Single-domain remote sensing image retrieval method based on weight adaptive combination robust loss

Abstract

The application relates to the technical field of remote sensing image retrieval, and particularly discloses a single-domain remote sensing image retrieval method based on a weight self-adaptive combined robust loss, which adopts the following steps: S1: inputting training data into a main model and extracting image features; the main model outputs the probability that the image belongs to each category according to the image features, and defines the ratio of the variance var of the probability to the information entropy entropy as a combination index; S2: inputting the combination index into a weight self-adaptive model to obtain loss weights alpha and beta; S3: substituting the loss weights alpha and beta into a combined robust loss function to obtain a loss value Lauto, and updating the parameters of the main model; and S4: substituting the same training data into the updated main model, again calculating the combined loss value Lauto, and updating the parameters of the weight self-adaptive model according to the combined loss value Lauto. The application solves the problem that the weight of the combined robust loss needs to be determined by manual testing, can effectively guarantee optimal precision, and overcomes the defects of time and labor consumption caused by manual testing of the weight.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image retrieval, in particular to a single-domain remote sensing image retrieval method based on a weight adaptive combination robust loss. BACKGROUND

[0002] In the face of the arrival of the era of remote sensing big data, the analysis and application range of massive remote sensing image data is expanding, and the timeliness and intelligence level of earth space information science data processing are also improving. The demand for automatic image retrieval from large remote sensing databases is becoming more and more urgent. Remote sensing image retrieval technology realizes the effective retrieval of massive remote sensing images through the similarity matching between image features, that is, "image searching image". Exploring how to efficiently retrieve the target or scene of interest from large-scale remote sensing image data is an important task to promote the sharing and effective use of remote sensing image data.

[0003] With the advent of deep learning, the feature representation of images gradually shifts from hand-crafted features to feature learning, and has shown great superiority in the field of remote sensing. It is widely used in various tasks of remote sensing image processing such as scene classification and image retrieval. However, deep learning technology relies heavily on large-scale labeled data. If the accuracy of data set labeling is too low, the performance of deep learning method will be severely reduced, and the retrieval accuracy will be reduced. Traditional data set labeling is usually manually labeled by experts in the relevant field, and the labeling accuracy can be guaranteed, but this method is very time-consuming and costly. In order to reduce the labeling cost and improve the labeling speed of remote sensing images, many scholars propose to use clustering and crowdsourcing to improve the labeling efficiency. However, these methods can save a lot of manual labeling cost and complete large-scale data labeling tasks quickly, but due to the lack of experience of labelers or the complexity of data, label noise, i.e. the class of incorrectly labeled samples, is inevitably introduced. According to statistics, the proportion of noise labels in existing data sets is between 8.0% and 38.5%. Therefore, it is necessary to improve the robustness of remote sensing image retrieval to label noise.

[0004] Existing deep learning-based optical remote sensing image retrieval methods rely on the basic assumption that the training data is clean, which is contrary to the real situation that label noise randomly exists in the actual application scene, resulting in the retrieval accuracy of most existing remote sensing image retrieval methods being reduced in the presence of label noise. Robust loss function is a simple and general anti-noise method. Applying robust loss function to remote sensing image retrieval method can effectively alleviate the problem of retrieval accuracy reduction caused by label noise, improve the generalization ability of the model. Considering that the remote sensing image retrieval model should balance the retrieval accuracy and generalization ability, in the actual application, as long as the model retrieval accuracy is ideal, as much label information as possible should be retained to improve the gain of sample diversity. The combination of robust loss can meet the above requirements, and the effect is superior to that of using a certain robust loss function alone. Therefore, the use of combined robust loss is the best, however, since the setting of the weight of the combined robust loss is related to the complexity of the data set, the loss weight that can obtain the optimal robustness under different label noise ratios of different data sets needs to be determined manually by testing, which has the disadvantages of difficulty to ensure the continuity of the optimal accuracy, time-consuming and laborious to obtain the optimal accuracy by manual testing.

[0005] Therefore, we propose a single-domain remote sensing image retrieval method based on weight adaptive combined robust loss to solve the above technical problems. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a single-domain remote sensing image retrieval method based on weight adaptive combined robust loss, which can automatically assign the weight of the combined robust loss, add a weight active learning network based on a multi-layer perception, and iteratively update the parameters of the main model and the adaptive weight learning network during the model training phase. Then save the optimal model in the training process through model verification, and finally test and evaluate it to realize robust remote sensing image retrieval under different proportions of label noise in the data set.

