Remote sensing image multi-label classification method and device, equipment, medium and product
By performing weak transformation and strong transformation processing on remote sensing images, combined with parallel multi-label classification model and multi-layer perceptron, correcting the noise label, the robustness problem of multi-label remote sensing image classification model under noise labels is solved, and the classification results are improved accuracy and generalization ability.
Patent Information
- Application Number
- CN202411124220.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-16
AI Technical Summary
The existing multi-label remote sensing image classification methods are susceptible to noise labels, making it difficult to train a robust multi-label remote sensing image classification model, resulting in inaccurate classification results.
The remote sensing image is processed through weak transformation and strong transformation to generate weak transformation images and strong transformation images. The multi-label remote sensing image classification model and multi-layer perceptron are used to extract image features and predict label vectors respectively, and correct noise labels are corrected by using the parallel multi-label remote sensing image classification model and multi-layer perceptron. , calculate the loss function to train the model.
The accuracy, generalization ability and robustness of the multi-label remote sensing image classification model are improved, and the accuracy of classification results is enhanced, especially in a weak supervision environment.
Smart Images

Figure CN118918385A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer vision technology, and in particular to a remote sensing image multi-label classification method, device, equipment, medium and product. Background Art
[0002] With the rapid development of remote sensing technology, remote sensing image classification systems based on artificial intelligence have been widely used in land detection, urban planning, vegetation cover detection and other fields. Compared with conventional images, remote sensing images are rich in a variety of semantic information. The scale and number of targets are quite different from those of conventional images. The original single-label classification method is often used to solve the problem of remote sensing image classification, which often results in the loss of semantic information. The multi-label remote sensing image classification algorithm aims to assign multiple labels to a remote sensing image. Using multi-label remote sensing image classification can make full use of remote sensing image semantics, which is of great significance for further interpretation of remote sensing images.
[0003] In the related technology, the research on multi-label remote sensing image classification algorithms can be roughly divided into traditional machine learning and deep learning-based methods. Early multi-label classification algorithms generally extract image features from remote sensing images, such as texture features, spectral features, etc., and then classify and distinguish labels through classifiers and design optimization algorithms. Since traditional artificial features are insensitive to changes in illumination and difficult to capture complex information, deep learning can overcome the shortcomings of traditional methods by automatically learning hierarchical feature representations in images and has stronger feature capture capabilities. However, image classification algorithms based on deep learning are highly sensitive to data and usually require a large amount of labeled data sets for training. This type of data labeling is affected by the subjective consciousness of the labeler, etc. It is difficult to avoid mislabeling and missing labels during the labeling process, resulting in noisy labels. Using noisy labels to train multi-label remote sensing image classification models will greatly affect the generalization ability of the model. Therefore, under the interference of noise, the accuracy and generalization ability of the multi-label remote sensing image classification model in the existing multi-label classification method based on deep learning are largely affected by the noise labels and it is also difficult to train a multi-label remote sensing image classification model with high robustness, resulting in inaccurate multi-label classification results. Summary of the invention
[0004] The purpose of this application is to provide a remote sensing image multi-label classification method, device, equipment, medium and product, which can improve the accuracy, generalization ability and robustness of the multi-label remote sensing image classification model, thereby improving the accuracy of the multi-label classification results.
[0005] To achieve the above objectives, this application provides the following solutions:
[0006] In a first aspect, the present application provides a remote sensing image multi-label classification method, comprising:
[0007] Acquire a training sample set, the training sample set includes a plurality of training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; the label vector of the remote sensing image is obtained according to a clean label and a noise label representing a category of the remote sensing image;
[0008] For a remote sensing image in any training sample in the training sample set, the remote sensing image is processed by using a weak transformation and a strong transformation respectively to obtain a weak transformation image and a strong transformation image corresponding to the remote sensing image;
[0009] Under the current training times, the weakly transformed image corresponding to the remote sensing image is input into the first branch of the sample model under the previous training times to obtain the output vector and implicit features of the classifier in the first branch under the current training times; the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times is obtained according to the output vector of the classifier in the first branch under the current training times; the sample model includes: a first branch and a second branch arranged in parallel; the first branch and the second branch both include a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model includes a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch;
[0010] Inputting the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; obtaining the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times according to the output vector of the classifier in the second branch at the current training times;
[0011] Correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number;
[0012] Calculate the value of the loss function at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number;
[0013] According to the value of the loss function under the current training times, the sample model under the previous training times is trained to obtain the sample model under the current training times, and it is judged whether the set training times are reached. If the set training times are not reached, the next training times are entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model is used to perform multi-label classification on the target remote sensing image.
