Classification method and device for multi-label incremental learning of remote sensing images

Abstract

The present disclosure provides a classification method and device for multi-label incremental learning of remote sensing images, comprising receiving a remote sensing image and a text description of the remote sensing image; extracting a first feature vector of the remote sensing image and a second feature vector of the text description; inputting the first feature vector into an existing session of a classification model to obtain a first classification of the remote sensing image; inputting the first feature vector of the remote sensing image into a preset number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; and fusing the first feature vector, the first classification and the second classification to obtain a classification prediction result of the image. The technical solution provided by the present disclosure continuously expands the inter-class decision boundary in the cosine space through an angle distance decision boundary refinement mechanism, effectively resisting boundary degradation under the condition of a small number of samples; and systematically corrects the gradient imbalance in the multi-label sparse scene through an asymmetric gradient rebalancing strategy, ensuring that the newly introduced classes obtain sufficient learning signals.

Description

Technical Field

[0001] This disclosure relates to the field of remote sensing image classification, and more particularly to a classification method and apparatus for performing multi-label incremental learning on remote sensing images. Background Technology

[0002] This section is intended to provide background or context for the embodiments of this disclosure as set forth in the claims. The description herein is not intended to be a prior art simply because it is included in this section.

[0003] The continuous advancements in remote sensing satellites and airborne sensors have made remote sensing imagery a core data source for Earth observation, urban planning, and agricultural management. Unlike natural images, remote sensing scenes inherently possess multi-label characteristics—a single image typically contains multiple land cover categories such as buildings, vegetation, water bodies, and roads, with systematic co-occurrence relationships between these categories. As remote sensing technology continues to develop, its semantic label space is expanding openly. Urbanization, ecological evolution, and disaster events continuously introduce new land cover categories, and these new categories often only have a very small number of labeled samples. These characteristics dictate that remote sensing image recognition systems must possess the ability to gradually integrate new categories with a small number of samples without retraining on historical data—that is, multi-label, small-sample incremental learning capability.

[0004] Existing incremental learning methods for classes have the following main limitations: First, most small-sample incremental learning methods are structurally incompatible with multi-label scenarios; second, existing multi-label incremental learning methods are all designed under the condition of full data, and there is no method specifically for the constraint that each class has only a very small number of samples; third, when the prototype of the new class is estimated by only a small number of samples in existing technical solutions, the inter-class decision boundary is very easy to degenerate; fourth, existing technical solutions do not design a special gradient balancing mechanism for the extreme label sparsity of remote sensing multi-label scenarios, which causes the model to continuously favor negative class predictions and makes it difficult to reliably identify newly introduced classes.

[0005] In summary, existing technologies lack a few-sample incremental learning method that can simultaneously address the problems of decision boundary degradation and severe gradient imbalance in multi-label open-world scenarios of remote sensing images. Summary of the Invention

[0006] In view of this, the purpose of this disclosure is to propose a classification technology for multi-label incremental learning of remote sensing images, which at least partially solves one of the technical problems in related technologies.

[0007] To achieve the above objectives, an exemplary embodiment of this disclosure provides a classification method for performing multi-label incremental learning on remote sensing images, comprising:

[0008] The system receives a remote sensing image and a text description of the remote sensing image; extracts a first feature vector from the remote sensing image and a second feature vector from the text description; inputs the first feature vector into an existing session of a classification model to obtain a first classification of the remote sensing image; inputs the first feature vector of the remote sensing image into a predetermined number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; wherein the incremental learning sessions are used for refined few-sample classification of the remote sensing image; each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training; and fuses the first feature vector, the first classification, and the second classification to obtain the classification prediction result of the image.