[0007] To solve the above technical problems, the present application provides a single-domain remote sensing image retrieval method based on weight adaptive combined robust loss, which adopts the following steps:

[0008] S1: input the training data into the main model and extract the image features, the main model outputs the probability that the image belongs to each category according to the image features and calculates the variance var and information entropy entropy of the main model output data, and defines the ratio of the variance var to the information entropy entropy as a combination index;

[0009] S2: input the combination index into the weight adaptive model, and obtain the loss weights alpha and beta according to the function approximation strategy of the weight adaptive model;

[0010] S3: substituting the loss weights a and b into the combined robust loss function to obtain a loss value Lauto, and updating the main model parameters;

[0011] S4: substituting the same training data into the updated main model, again calculating the combined loss value Lauto and updating the parameters of the weight adaptive model.

[0012] Preferably, in step S1, the main model is a deep convolutional neural network, which uses convolution and pooling means to extract features layer by layer from low-level visual features to obtain abstract features containing high-level semantic information, so as to be able to obtain feature extraction of the training data.

[0013] Preferably, the deep convolutional neural network is a residual artificial neural network, which includes a convolution module for changing the dimension of the network and a discriminative module for increasing the depth of the network, and both the convolution module and the discriminative module include batch normalization, so as to be able to maintain the convergence speed of the deep convolutional neural network.

[0014] Preferably, in step S1, the formula for calculating the variance var is as follows:

[0015]

[0016] The formula for calculating the information entropy entropy is as follows:

[0017]

[0018] Where C represents the total number of labels, x i represents the floating point value of the model output, represents the average value of x i , p i represents the probability value obtained after the model output is activated by a softmax function, and the formula for calculating the softmax function is as follows:

[0019]

[0020] Here, the softmax function is an activation function, and the activation layer formed by the function can give the deep neural network the ability to express nonlinear features.

[0021] Preferably, in step S2, the weight adaptive model includes an input layer composed of combination indexes, a hidden layer composed of a plurality of neurons, and an output layer composed of loss weights a and b, and the input layer, the hidden layer and the output layer are communicatively connected, so as to be able to output the loss weights a and b through the action of the neurons.

[0022] Preferably, the loss weight alpha and the loss weight beta are updated synchronously.

[0023] Preferably, in step S3, the loss value Lauto is calculated according to the following formula:

[0024] L auto = alpha * L1 + beta * L2

[0025] Wherein, alpha and beta are loss weights, L1 is an active loss function, specifically a normalized cross-entropy loss function, and the calculation formula is as follows:

[0026]

[0027] L2 is a passive loss function, specifically a reverse cross-entropy loss function, and the calculation formula is as follows:

[0028]

[0029] Wherein, y i is the one-hot encoding of the true label, p i is the sample prediction probability

[0030] Preferably, in step S3, the parameters of the master model are updated using a gradient descent algorithm.

[0031] Preferably, in step S5, after completing the training rounds of all the training data, the validation set is used to validate all the trained master models and save the optimal master model, and the optimal master model is tested and evaluated in terms of the average precision mean value mAP, and the calculation formula of the average precision mean value mAP is as follows:

[0032]

[0033]

[0034] Wherein, Q represents the number of to-be-queried remote sensing images; AveP(q) represents the average precision of the qth to-be-queried remote sensing image; k represents the kth ranking position of the query result; n represents the number of returned query results; M represents the number of the same categories as the to-be-queried image in the remote sensing image query library; P(k) represents the accuracy of the top k results; and rel(k) represents whether the kth image in the query result is the same as the query image category, 1 for same and 0 for different.

[0035] Preferably, the training data contains label noise.

[0036] Through the above technical solutions, the present application has the following advantages:

[0037] In the present application, firstly, the training data is input into the main model and the image features are extracted, the main model outputs the probability of the image belonging to each category according to the image features and calculates the variance var and the information entropy entropy of the main model output data, and the ratio of the variance var and the information entropy entropy is defined as the combination index, secondly, the combination index is input into the weight adaptive model, the loss weights alpha and beta are obtained according to the function approximation strategy of the weight adaptive model, then the loss weights alpha and beta are substituted into the combined robust loss function to obtain the loss value Lauto, and the main model parameters are updated, further, the same training data is substituted into the updated main model, the combined loss value Lauto is calculated again and the parameters of the weight adaptive model are updated, finally, the optimal main model in the training process is saved, the optimal main model is tested and the average precision mean mAP is evaluated, the method is based on the combined robust loss, the weight active learning network is designed based on the multilayer perception, the weight adaptive combined robust loss function is constructed, the single domain remote sensing image retrieval under the condition of labeled noise is realized, the problem that the weight of the combined robust loss needs to be determined by manual test is solved, the continuity of the optimal precision can be effectively guaranteed, and the time-consuming and labor-consuming defects of manual test weight are overcome.