[0014] In a second aspect, the present application provides a remote sensing image multi-label classification device, comprising:
[0015] An acquisition module is used to acquire a training sample set, wherein the training sample set includes a plurality of training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; the label vector of the remote sensing image is obtained according to a clean label and a noise label representing a category of the remote sensing image;
[0016] A data enhancement module, for processing a remote sensing image in any one of the training samples in the training sample set by using a weak transformation and a strong transformation respectively, to obtain a weak transformation image and a strong transformation image corresponding to the remote sensing image;
[0017] The first branch calculation module is used to input the weakly transformed image corresponding to the remote sensing image into the first branch of the sample model under the previous training times under the current training times to obtain the output vector and implicit features of the classifier in the first branch under the current training times; obtain the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times according to the output vector of the classifier in the first branch under the current training times; the sample model includes: a first branch and a second branch set in parallel; the first branch and the second branch both include a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model includes a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch;
[0018] The second branch calculation module is used to input the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; and obtain the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times according to the output vector of the classifier in the second branch at the current training times;
[0019] A correction module, used for correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number;
[0020] A loss function calculation module is used to calculate the value of the loss function at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number;
[0021] The training module is used to train the sample model under the previous training times according to the value of the loss function under the current training times to obtain the sample model under the current training times, and to determine whether the set training times have been reached. If the set training times have not been reached, the next training times will be entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model will be used to perform multi-label classification on the target remote sensing image.
[0022] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the multi-label classification methods for remote sensing images described above.
[0023] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the multi-label classification methods for remote sensing images described above.
[0024] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the multi-label classification methods for remote sensing images described above.
[0025] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0026] The present application provides a remote sensing image multi-label classification method, device, equipment, medium and product. The remote sensing image is processed by using weak transformation and strong transformation respectively to obtain the weak transformation image and strong transformation image corresponding to the remote sensing image, and the label vector of the remote sensing image is corrected according to the output vector of the classifier in the first branch under the current training times and the output vector of the classifier in the second branch under the current training times to obtain the corrected label vector of the remote sensing image under the current training times. Noise labels are corrected as much as possible in complex noise label training to improve the performance of multi-label classification of remote sensing images in a weakly supervised environment. The problem that the existing multi-label remote sensing image classification method is easily affected by noise labels and it is difficult to train a robust multi-label remote sensing image classification model is solved, and the technical effect of improving the accuracy, generalization ability and robustness of the multi-label remote sensing image classification model and thus improving the accuracy of the multi-label classification results is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0028] Figure 1 This is a flowchart of a multi-label classification method for remote sensing images in one embodiment of the present application;
[0029] Figure 2 This is a schematic diagram of a multi-label classification method for remote sensing images in one embodiment of the present application;
[0030] Figure 3 This is a sample model data processing flow chart of a remote sensing image multi-label classification method in one embodiment of the present application;
[0031] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0032] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0033] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0034] In an exemplary embodiment, a remote sensing image multi-label classification method is provided, specifically a robust weakly supervised remote sensing image multi-label classification method, such as Figure 2 As shown, firstly, the training set images (including multiple remote sensing images) are obtained, and two kinds of transformations are performed on the remote sensing images through strong transformation and weak transformation processing to obtain two types of views with large differences, and the deep features of the two views are respectively extracted through two backbone networks with shared weights; then, based on the two image features, the classifiers are used to predict the category scores (predicted label vectors), and the two image features are projected into the latent space by a multilayer perceptron to obtain the implicit features; then, the internal consistency constraints are defined based on the implicit features of the two types of views, and the structural consistency constraints are defined based on the implicit features and predicted label vectors of different images, and these two constraints are introduced into the binary cross entropy loss function; finally, the sorting error is defined according to the predicted label vectors of the two enhanced images to identify and correct the noise labels, and the network training is performed. As shown Figure 1 As shown, the specific steps include:
[0035] Step 201: Obtain a training sample set, the training sample set includes multiple training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; in order to train a robust multi-label remote sensing image classification model, it is necessary to obtain remote sensing images with noise labels as training samples, so the label vector of the remote sensing image is obtained based on the clean label and the noise label representing the remote sensing image category, and the noise label should include two categories, pseudo positive labels and pseudo negative labels. The training sample set is represented by D, D = {x i ,y i )|1≤i≤N}, where N represents the total number of training samples, x i is the remote sensing image in the i-th training sample, y i ={y i1 ,y i2 ,...,y iq}∈{0,1} n×q is the label vector of the remote sensing image in the i-th training sample, where y iq =1 means that the qth label is the true category of the remote sensing image in the ith training sample, y iq=0 means that the qth label is not the true category of the remote sensing image in the ith training sample. For example, {0, 1, 0, 1, 0, 0} means that the image set has a total of 6 label categories. The remote sensing image does not have the 1st, 3rd, 5th, and 6th category labels, which can be called the negative label of the sample; while the 2nd and 4th category labels belong to the sample, which can be called the positive label of the sample. Before the sample training, in order to ensure that the noise label is introduced into the training sample, the clean label needs to be denoised. The denoising process is: select a certain noise ratio, and randomly add additive noise and subtractive noise in proportion. Among them, the addition of additive noise is achieved by flipping 0 in the selected label to 1, while the subtractive noise is to flip 1 to 0, ensuring that both the pseudo positive label and the pseudo negative label are introduced into the training set as the introduced noise label.
