIR methods only distinguish scenes, which is called scene-level CBRSIR, which is usually insufficient to meet the requirements. For example, many times users want to retrieve not only ships from remote sensing databases, but also specific types of ships. We call this task content-based remote sensing object retrieval (CBRSOR). Some works have focused on the CBRSOR problem, however, the existing CBRSOR methods ignore the following important problem: due to the limitation of GPU memory during training, the existing CBRSOR methods can only sample a batch of data at a time and calculate the loss function based on limited data. There are a large number of simple samples in the training process of the neural network. The sampling-based local learning causes the neural network to overfit the relationship of simple samples. Therefore, the model cannot effectively learn discriminative features, making it difficult to perform well on fine-grained CBRSOR tasks.

[0003] In summary, how to provide a more accurate remote sensing target retrieval method is a problem that needs to be solved at present. Summary of the invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problem of low precision of the remote sensing target retrieval method in the prior art.

[0005] In order to solve the above technical problems, the present invention provides a remote sensing target retrieval method, comprising:

[0006] Obtain a target query image and a database image set;

[0007] The target query image and the database image set are input into the trained remote sensing target retrieval model to obtain the query result.

[0008] Wherein, the training method of the remote sensing target retrieval model includes:

[0009] Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels;

[0010] Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space;

[0011] Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function;

[0012] The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

[0013] Preferably, the step of inputting the target query image and the database image set into a trained remote sensing target retrieval model to obtain the query result comprises:

[0014] Inputting the target query image and the database image set into a trained remote sensing target retrieval model to obtain a similarity estimate between the target query image and each database image;

[0015] sorting the database image set according to the similarity estimate;

[0016] According to the sorting results, the database image corresponding to the highest similarity estimate is determined as the final query result.

[0017] Preferably, the step of constructing a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels, comprises:

[0018] A plurality of query image-database image pairs are obtained, and paired query image-database image pairs are defined as positive sample pairs, and unpaired query image-database image pairs are defined as negative sample pairs.

[0019] Preferably, the extracting of sample features of each sample pair respectively, and calculating the statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space includes:

[0020] Extract sample features of each sample pair respectively;

[0021] Using a maximum likelihood estimation method, based on the sample characteristics, the covariance matrix of the positive sample pair and the negative sample pair is estimated to obtain statistical distribution parameter estimates of the positive sample pair and the negative sample pair;

[0022] The generalized likelihood ratio corresponding to the statistical distribution parameter estimation is defined as a metric space.

[0023] Preferably, constructing a similarity estimation function of sample pairs based on the metric space, and constructing a loss function based on the similarity estimation function comprises:

[0024] Calculating a differential feature of each sample pair based on the sample features;

[0025] Calculating multivariate Gaussian distributions of positive sample pairs and negative sample pairs based on the differential features respectively;

[0026] Constructing a similarity estimation function based on the multivariate Gaussian distribution;

[0027] Defining the differential feature as a zero-mean distribution, and updating the similarity estimation function based on the metric space;

[0028] Construct a loss function based on the updated similarity estimation function.

[0029] Preferably, defining the differential feature as a zero mean distribution and updating the similarity estimation function based on the metric space comprises:

[0030] The product of the differential feature, the transpose of the differential feature, and the difference between the inverse function of the statistical distribution parameter estimate of the positive sample pair and the inverse function of the statistical distribution parameter estimate of the negative sample pair is defined as an updated similarity estimation function.

[0031] Preferably, constructing a loss function based on the updated similarity estimation function includes:

[0032] The cosine similarity function in the weighted sum form of unified loss and identity loss is replaced with the updated similarity estimation function to obtain the loss function.

[0033] The present invention also provides a remote sensing target retrieval device, comprising:

[0034] A data acquisition module, used to acquire a target query image and a database image set;

[0035] The target retrieval module is used to input the target query image and the database image set into the trained remote sensing target retrieval model to obtain the query result.

[0036] Wherein, the training method of the remote sensing target retrieval model includes:

[0037] Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels;

[0038] Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space;

[0039] Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function;

[0040] The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

[0041] The present invention also provides a remote sensing target retrieval device, comprising:

[0042] Memory for storing computer programs;

[0043] A processor is used to implement the above-mentioned remote sensing target retrieval method steps when executing the computer program.

[0044] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned remote sensing target retrieval method are implemented.