[0038] Other features and advantages of the present application will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0039] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, and are used together with the following specific embodiments to explain the present application, but do not constitute a limitation on the present application. In the drawings:

[0040] Fig. 1 is a flow chart of the single domain remote sensing image retrieval method based on the weight adaptive combined robust loss of the present application;

[0041] Fig. 2 is a schematic diagram of the weight adaptive model of the single domain remote sensing image retrieval method based on the weight adaptive combined robust loss of the present application;

[0042] Fig. 3 is a comparison diagram of the noise robust remote sensing image retrieval method of manual weighting and automatic weighting. DETAILED DESCRIPTION

[0043] The specific embodiments of the present application will be described in detail below in combination with the drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0044] First of all, it needs to be explained that in the following description, some orientation words related to the technical scheme of the present application are only for the convenience of describing the present application and simplifying the description, and are not intended to indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0045] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be direct connection, or indirect connection through intermediate medium, it can be internal communication of two elements or interaction relationship of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0046] The present application provides a single-domain remote sensing image retrieval method based on weight adaptive combination robust loss, as shown in Figs. 1-3 The method adopts the following steps:

[0047] S1: input the training data into the main model and extract the image features, the main model outputs the probability of the image belonging to each category according to the image features and calculates the variance var and information entropy entropy of the main model output data, and defines the ratio of the variance var and the information entropy entropy as the combination index;

[0048] S2: input the combination index into the weight adaptive model, and obtain the loss weights alpha and beta according to the function approximation strategy of the weight adaptive model;

[0049] S3: substitute the loss weights alpha and beta into the combination robust loss function to obtain the loss value Lauto, and update the main model parameters;

[0050] S4: substitute the same training data into the updated main model, calculate the combination loss value Lauto again, and update the parameters of the weight adaptive model.

[0051] It should be noted that in step S1, the training data is input into the main model and the image features are extracted, the main model outputs the probability of the image belonging to each category according to the image features and calculates the variance var and information entropy entropy of the main model output data, and defines the ratio of the variance var and the information entropy entropy as the combination index:

[0052] The main model is a deep convolutional neural network. The neural network model can be any one of a visual geometry group network, a residual artificial neural network, a densely connected network, or a conventional convolutional neural network model or a variant model thereof. Specifically, the neural network model is configured to extract features from low-level visual features by convolution, pooling, and other operations, and finally obtain abstract features containing high-level semantic information, thereby realizing feature extraction of the training data. Preferably, the neural network model is a residual artificial neural network, which includes a convolution module and a discriminator module. The former cannot be connected in series and is configured to change the network dimension. The latter can be connected in series and is configured to increase the network depth. Both the convolution module and the discriminator module include batch normalization, which can maintain a faster network convergence speed. The number of training data is set to 128. It should be noted that for an image, the variance var is calculated according to the following formula:

[0053]

[0054] The information entropy entropy is calculated according to the following formula:

[0055]

[0056] where C represents the total number of labels, x i represents a floating-point value output by the model, represents the average value of x i , and p i represents a probability value obtained by activating the model output by using a softmax function. The calculation formula of the softmax function is as follows:

[0057]

[0058] Here, the softmax function is an activation function, and an activation layer formed by the function can give the deep neural network the ability to express nonlinear features.

[0059] Generally, the more diverse the corresponding labels of the training data are, the smaller the variance var is, and the higher the information entropy entropy is. Therefore, the ratio of the information entropy entropy to the variance var is used as a combined index for measuring the diversity of the sample. The trend of the index changing with the diversity of the data labels is consistent with the information entropy entropy, and the influence of the diversity of the data sample can be further expanded. Then, the combined index of the corresponding data is saved in the form of a floating-point tensor, which is input into the weight adaptive model (AWNet) in batches, and finally the reasonable adjusted weight hyperparameters are returned and fed back to the combined robust loss function.