[0036] Step 202: For any remote sensing image in the training sample set, weak transformation (through simple geometric transformations such as cropping, rotation and scaling) and strong transformation (first geometric transformation and then non-geometric transformations such as brightness, contrast and saturation) are used to process the remote sensing image to obtain a weak transformation image and a strong transformation image corresponding to the remote sensing image. Through weak transformation and strong transformation, two images with large differences are obtained. In order to ensure the consistency of the size of the training sample images and to meet the requirements of the multi-label remote sensing image classification model for the input image, the training sample images are weakly transformed.
[0037] Step 203: Under the current training times, the weakly transformed image corresponding to the remote sensing image is input into the first branch of the sample model under the previous training times to obtain the output vector and implicit features of the classifier in the first branch under the current training times; the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times is obtained according to the output vector of the classifier in the first branch under the current training times. The sample model includes: a first branch and a second branch set in parallel; the first branch and the second branch both include a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model includes a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch. The classifier includes a flattening layer and a fully connected layer, wherein the fully connected layer obtains an output vector z, and then z is subjected to softmax to obtain a posterior probability, and then a predicted label vector is obtained according to the posterior probability and a set threshold.
[0038] Step 204: input the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; obtain the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times based on the output vector of the classifier in the second branch at the current training times.
[0039] Step 205: Correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training times and the output vector of the classifier in the second branch at the current training times to obtain the corrected label vector of the remote sensing image at the current training times.
[0040] Step 206: Calculate the value of the loss function at the current training times according to the corrected label vector of each remote sensing image in the training sample set at the current training times, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training times, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training times.
[0041] Step 207: According to the value of the loss function under the current training times, the sample model under the previous training times is trained to obtain the sample model under the current training times, and it is determined whether the set training times have been reached. If the set training times have not been reached, the next training times are entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model is used to perform multi-label classification on the target remote sensing image. Specifically, the target remote sensing image is input into the trained multi-label remote sensing image classification model to obtain an output vector, and the posterior probability of each category of the target remote sensing image is obtained according to the output vector and the softmax function. If the posterior probability of a certain category is greater than the set threshold, then this category is the label of the target remote sensing image, i.e., the positive label, otherwise it will be determined as an irrelevant label, i.e., a negative label.
[0042] In another exemplary embodiment of the present application, step 203 and step 204 are specifically as follows:
[0043] For the weakly transformed images and strongly transformed images corresponding to the remote sensing images, the feature encoder with shared weights (ResNet18 network) can be used as the backbone framework to perform deep feature extraction on the input images (weakly transformed images and strongly transformed images) to obtain the deep features of the weakly transformed images and strongly transformed images corresponding to the remote sensing images.
[0044] Then, the deep features of the weakly transformed image corresponding to the remote sensing image are input into the classifier and multi-layer perceptron in the first branch for further extraction and transformation to obtain the output vector and implicit features of the weakly transformed image corresponding to the remote sensing image. The multi-layer perceptron further projects the representation vector into the latent space to obtain the implicit features. The classifier makes predictions based on the encoder output to obtain the output vector. Similarly, the deep features of the strongly transformed image corresponding to the remote sensing image are input into the classifier and multi-layer perceptron in the second branch to obtain the output vector and implicit features of the weakly transformed image corresponding to the remote sensing image. The predicted label vector is obtained based on the output vector.
[0045] This application adopts a collaborative learning model to extract deep features of images through a feature encoder to obtain deep features of the image; the high-level features of the network often contain rich semantic information. In order to extract rich semantic information in noisy scenes, based on the feature encoder, a classifier and a multi-layer perceptron are used to output output vectors and implicit features respectively, which are used for subsequent supervised training and similarity constraint training.