[0045] The above technical solution of the present invention has the following advantages compared with the prior art:

[0046] The remote sensing target retrieval method described in the present invention introduces a hypothetical distribution in the training process of the remote sensing target retrieval model, defines the generalized likelihood ratio corresponding to the statistical distribution parameter estimation as a metric space, constructs a loss function based on the metric space, and trains the remote sensing target retrieval model; the present invention introduces statistical information of global data in the training stage, avoids overfitting of the neural network, improves the feature expression ability of the neural network, and enables high-precision retrieval of corresponding targets in remote sensing images based on the content of a given target image. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0048] Figure 1 It is a flow chart of the implementation of a remote sensing target retrieval method provided by the present invention;

[0049] Figure 2 It is an overall flow chart of a remote sensing target retrieval method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The core of the present invention is to provide a remote sensing target retrieval method, device, equipment and computer storage medium, which effectively improves the accuracy of remote sensing target retrieval.

[0051] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] Please refer to Figure 1 and Figure 2 , Figure 1 A flowchart of a remote sensing target retrieval method provided by the present invention; Figure 2 The overall flow chart of a remote sensing target retrieval method provided by an embodiment of the present invention; the specific operation steps are as follows:

[0053] S101: Acquire a target query image and a database image set;

[0054] S102: Input the target query image and the database image set into the trained remote sensing target retrieval model to obtain the query result.

[0055] Wherein, the training method of the remote sensing target retrieval model includes:

[0056] Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels;

[0057] Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space;

[0058] Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function;

[0059] The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

[0060] Based on the above embodiments, this embodiment describes step S102 in detail:

[0061] Inputting the target query image and the database image set into a trained remote sensing target retrieval model to obtain a similarity estimate between the target query image and each database image;

[0062] sorting the database image set according to the similarity estimate;

[0063] According to the sorting results, the database image corresponding to the highest similarity estimate is determined as the final query result.

[0064] Based on the above embodiments, this embodiment describes in detail the training process of the remote sensing target retrieval model:

[0065] The present invention proposes a generalized likelihood ratio test based metric learning (GLRTML) method for content-based remote sensing target retrieval (CBRSOR). GLRTML models the retrieval task as a hypothesis testing problem. By introducing the hypothesis distribution in the training process, the neural network can better learn the global relationship of the data.

[0066] Based on the above embodiment, the construction of a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels, includes:

[0067] Obtain multiple query image-database image pairs, and the GLRTML proposed in the present invention models the CBRSOR problem as a hypothesis testing problem. The case of unpaired query image and gallery image is defined as the null hypothesis H0, that is, the unpaired query image-database image pair is defined as a negative sample pair, and correspondingly, the case of paired query image and gallery image is defined as the alternative hypothesis H1, that is, the paired query image-database image pair is defined as a positive sample pair.

[0068] Based on the above embodiment, constructing a similarity estimation function of sample pairs based on the metric space, and constructing a loss function based on the similarity estimation function includes:

[0069] Calculating a differential feature of each sample pair based on the sample features;

[0070] For any query image I i and database image I j For (i, j), the likelihood ratio between the null hypothesis H0 distribution p(·|H0) and the alternative hypothesis H1 distribution p(·|H1) is used to define the similarity s(i, j) between the query image and the gallery image:

[0071]

[0072] Thanks to the powerful representation ability of neural networks, we can calculate similarity in feature space instead of original image space:

[0073]

[0074] where x θ，i and x θ，jRepresents the features of the image obtained by neural network inference with parameter θ. From the perspective of statistical inference, a high value of s(i, j) means accepting hypothesis H1 with high confidence, that is, believing that I i and I j are similar. Conversely, a low value of s(i, j) means dissimilarity. Considering that s(i, j) should be equivalent to s(j, i), we will θ，i and x θ，j Replaced by the differential feature x θ，ij =x θ，i -x θ，j , which eliminates the bias caused by the position of the elements in the feature and introduces symmetry:

[0075]

[0076] Calculating multivariate Gaussian distributions of positive sample pairs and negative sample pairs based on the differential features;

[0077] The distribution of the assumptions H0 and H1 and the null hypothesis distribution are often difficult to obtain in advance, so they need to be calculated through parameter estimation. and It means p(x θ，ij |H1) and p(x θ，ij |H0), that is, the estimated distribution. Where θ1 and θ0 are the estimated distribution parameters. We assume and are all d-dimensional multivariate Gaussian distributions, where the mean vectors are μ1, μ0, and the covariance matrices are ∑1, ∑0:

[0078]

[0079] · constructing a similarity estimation function based on the multivariate Gaussian distribution;

[0080] Substituting Formula 4 and Formula 5 into Formula 3 to replace p(·|H1) and p(·|H0), we obtain:

[0081]

[0082] Defining the differential feature as a zero-mean distribution, and updating the similarity estimation function based on the metric space;

[0083] C0 is the statistically irrelevant part. Considering the final use for ranking, and paying attention to the symmetry of the probability space (i.e., s(i, j) should be equivalent to s(j, i)), we assume that the differential feature is a zero-mean distribution, we further simplify Formula 6 and obtain the similarity score:

[0084] s(i, j0=x θ，ij T(∑0 -1 -∑1 -1 )x θ，ij . (7)

[0085] Based on the above embodiments, in order to better understand “updating the similarity estimation function based on the metric space”, we supplement the definition of the metric space, that is, extracting the sample features of each sample pair respectively, and calculating the statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as the metric space includes:

[0086] Extract sample features for each sample pair separately;

[0087] Using the maximum likelihood estimation method, based on the sample characteristics, the covariance matrix of the positive sample pair and the negative sample pair is estimated to obtain the statistical distribution parameter estimates of the positive sample pair and the negative sample pair;

[0088] The present invention uses maximum likelihood estimation (MLE) to estimate the covariance matrix from the data. Assume that N1 and N0 hypothesized H1 samples and H0 samples (i.e., positive sample pairs and negative sample pairs) can be constructed in the training set, respectively, and X1 and X0 represent the positive sample pair set and the negative sample pair set, respectively. For ease of expression, we represent the features of the lth positive or negative sample as and According to MLE:

[0089]

[0090] make We can get an estimate of ∑1:

[0091]

[0092] Similarly, we can get:

[0093]

[0094] The generalized likelihood ratio corresponding to the statistical distribution parameter estimation is defined as a metric space.

[0095] Therefore, the similarity estimation function is updated based on the metric space, that is, the estimated similarity score of Formula 7 is rewritten:

[0096]

[0097] The statistical distribution parameter estimates are parameters in the metric space (i.e., generalized likelihood ratio). The metric space can be updated by updating the statistical distribution parameter estimates, thereby obtaining an updated similarity estimation function; that is, Formula 11 defines not only the metric space, but also the similarity estimation function.

[0098] The present invention also cuts out the The spectrum to ensure It is positive.

[0099] Based on the above embodiment, constructing a loss function based on the updated similarity estimation function includes:

[0100] In order to train the neural network end-to-end, we construct the loss function based on Formula 11. The baseline of our loss function is the weighted sum of unified loss and identity loss, and we only replace the cosine similarity with the similarity score defined in Formula 11:

[0101]

[0102] Where M1, M0, and M represent the total number of positive sample pairs, negative sample pairs, and images in the batch, respectively. and y are the estimated similarity scores of the positive and negative pairs, respectively, and m, n, and i index the positive pair, negative pair, and image, respectively. i is x θ，i The category label of is the classifier that recognizes x θ,i For category y i The predicted probability of , v is the temperature coefficient, and α is the weighting coefficient.

[0103] Based on the above embodiment, the remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model, including:

[0104] To calculate the loss function in Formula 12, we must first estimate the covariance matrix according to Formula 9 and Formula 10. After the loss function is calculated and the gradient backpropagation is completed, the distribution of each sample feature changes due to the update of the neural network parameters, and the statistical distribution parameters need to be re-estimated. Therefore, we designed an iterative training strategy to collaboratively update the neural network parameters and statistical distribution parameters during the training process. In addition, we designed a two-stage training strategy, which first uses the label learning ability of identity loss to obtain a pre-trained model, and then uses Formula 12 as the loss function for the second stage of training.

[0105] In the testing phase, the features of the query image and the database image are obtained through the trained neural network, and then the estimated similarity score is calculated according to Formula 11. and is estimated on the training set after the last training epoch. Finally, we sort the estimated similarity scores in descending order to obtain the retrieval results.

[0106] Based on the above embodiments, the GLRTML provided by the present invention pays more attention to difficult samples or rare samples in the data during training, rather than ordinary simple samples.