[0060] It should be noted that in the S2 step, the combined index is input into the weight adaptive model, and the loss weights a and b are obtained according to the function approximation strategy of the weight adaptive model:

[0061] The weight adaptive model comprises an input layer, a hidden layer and an output layer, the input layer is a combined index of each batch of training data, the hidden layer comprises 20 neurons, and the output layer is a loss weight a and a loss weight β of different loss functions. The input layer, the hidden layer and the output layer are connected and arranged, so that the combined index can be output as the loss weight a and the loss weight β through the action of the neurons. Here, the loss weight a and the loss weight β are updated synchronously, and the combined robust loss can effectively alleviate the adverse effects of training data containing noise labels on the remote sensing image retrieval model. In the traditional combined loss, the weights of each loss term need to be determined manually, see Fig. 2 The manual weighting noise robust remote sensing image retrieval method has the following process: before training, the weights of each loss term in the combined loss function and other hyperparameters are manually assigned, then the training data are input into the main model (Model) to start training, and the loss value of the model output (Output) is calculated by using the combined loss, so as to update the main model parameters by back propagation and gradient descent. The main model is iteratively updated according to the training batch, until the training of this round is completed. Then, the trained main model is verified for accuracy, and after completing the training of all rounds, the model with the highest verification accuracy is saved as the optimal main model, and the optimal main model is evaluated by using the test data to obtain the final retrieval accuracy. However, if the combined robust loss is to be optimized, the method needs to manually try different combinations of loss term weights multiple times, which is time-consuming and labor-intensive. Compared with the manual weighting method, the automatic weighting noise robust remote sensing image retrieval method does not need to consider each loss weight in the combined loss as a hyperparameter. In the training process, the weights of each loss term can be automatically generated by the model output and the weight adaptive model, and the main model and the weight adaptive model are iteratively updated by using the weighted combined loss, so as to realize the adaptive weighting of the combined loss, ensure the optimal effect of the combined robust loss, and reduce the time and labor cost of manual weight testing.

[0062] It should be noted that in the S3 step, the loss weights a and β are substituted into the combined robust loss function to obtain the loss value Lauto, and the main model parameters are updated.

[0063] The calculation formula of the loss value Lauto is as follows:

[0064] L auto =α·L1+β·L2

[0065] Wherein, a and β are loss weights, L1 is an active loss function, which can be any one of a cross-entropy loss function, a normalized cross-entropy loss function, a focal loss function, a normalized focal loss function, etc. Here, preferably, the active loss function is a normalized cross-entropy loss function, and its calculation formula is as follows:

[0066]

[0067] L2 is a passive loss function, which can be any one of mean absolute error loss function, normalized mean absolute error loss function, reverse cross-entropy loss function, normalized reverse cross-entropy loss function, and here, preferably, the passive loss function is the reverse cross-entropy loss function, and its calculation formula is as follows:

[0068]

[0069] where y i is the one-hot encoding of the true label, p i is the sample prediction probability.

[0070] The loss value Lauto is updated by the back propagation strategy and the gradient descent algorithm, the gradient descent algorithm is an optimization algorithm for finding the minimum value of a function, the process of training the model is to make the loss function smaller and smaller through continuous iteration calculation to obtain the minimum value, the smaller the loss function is, the better, and theoretically, the value of the loss function is 0, which represents the optimal main model, that is, the actual output of the neural network is as close as possible to the expected output, in this process, the weights of each layer of network parameters will be adjusted, which is also the training principle of neural network.

[0071] It should be noted that in the S4 step, the same training data is substituted into the updated main model, and the combined loss value Lauto is calculated again to update the parameters of the weight adaptive model.

[0072] In the process of weight updating, two updates are designed, namely the parameter update of the main model and the parameter update of the weight adaptive model. First, the training data is input into the main model, the training data contains label noise, the features contained in the input data are extracted and encoded through the convolution layer, combined with the downsampling operation of the pooling layer and the nonlinear mapping ability of the activation layer, the floating-point main model output result is output from the output layer, and then the combined index is input into the weight adaptive model in batches according to the calculation formula of variance var and information entropy entropy, and through the nonlinear mapping ability of the hidden layer neurons in the weight adaptive model, the loss weights a and b of the updated combined robust loss function are output, after the loss weights a and b are updated, the combined loss value Lauto is calculated through the combined robust loss function to update the parameters of the main model, then the same training data is input into the updated main model, the combined loss value Lauto is calculated again, and the weight adaptive model is updated, the above iteration update process is repeated until all training rounds of the training set are completed. Whenever a round of training is completed, the model is verified through the validation set, and the optimal model in all training rounds is saved. Finally, the optimal main model saved in the verification stage is tested and precision evaluated through the test set.