[0046] In another exemplary embodiment of the present application, step 205 is replaced by the following:
[0047] Construct the positive label set L of the remote sensing image (the remote sensing image in the i-th training sample set) i and negative label set The positive label set includes the real categories of the remote sensing image; the negative label set includes the categories that the remote sensing image does not belong to. Both sets may contain pseudo positive labels and pseudo negative labels.
[0048] Based on the output vector of the classifier in the first branch under the current number of training times, the posterior probability of the classifier in the first branch for each positive label in the positive label set and the posterior probability of the classifier in the first branch for each negative label in the negative label set are obtained. The posterior probability of the classifier in the cth branch for the i-th label is the posterior probability obtained by softmax mentioned above, according to the formula Calculate, where z i (c) and z j (c) They represent the i-th and j-th elements in the output vector z of the classifier in the c-th branch respectively.
[0049] For any positive label in the positive label set, the probability that the positive label is a pseudo-positive label is obtained from the perspective of the classifier in the first branch based on the posterior probability of the positive label by the classifier in the first branch and the posterior probability of each negative label in the negative label set by the classifier in the first branch. Specifically: Calculate the probability that the positive label l is a pseudo-positive label from the perspective of the classifier in the cth branch, where c is 1 or 2, where and Respectively represent the classifier in the cth branch for positive label l and negative label The posterior probability of .
[0050] For any negative label in the negative label set, the probability that the negative label is a pseudo-negative label from the perspective of the classifier in the first branch is obtained based on the posterior probability of the negative label by the classifier in the first branch and the posterior probability of each positive label in the positive label set by the classifier in the first branch. Specifically, according to the formula Calculate the negative label from the perspective of the classifier in the cth branch is the probability of a false negative label.
[0051] Based on the output vector of the classifier in the second branch at the current training times, the posterior probability of the classifier in the second branch for each positive label in the positive label set and the posterior probability of the classifier in the second branch for each negative label in the negative label set are obtained.
[0052] and They measure l and is the probability of pseudo positive and pseudo negative labels. Therefore, for any positive label in the positive label set, the probability that the positive label is a pseudo positive label from the perspective of the classifier in the second branch is obtained according to the posterior probability of the positive label by the classifier in the second branch and the posterior probability of each negative label in the negative label set by the classifier in the second branch.
[0053] For any negative label in the negative label set, the probability that the negative label is a pseudo-negative label from the perspective of the classifier in the second branch is obtained based on the posterior probability of the negative label by the classifier in the second branch and the posterior probability of each positive label in the positive label set by the classifier in the second branch.
[0054] Determine the probability that each positive label in the positive label set is a pseudo positive label from the perspective of the classifier in the first branch and the maximum positive label corresponding to the probability that each positive label in the positive label set is a pseudo positive label from the perspective of the classifier in the second branch. Specifically, according to the formula Determine the maximum positive label l * .
[0055] Determine the probability that each negative label in the negative label set is a pseudo-negative label from the perspective of the classifier in the first branch and the negative label corresponding to the maximum probability among the probabilities that each negative label in the negative label set is a pseudo-negative label from the perspective of the classifier in the second branch as the maximum negative label. Specifically, according to the formula Determine the maximum negative label
[0056] The value corresponding to the maximum positive label and the value corresponding to the maximum negative label in the label vector of the remote sensing image are corrected (flipped to "0-1" or "1-0") to obtain the corrected label vector of the remote sensing image under the current training times. When using the supervised loss function for network training, the noise label as the expected output will mislead the model, so the expected output needs to be corrected.
[0057] In another exemplary embodiment of the present application, step 206 is replaced by the following:
[0058] According to the implicit features of the weakly transformed images corresponding to each remote sensing image in the training sample set at the current training times and the implicit features of the strongly transformed images corresponding to each remote sensing image in the training sample set at the current training times, the internal consistency function value at the current training times is calculated.
[0059] The structural consistency function value at the current training number is calculated based on the predicted label vectors and implicit features of the weakly transformed images corresponding to each remote sensing image in the training sample set at the current training number and the predicted label vectors and implicit features of the strongly transformed images corresponding to each remote sensing image in the training sample set at the current training number.
[0060] The supervised loss function value for the current training number is calculated based on the corrected label vector of each remote sensing image in the training sample set for the current training number, the predicted label vector of the weakly transformed image corresponding to each remote sensing image in the training sample set for the current training number, and the predicted label vector of the strongly transformed image corresponding to each remote sensing image in the training sample set for the current training number.