[0107] prove:

[0108] Consider a training batch containing M1 and M0 positive samples and negative samples Including difficult positive samples And difficult negative examples where m, l, m * , l * is the index variable. Consider and Predict the difficulty and rarity of samples under hypothesis H1 and hypothesis H0:

[0109]

[0110] Therefore, suppose Represents the regular positive sample The estimated similarity of Represents the difficult positive samples The estimated similarity of is:

[0111]

[0112] Similarly, we can get:

[0113]

[0114] In order to facilitate writing, The loss function for the GLRTML part is about and Taking partial derivatives we get:

[0115]

[0116]

[0117] Considering the scaling property of the exponential function:

[0118]

[0119] Therefore, the training process of GLRTML will pay more attention to rare and difficult samples to avoid overfitting.

[0120] Based on the above embodiments, the present invention theoretically analyzes the optimality of GLRTML given the accurate estimation of the false alarm rate and the hypothesis distribution. Experimental results show that GLRTML can significantly improve the performance of the CBRSOR task.

[0121] Experimental conditions description:

[0122] FGSRSI-23 is a fine-grained ship recognition dataset released by Xiong et al., containing 23 fine-grained ship subcategories and a total of 1067 images. The images of FGSRSI-23 are mainly from Google Earth and SuperView-1 (SV-1) satellites. The advantage of FGSRSI-23 is that there is only one ship target in each image, which is suitable for CBRSOR tasks. In order to minimize the randomness of the experimental results, we removed categories with less than 10 samples, and the processed dataset contains 16 categories and 1038 images. According to the experimental settings of CBRSOR, we divide the dataset into a training set for training, and a query set and gallery set for testing. The training set, query set, and gallery set have 516, 202, and 320 ship samples, respectively, and all contain all ship categories. In addition, we also added 2065 background images to the gallery set to simulate a more realistic CBRSOR scene. This reorganized version of the dataset is called CBRSOR-FGSRSI.

[0123] The present invention ensures a fair backbone network and consistent hyperparameters. In terms of the backbone network, the present invention uses the popular convolution-based neural network Resnet-50 as a feature extractor. The hyperparameter settings are reported as follows. The batch size M is set to 64, and the feature dimension d is 64. The epochs of the first training stage are 100, and the epochs of the second training stage are 50. The momentum of SGD is 0.9, and the learning rates of the first and second training stages are 0.05 and 0.005, respectively. In the loss function, the temperature coefficient v is 0.001, the weight coefficient α is 1, and the weight decay coefficient is 0.0005.

[0124] In terms of experimental environment, for fair comparison, all experiments were conducted in Python 3.8.18 and PyTorch 2.1.2, and the computer used Intel(R) Xeon(R) Platinum 8358P CPU and NVIDIA A800-80GB GPU. The operating system was Ubuntu 20.04, and the CUDA toolkit version was 11.4.

[0125] The present invention illustrates the effectiveness of GLRTML through ablation experiments. The experimental results are shown in Table 1:

[0126] Table 1 Ablation experiment results on CBRSOR-FGSRSI dataset

[0127]

[0128] GLRTML-TR and GLRTML-TE represent the use of statistical models in the training phase and the use of statistical models in the testing phase, respectively. If the use of statistical models is cancelled in the testing phase, cosine similarity is used to calculate the similarity between embeddings. By adding GLRTML-TR to the baseline method, significant performance improvements can be obtained even when cosine similarity is used as the metric space in the testing phase. This shows that GLRTML introduces statistical information of global data in the training phase, avoids overfitting of neural networks, and improves the feature expression ability of neural networks. By introducing statistical models in the testing phase, the performance is further significantly improved, which is intuitive because the metric space of the training and testing processes remains consistent under this condition, reflecting the superiority of the metric space derived based on the likelihood ratio.

[0129] An embodiment of the present invention provides a remote sensing target retrieval device; the specific device may include:

[0130] A data acquisition module, used to acquire a target query image and a database image set;

[0131] The target retrieval module is used to input the target query image and the database image set into the trained remote sensing target retrieval model to obtain the query result.

[0132] Wherein, the training method of the remote sensing target retrieval model includes:

[0133] Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels;

[0134] Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space;

[0135] Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function;

[0136] The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

[0137] The remote sensing target retrieval device of this embodiment is used to implement the aforementioned remote sensing target retrieval method. Therefore, the specific implementation method of the remote sensing target retrieval device can be seen in the embodiment part of the aforementioned remote sensing target retrieval method. For example, the data acquisition module and the target retrieval module are respectively used to implement steps S101 and S102 in the aforementioned remote sensing target retrieval method. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part, which will not be repeated here.