[0073] It also needs to be explained that when the training round of all training data is completed, the validation set is used to verify the main model obtained by training and save the optimal main model, and the optimal main model is tested and the average precision mean mAP is evaluated, and the calculation formula of the average precision mean mAP is as follows:

[0074]

[0075]

[0076] Wherein, Q represents the number of the to-be-queried remote sensing image; AveP(q) represents the average precision of the qth to-be-queried remote sensing image; k represents the kth ranking position of the query result; n represents the number of returned query results; M represents the number of the same categories as the to-be-queried image in the remote sensing image query library; P(k) represents the accuracy of the first k results; rel(k) represents whether the kth image in the query result is the same as the query image category, and 1 represents the same, and 0 represents the different.

[0077] The present application adds 5%, 10%, 20% and 30% four kinds of artificial noise labels in the land use remote sensing image data set to simulate the random noise scene in the actual application, and constructs the simulated noise data set, and the specific method is to change the real label of the training data to any other category label except the real label according to the noise ratio, so as to simulate the noise label in the real scene. The experiment follows the data division standard of small-scale data set 60%, 20%, 20%, and randomly selects 60% of the images in the data set as the training set, 20% as the validation set, and 20% as the test set. The training set and the validation set data are noisy data, the test set data are clean data, and the test set data are only used for evaluation of various methods, not participating in the training process, and the experimental results are shown in Table 1:

[0078] Table 1 Average precision mean (mAP) of land use remote sensing image data set under different noise rates

[0079]

[0080] The experimental results are shown in Table 1, which shows the retrieval accuracy of the traditional remote sensing image retrieval method (ResNet50 not robust), the active loss weight a e [0.1, 1, 10], the passive loss weight β e [0.1, 1, 10, 100], the 12 kinds of manual weighting combinations and the weight adaptive method APL auto proposed in this paper under the conditions of label noise rate of 5%, 10%, 20% and 30%. The difference between the highest accuracy and the lowest accuracy of the 12 kinds of manual weighting combinations under each noise rate is calculated, and the result is recorded as the accuracy fluctuation of the manual weighting method.

[0081] Through observation and analysis, it can be seen that the traditional remote sensing image retrieval method (ResNet50 not robust) does not have robustness, and as the noise rate increases, the retrieval accuracy will decrease significantly. The active loss robust loss function needs to use artificial testing to determine the active loss weight a and the passive loss weight β. The disadvantage of this method is that it is difficult to ensure the persistence of optimal accuracy. As shown in Table 1, among the 12 kinds of manual weighting combinations of active loss weight a e [0.1, 1, 10] and passive loss weight β e [0.1, 1, 10, 100], the retrieval accuracy of different weight combinations will fluctuate by 2.05%-7.10%, and the testing weight process is very time-consuming. The weight adaptive method APL auto proposed in this paper achieves results comparable to the optimal retrieval accuracy of the manual weighting APL, but takes only the same time as testing one artificial weight combination. It maintains the stability of the optimal accuracy of the robust loss while greatly saving time and labor costs. In addition, compared with the original label noise not robust method, the method proposed in this paper has obvious label noise robust effect. For example, when the data set contains 20% label noise, the retrieval accuracy of the method proposed in this paper is 92.85%, compared with the not robust retrieval method of 79.18%, which improves the average precision mAP index by 13.67%.

[0082] From the above description, it can be seen that, in the present application, first, the training data is input into the main model and the image features are extracted, the main model outputs the probability of the image belonging to each category according to the image features and calculates the variance var and the information entropy entropy of the main model output data, and the ratio of the variance var and the information entropy entropy is defined as the combination index, second, the combination index is input into the weight adaptive model, the weight adaptive model obtains the loss weights a and b according to the function approximation strategy, then, the loss weights a and b are substituted into the combined robust loss function to obtain the loss value Lauto, the loss value Lauto updates the main model parameters through the back propagation strategy and the gradient descent algorithm, then, the same training data is substituted into the updated main model, the combined loss value Lauto is calculated again and the parameters of the weight adaptive model are updated, finally, the optimal main model in the training process is saved, the optimal main model is tested and the average precision mean mAP is evaluated, the present application is based on the combined robust loss, designs the weight active learning network based on the multilayer perceptron, constructs the weight adaptive combined robust loss function, realizes the single domain remote sensing image retrieval under the condition of labeled noise, solves the problem that the weight of the combined robust loss needs to be determined by manual testing, can effectively guarantee the persistence of the optimal precision, and overcomes the time-consuming and laborious shortcomings of manual testing weight.