[0061] The value of the loss function at the current number of training times is calculated based on the internal consistency function value at the current number of training times, the structural consistency function value at the current number of training times, and the supervised loss function value at the current number of training times.
[0062] In another exemplary embodiment of the present application, since the weakly transformed image and the strongly transformed image of the same remote sensing image come from the same remote sensing image, similar features should be obtained after different branch networks, and the internal similarity is used to constrain the implicit features of the two transformations, and the internal consistency function is defined based on the two implicit features. The internal similarity means that the same data from different transformations has the same mapping, and its similarity measure, i.e., the internal consistency function, is defined as:
[0063]
[0064] and The consistency measure between the two is defined as: express and The pairwise similarity of .
[0065] Therefore, according to the implicit features of the weakly transformed images corresponding to each remote sensing image in the training sample set under the current training times and the implicit features of the strongly transformed images corresponding to each remote sensing image in the training sample set under the current training times, the internal consistency function value under the current training times is calculated, specifically:
[0066] According to the formula Calculate the internal consistency function value, where C int represents the internal consistency function value, N represents the total number of remote sensing images in the training sample set, represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the jth training sample, τ represents the temperature parameter, exp() represents the exponential function with the natural constant e as the base, express The transpose of .
[0067] In another exemplary embodiment of the present application, the classifier makes predictions based on the output of the feature encoder. Since the outputs of the classifier and the multi-layer perceptron come from the same enhanced image (weakly transformed image or strongly transformed image), the prediction results of the classifier should be consistent with the key structural information in the implicit features, that is, there is a point-to-point structural similarity between the classifier output and the implicit features. Structural consistency constraints are defined based on implicit features and predicted label vectors. This constraint enhances the representation ability of features to a certain extent, improves the accuracy of predicted label vectors, and thus reduces the interference of noise labels. The specific structural consistency KL divergence constraint between the classifier output and the multi-layer perceptron output is: in, represents the structural consistency loss of the classifier and perceptron in the c-th branch (c=1,2); express and The Euclidean distance between express and The Euclidean distance between The parameters μ and σ represent the mean and variance of the sample x, respectively, and impose point-to-point structural similarity constraints on the classifier output and implicit features in the two-branch network. The structural similarity measure, i.e., the structural consistency function, is defined as
[0068] Therefore, the structural consistency function value under the current training times is calculated according to the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set under the current training times and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set under the current training times, specifically:
[0069] According to the formula Calculate the structural consistency function value, where C str represents the structural consistency function value, express and The Euclidean distance between them, μ represents the mean, σ represents the variance, express and The Euclidean distance between express and The Euclidean distance between express and The Euclidean distance between represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the jth training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the jth training sample, Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the jth training sample, represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, Represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the jth training sample.
[0070] In another exemplary embodiment of the present application, for the supervised loss L commonly used in network training sup , that is, BCELoss performs supervised learning on the outputs of the two classifiers respectively, which is defined as: represents the loss function based on the classifier output in the cth branch (c = 1, 2). The final supervised loss function is defined as: Therefore, the supervised loss function value under the current training times is calculated according to the corrected label vector of each remote sensing image in the training sample set under the current training times, the predicted label vector of the weakly transformed image corresponding to each remote sensing image in the training sample set under the current training times, and the predicted label vector of the strongly transformed image corresponding to each remote sensing image in the training sample set under the current training times, which is specifically:
[0071] Among them, L sup represents the supervised loss function value, N represents the total number of remote sensing images in the training sample set, y' i represents the label vector of the corrected remote sensing image in the i-th training sample set, Represents y' i The transpose of Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, (1-y' i ) T means (1-y' i ), Represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the i-th training sample.
[0072] In another exemplary embodiment of the present application, two consistency constraints are introduced into the objective function to guide the network to learn more discriminative features and reduce the negative impact of noise labels in the BCELoss of the objective function, that is, supervised training of the network is performed by combining supervised binary cross entropy loss with internal consistency loss and structural consistency loss. The loss function formula is: L = L sup +C str +C int .
[0073] In practical applications, the loss function is combined with the Adam optimizer algorithm during training, the learning rate is set to 0.001, and the model parameters are updated by minimizing the loss function until the model converges.