[0138] A specific embodiment of the present invention further provides a remote sensing target retrieval device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the above-mentioned remote sensing target retrieval method when executing the computer program.

[0139] A specific embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned remote sensing target retrieval method are implemented.

[0140] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0141] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0142] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0144] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A remote sensing target retrieval method, characterized in that: include: Obtain a target query image and a database image set; The target query image and the database image set are input into the trained remote sensing target retrieval model to obtain the query result. Wherein, the training method of the remote sensing target retrieval model includes: Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels; Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space; Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function; The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

2. The remote sensing target retrieval method according to claim 1, characterized in that: The step of inputting the target query image and the database image set into a trained remote sensing target retrieval model to obtain the query result comprises: Inputting the target query image and the database image set into a trained remote sensing target retrieval model to obtain a similarity estimate between the target query image and each database image; sorting the database image set according to the similarity estimate; According to the sorting results, the database image corresponding to the highest similarity estimate is determined as the final query result.

3. The remote sensing target retrieval method according to claim 1, characterized in that: The construction of the query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels, comprises: A plurality of query image-database image pairs are obtained, and paired query image-database image pairs are defined as positive sample pairs, and unpaired query image-database image pairs are defined as negative sample pairs.

4. The remote sensing target retrieval method according to claim 1, characterized in that: The sample features of each sample pair are extracted respectively, and the statistical distribution parameter estimates of the positive sample pair and the negative sample pair are calculated respectively based on the sample features, and the generalized likelihood ratio corresponding to the statistical distribution parameter estimate is defined as a metric space including: Extract sample features of each sample pair respectively; Using a maximum likelihood estimation method, based on the sample characteristics, the covariance matrix of the positive sample pair and the negative sample pair is estimated to obtain statistical distribution parameter estimates of the positive sample pair and the negative sample pair; The generalized likelihood ratio corresponding to the statistical distribution parameter estimation is defined as a metric space.

5. The remote sensing target retrieval method according to claim 1, characterized in that: The constructing of a similarity estimation function of sample pairs based on the metric space, and constructing a loss function based on the similarity estimation function comprises: Calculating a differential feature of each sample pair based on the sample features; Calculating multivariate Gaussian distributions of positive sample pairs and negative sample pairs based on the differential features respectively; Constructing a similarity estimation function based on the multivariate Gaussian distribution; Defining the differential feature as a zero-mean distribution, and updating the similarity estimation function based on the metric space; Construct a loss function based on the updated similarity estimation function.

6. The remote sensing target retrieval method according to claim 5, characterized in that: Defining the differential feature as a zero mean distribution and updating the similarity estimation function based on the metric space includes: The product of the differential feature, the transpose of the differential feature, and the difference between the inverse function of the statistical distribution parameter estimate of the positive sample pair and the inverse function of the statistical distribution parameter estimate of the negative sample pair is defined as an updated similarity estimation function.

7. The remote sensing target retrieval method according to claim 5, characterized in that: The constructing of the loss function based on the updated similarity estimation function comprises: The cosine similarity function in the weighted sum form of unifiedloss and identityloss is replaced with the updated similarity estimation function to obtain the loss function.

8. A remote sensing target retrieval device, characterized in that: include: A data acquisition module, used to acquire a target query image and a database image set; The target retrieval module is used to input the target query image and the database image set into the trained remote sensing target retrieval model to obtain the query result. Wherein, the training method of the remote sensing target retrieval model includes: Construct a query image-database image pair training set, including positive sample pairs and negative sample pairs with predefined labels; Extracting sample features of each sample pair respectively, and calculating statistical distribution parameter estimates of the positive sample pair and the negative sample pair respectively based on the sample features, and defining the generalized likelihood ratio corresponding to the statistical distribution parameter estimate as a metric space; Constructing a similarity estimation function of the sample pair based on the metric space, and constructing a loss function based on the similarity estimation function; The remote sensing target retrieval model is trained based on the loss function, and the neural network parameters are updated in each training cycle, thereby updating the metric space, and the training is iterated until the training converges to obtain a trained remote sensing target retrieval model.

9. A remote sensing target retrieval device, characterized in that: include: Memory for storing computer programs; A processor, used to implement the steps of a remote sensing target retrieval method as claimed in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a remote sensing target retrieval method as claimed in any one of claims 1 to 7 are implemented.

Images