[0083] The preferred embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the specific details in the above-described embodiments, and various simple modifications can be made to the technical solutions of the present application within the technical concept of the present application, and these simple modifications all belong to the protection scope of the present application.

[0084] In addition, it should be noted that each specific technical feature described in the above specific embodiments can be combined in any appropriate manner without contradiction, and in order to avoid unnecessary repetition, the present application will not further describe various possible combinations.

[0085] In addition, various different embodiments of the present application can also be combined in any appropriate manner, as long as they do not deviate from the idea of the present application, and they should also be considered as disclosed by the present application.

Claims

1. A single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss, characterized in that, The method adopts the following steps: S1: input training data into a main model and extract image features, the main model outputs the probability of the image belonging to each category according to the image features and calculates the variance var and the information entropy entropy of the main model output data, and defines the ratio of the variance var to the information entropy entropy as a combination index; S2: input the combination index into a weight adaptive model to obtain loss weights a and b according to the function approximation strategy of the weight adaptive model; the weight adaptive model includes an input layer composed of the combination index, a hidden layer composed of a plurality of neurons, and an output layer composed of the loss weights a and b, and the input layer, the hidden layer and the output layer are communicatively connected to enable the combination index to be output through the neurons to output the loss weights a and b; S3: substitute the loss weights a and b into a combined robust loss function to obtain a loss value Lauto, and update the parameters of the main model; S4: substitute the same training data into the updated main model, calculate the combined loss value Lauto again, and update the parameters of the weight adaptive model.

2. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, characterized in that, In step S1, the main model is a deep convolutional neural network, which uses convolution and pooling methods to extract features from low-level visual features at a shallow level, layer by layer, to obtain abstract features containing high-level semantic information, so as to enable feature extraction of the training data.

3. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 2, characterized in that, The deep convolutional neural network is a residual artificial neural network, which includes a convolution module for changing the dimension of the network and a discrimination module for increasing the depth of the network, and both the convolution module and the discrimination module include batch normalization to maintain the convergence speed of the deep convolutional neural network.

4. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, characterized in that, In step S1, the calculation formula of the variance var is as follows: The calculation formula of the information entropy entropy is as follows: wherein, denotes the total number of labels, denotes the float value output by the model, denotes the average value of denotes the probability value obtained after the model output is activated by a softmax function, and the calculation formula of the softmax function is as follows:​ Wherein, the softmax function is an activation function, and the activation layer formed by the function can give the deep neural network the ability to express nonlinear characteristics.

5. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, wherein, The loss weights a and b are updated synchronously.

6. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, wherein, In step S3, the calculation formula of the loss value Lauto is as follows: wherein a and b are loss weights, is an active loss function, specifically a normalized cross-entropy loss function, whose calculation formula is as follows: The passive loss function is specifically a reverse cross-entropy loss function, and its calculation formula is as follows: wherein, is a one-hot encoding of the true label, is the sample prediction probability.

7. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, characterized in that, In step S3, the parameters of the main model are updated by using the gradient descent algorithm.

8. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to claim 1, characterized in that, After completing all the training rounds of the training data, the validation set is used to verify all the trained main models and save the optimal main model, and the optimal main model is tested and evaluated by the average precision mean mAP, and the calculation formula of the average precision mean mAP is as follows: Wherein, Q represents the number of to-be-queried remote sensing images; AveP(q) represents the average precision of the qth to-be-queried remote sensing image; k represents the kth ranking position of the query result; n represents the number of returned query results; M represents the number of the same categories in the remote sensing image query library as the to-be-queried image; P(k) represents the accuracy of the first k results; rel(k) represents whether the kth image in the query result is the same as the query image category, and 1 represents the same and 0 represents the different.

9. The single-domain remote sensing image retrieval method based on weight-adaptive combination robust loss according to any one of claims 1-8, characterized in that, The training data contains label noise.

Images