[0074] Combine the following Figure 3 The data processing process of the sample model provided in this application is illustrated by way of example:
[0075] like Figure 3 As shown, the training image (remote sensing image) x is obtained. i , strong transformation and weak transformation are performed on the same training image. Specifically, the training image is randomly cropped, randomly flipped, rotated and normalized. The input image size of the model is adjusted to 3×256×256, where 3 represents the three channels of the sample; the strong transformation operation adjusts the brightness, contrast and saturation of the image, and the adjustment probability is set to 0.8 and the adjustment intensity is set to {0.4, 0.4, 0.4} respectively. After preprocessing, the order of the training set sample images is randomly disrupted and input into the multi-label remote sensing image classification model. Deep feature extraction is performed through the feature encoder, and the convolution block of the ResNet18 network is used as the backbone framework. The input dimension size of this module is 3×224×224, and the output dimension size is 512×1×1. Classifiers and multi-layer perceptrons are used for further feature processing, where the classifier consists of a flattening layer and a fully connected layer. The multi-layer perceptron is a three-layer perceptron that nonlinearizes deep features, mainly consisting of a linear layer, a ReLU activation layer and a linear layer. This application can maintain a high classification accuracy even when the noise accounts for 50%, and the model runs efficiently, which can simultaneously meet the requirements of speed and accuracy in real applications.
[0076] This application combines two data enhancement methods of different strengths (weak transformation and strong transformation) to generate diverse training data so that the network can learn different image features of the same image; a collaborative learning model is adopted in feature extraction, and two similarity constraints are used to train two enhanced images; finally, a sorting loss is introduced to select and correct the noise labels. The final loss function uses the commonly used multi-label binary cross entropy loss supervision and is weighted with two similarity constraints to guide the model to perform robust learning. Not only can the end-to-end sample model be used, deep feature networks are used to extract deep features, and noisy labels are corrected as much as possible in complex noise label training to improve the performance of multi-label classification of remote sensing images in weakly supervised environments, but it has excellent classification accuracy and generalization ability in practical applications.
[0077] Based on the same inventive concept, the embodiment of the present application also provides a remote sensing image multi-label classification device for implementing the remote sensing image multi-label classification method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more remote sensing image multi-label classification device embodiments provided below can refer to the limitations of the remote sensing image multi-label classification method above, and will not be repeated here.
[0078] In an exemplary embodiment, a remote sensing image multi-label classification device is provided, comprising:
[0079] The acquisition module is used to acquire a training sample set, wherein the training sample set includes multiple training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; the label vector of the remote sensing image is obtained based on a clean label and a noise label representing a remote sensing image category.
[0080] The data enhancement module is used to process the remote sensing image in any training sample in the training sample set by using weak transformation and strong transformation respectively to obtain the weak transformation image and strong transformation image corresponding to the remote sensing image.
[0081] The first branch calculation module is used to input the weakly transformed image corresponding to the remote sensing image into the first branch of the sample model at the previous training times under the current training times to obtain the output vector of the classifier in the first branch under the current training times and the implicit features of the weakly transformed image corresponding to the remote sensing image under the current training times; and obtain the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times according to the output vector of the classifier in the first branch under the current training times. The sample model includes: a first branch and a second branch arranged in parallel; the first branch and the second branch both include a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model includes a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch.
[0082] The second branch calculation module is used to input the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; and obtain the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times according to the output vector of the classifier in the second branch at the current training times.
[0083] The correction module is used to correct the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number.
[0084] The loss function calculation module is used to calculate the value of the loss function at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number.
[0085] The training module is used to train the sample model under the previous training times according to the value of the loss function under the current training times to obtain the sample model under the current training times, and to determine whether the set training times have been reached. If the set training times have not been reached, the next training times will be entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model will be used to perform multi-label classification on the target remote sensing image.
[0086] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store multi-label classification data of remote sensing images. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a multi-label classification method for remote sensing images is implemented.
[0087] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the above-mentioned method embodiments when executing the computer program.
[0088] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.
[0089] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the above-mentioned method embodiments are implemented.
[0090] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0091] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0092] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.
[0093] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0094] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A multi-label classification method for remote sensing images, It is characterized in that The remote sensing image multi-label classification method comprises: Acquire a training sample set, the training sample set includes a plurality of training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; the label vector of the remote sensing image is obtained according to a clean label and a noise label representing a category of the remote sensing image; For a remote sensing image in any training sample in the training sample set, the remote sensing image is processed by using a weak transformation and a strong transformation respectively to obtain a weak transformation image and a strong transformation image corresponding to the remote sensing image; Under the current training times, the weakly transformed image corresponding to the remote sensing image is input into the first branch of the sample model under the previous training times to obtain the output vector of the classifier in the first branch under the current training times and the implicit features of the weakly transformed image corresponding to the remote sensing image under the current training times; the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times is obtained according to the output vector of the classifier in the first branch under the current training times; the sample model comprises: a first branch and a second branch arranged in parallel; the first branch and the second branch both comprise a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model comprises a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch; Inputting the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; obtaining the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times according to the output vector of the classifier in the second branch at the current training times; Correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number; Calculate the value of the loss function at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number; According to the value of the loss function under the current training times, the sample model under the previous training times is trained to obtain the sample model under the current training times, and it is judged whether the set training times are reached. If the set training times are not reached, the next training times are entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model is used to perform multi-label classification on the target remote sensing image.
2. The remote sensing image multi-label classification method according to claim 1, It is characterized in that The step of correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number specifically includes: Constructing a positive label set and a negative label set of the remote sensing image; the positive label set includes the real category of the remote sensing image; the negative label set includes the category that the remote sensing image does not belong to; Obtaining, based on the output vector of the classifier in the first branch at the current number of training times, the posterior probability of the classifier in the first branch for each positive label in the positive label set and the posterior probability of the classifier in the first branch for each negative label in the negative label set; For any positive label in the positive label set, the probability that the positive label is a pseudo-positive label is obtained from the perspective of the classifier in the first branch according to the posterior probability of the positive label by the classifier in the first branch and the posterior probability of each negative label in the negative label set by the classifier in the first branch; For any negative label in the negative label set, the probability that the negative label is a pseudo-negative label is obtained from the perspective of the classifier in the first branch according to the posterior probability of the negative label by the classifier in the first branch and the posterior probability of each positive label in the positive label set by the classifier in the first branch; Obtaining, based on the output vector of the classifier in the second branch at the current number of training times, the posterior probability of the classifier in the second branch for each positive label in the positive label set and the posterior probability of the classifier in the second branch for each negative label in the negative label set; For any positive label in the positive label set, the probability that the positive label is a pseudo-positive label is obtained from the perspective of the classifier in the second branch according to the posterior probability of the positive label by the classifier in the second branch and the posterior probability of each negative label in the negative label set by the classifier in the second branch; For any negative label in the negative label set, the probability that the negative label is a pseudo-negative label is obtained from the perspective of the classifier in the second branch according to the posterior probability of the negative label by the classifier in the second branch and the posterior probability of each positive label in the positive label set by the classifier in the second branch; Determine the probability that each positive label in the positive label set is a pseudo positive label from the perspective of the classifier in the first branch, and the positive label corresponding to the maximum probability among the probabilities that each positive label in the positive label set is a pseudo positive label from the perspective of the classifier in the second branch as the maximum positive label; Determine the probability that each negative label in the negative label set is a pseudo-negative label from the perspective of the classifier in the first branch, and the negative label corresponding to the maximum probability among the probabilities that each negative label in the negative label set is a pseudo-negative label from the perspective of the classifier in the second branch as the maximum negative label; The value corresponding to the maximum positive label and the value corresponding to the maximum negative label in the label vector of the remote sensing image are corrected to obtain the corrected label vector of the remote sensing image under the current training times.
3. The remote sensing image multi-label classification method according to claim 1, It is characterized in that The value of the loss function under the current training times is calculated according to the corrected label vector of each remote sensing image in the training sample set under the current training times, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set under the current training times, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set under the current training times, specifically including: Calculate the internal consistency function value at the current training times according to the implicit features of the weakly transformed images corresponding to each remote sensing image in the training sample set at the current training times and the implicit features of the strongly transformed images corresponding to each remote sensing image in the training sample set at the current training times; Calculating the structural consistency function value at the current training number according to the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number; Calculate the supervised loss function value at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number; The value of the loss function at the current number of training times is calculated based on the internal consistency function value at the current number of training times, the structural consistency function value at the current number of training times, and the supervised loss function value at the current number of training times.
4. The remote sensing image multi-label classification method according to claim 3, It is characterized in that The internal consistency function value under the current training times is calculated according to the implicit features of the weakly transformed images corresponding to each remote sensing image in the training sample set under the current training times and the implicit features of the strongly transformed images corresponding to each remote sensing image in the training sample set under the current training times, specifically: According to the formula Calculate the internal consistency function value, where C int represents the internal consistency function value, N represents the total number of training samples in the training sample set, represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the jth training sample, τ represents the temperature parameter, exp() represents the exponential function with the natural constant e as the base, express The transpose of .
5. The remote sensing image multi-label classification method according to claim 3, It is characterized in that The calculation of the structural consistency function value under the current training times according to the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set under the current training times and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set under the current training times is specifically: According to the formula Calculate the structural consistency function value, where C str represents the structural consistency function value, express and The Euclidean distance between them, μ represents the mean of sample x, σ represents the standard deviation of sample x, express and The Euclidean distance between express and The Euclidean distance between express and The Euclidean distance between represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the weakly transformed image corresponding to the remote sensing image in the jth training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, represents the implicit features of the strongly transformed image corresponding to the remote sensing image in the jth training sample, Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the jth training sample, represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the i-th training sample, Represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the jth training sample.
6. The remote sensing image multi-label classification method according to claim 3, It is characterized in that The supervised loss function value under the current training times is calculated according to the corrected label vector of each remote sensing image in the training sample set under the current training times, the predicted label vector of the weakly transformed image corresponding to each remote sensing image in the training sample set under the current training times, and the predicted label vector of the strongly transformed image corresponding to each remote sensing image in the training sample set under the current training times, specifically: Among them, L sup represents the supervised loss function value, N represents the total number of remote sensing images in the training sample set, y' i Represents the label vector of the corrected remote sensing image in the i-th training sample, y' i T Represents y' i The transpose of Represents the predicted label vector of the weakly transformed image corresponding to the remote sensing image in the i-th training sample, (1-y' i ) T means (1-y' i ), Represents the predicted label vector of the strongly transformed image corresponding to the remote sensing image in the i-th training sample.
7. A remote sensing image multi-label classification device, It is characterized in that The remote sensing image multi-label classification device comprises: An acquisition module is used to acquire a training sample set, wherein the training sample set includes a plurality of training samples, each training sample includes a remote sensing image and a label vector of the remote sensing image; the label vector of the remote sensing image is obtained according to a clean label and a noise label representing a category of the remote sensing image; A data enhancement module, for processing a remote sensing image in any one of the training samples in the training sample set by using a weak transformation and a strong transformation respectively, to obtain a weak transformation image and a strong transformation image corresponding to the remote sensing image; The first branch calculation module is used to input the weakly transformed image corresponding to the remote sensing image into the first branch of the sample model under the previous training times under the current training times to obtain the output vector and implicit features of the classifier in the first branch under the current training times; obtain the predicted label vector of the weakly transformed image corresponding to the remote sensing image under the current training times according to the output vector of the classifier in the first branch under the current training times; the sample model includes: a first branch and a second branch set in parallel; the first branch and the second branch both include a multi-label remote sensing image classification model and a multi-layer perceptron; the multi-label remote sensing image classification model includes a feature encoder and a classifier connected in sequence; the multi-layer perceptron in the first branch is connected to the output end of the feature encoder in the first branch; the multi-layer perceptron in the second branch is connected to the output end of the feature encoder in the second branch; The second branch calculation module is used to input the strongly transformed image corresponding to the remote sensing image into the second branch of the sample model at the previous training times to obtain the output vector and implicit features of the classifier in the second branch at the current training times; and obtain the predicted label vector of the strongly transformed image corresponding to the remote sensing image at the current training times according to the output vector of the classifier in the second branch at the current training times; A correction module, used for correcting the label vector of the remote sensing image according to the output vector of the classifier in the first branch at the current training number and the output vector of the classifier in the second branch at the current training number to obtain the corrected label vector of the remote sensing image at the current training number; A loss function calculation module is used to calculate the value of the loss function at the current training number according to the corrected label vector of each remote sensing image in the training sample set at the current training number, the predicted label vector and implicit features of the weakly transformed image corresponding to each remote sensing image in the training sample set at the current training number, and the predicted label vector and implicit features of the strongly transformed image corresponding to each remote sensing image in the training sample set at the current training number; The training module is used to train the sample model under the previous training times according to the value of the loss function under the current training times to obtain the sample model under the current training times, and to determine whether the set training times have been reached. If the set training times have not been reached, the next training times will be entered until the set training times are reached to obtain the trained sample model, and the multi-label remote sensing image classification model in the first branch or the second branch of the trained sample model will be used to perform multi-label classification on the target remote sensing image.
8. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the remote sensing image multi-label classification method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the remote sensing image multi-label classification method described in any one of claims 1 to 6 is implemented.
10. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the remote sensing image multi-label classification method described in any one of claims 1 to 6 is implemented.
Citation Information
Patent Citations
Method of and system for joint data augmentation and classification learning
CA3070817A1
Model training method and device and storage medium
CN110163234A
Self-supervision field adaptive deep learning method based on consistency training
CN111144565A
Remote sensing image weak supervision target detection method for optimizing feature extraction
CN116957051A
Remote sensing image cross-domain small sample classification method based on pseudo label uncertainty perception
CN117152503A