[0009] Based on the same inventive concept, a second aspect of the exemplary embodiments of this disclosure provides a classification apparatus for performing multi-label incremental learning on remote sensing images, comprising: A receiving module configured to receive remote sensing images and text descriptions of the remote sensing images; a vector extraction module configured to extract a first feature vector of the remote sensing images and a second feature vector of the text descriptions; a first classification module configured to input the first feature vectors into an existing session of a classification model to obtain a first classification of the remote sensing images; a second classification module configured to input the first feature vectors of the remote sensing images into a predetermined number of incremental learning sessions of a classification model to obtain a second classification of the remote sensing images; wherein the incremental learning sessions are used for few-sample classification of the remote sensing images; each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training; and a result output module configured to fuse the first feature vectors, the first classification, and the second classification to obtain a classification prediction result for the image.

[0010] Based on the same inventive concept, a third aspect of the exemplary embodiments of this disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in the first aspect.

[0011] Based on the same inventive concept, a fourth aspect of the exemplary embodiments of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method as described in the first aspect.

[0012] Based on the same inventive concept, a fifth aspect of the exemplary embodiments of this disclosure provides a computer program product including computer program instructions that, when run on a computer, cause the computer to perform the method as described in the first aspect.

[0013] As can be seen from the above, the technical solution provided by this disclosure completely eliminates catastrophic forgetting at the architectural level by freezing the backbone network and the session isolation classification head, without the need for replaying historical data; it continuously expands the inter-class decision boundary in cosine space through the angle spacing decision boundary refinement mechanism, effectively resisting boundary degradation under conditions of few samples; it systematically corrects the gradient imbalance in multi-label sparse scenarios through the asymmetric gradient rebalancing strategy, ensuring that newly introduced categories obtain sufficient learning signals; and the synergistic effect of each module forms a complete technical system for open-world remote sensing scenarios, with good engineering deployment adaptability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of a classification method for multi-label incremental learning of remote sensing images according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram of the training process of an incremental learning session according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram of the process for calculating the adaptive cosine interval loss value for classification prediction in an embodiment of this disclosure. Figure 4 This is a schematic diagram of the process for calculating the scalar loss value of classification prediction according to an embodiment of this disclosure; Figure 5 This is a schematic diagram of a classification device for performing multi-label incremental learning on remote sensing images according to an embodiment of the present disclosure; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0016] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0017] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this application's technical solution, based on the prompt message.

[0018] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0019] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0020] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0021] To make the objectives, technical solutions, and advantages of this disclosure clearer, the principles and spirit of this disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0022] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0023] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar words used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly. The article "a" or "an" preceding an element does not exclude the existence of multiple such elements.

[0024] The principles and spirit of this disclosure will be explained in detail below with reference to several representative embodiments. Invention Overview As described in the background section, with the continuous development of remote sensing technology, its semantic label space is expanding openly. Urbanization, ecological evolution, and disaster events continuously introduce new land cover categories, and these new categories often only have a very small number of labeled samples. These characteristics dictate that remote sensing image recognition systems must have the ability to gradually integrate new categories with a small number of samples without retraining on historical data, i.e., multi-label small-sample incremental learning capability.

[0026] The relevant technologies have the following limitations: most small-sample incremental learning methods are based on the single-label classification assumption and use a softmax classifier to handle mutually exclusive categories, which is structurally incompatible with multi-label scenarios; existing multi-label incremental learning methods are all designed under the condition of full data, and there is no method specifically for the constraint that each class has only a very small number of samples; and the high inter-class visual similarity unique to remote sensing images (such as spectral confusion between bare soil and dry grass, and texture similarity between building roofs and bare ground) and extreme positive and negative sample imbalance jointly compress the decision space—when the prototype of a new category is estimated by only a small number of samples, the inter-class decision boundary is very easy to degenerate; existing methods have not designed a special gradient equalization mechanism for the extreme label sparsity of remote sensing multi-label scenarios, and the gradient signal generated by a large number of simple negative class samples masks the effective learning signal of a few positive class samples, causing the model to continuously favor negative class prediction and making it difficult to reliably identify newly introduced categories.

[0027] To address the aforementioned issues, this disclosure provides a classification method for multi-label incremental learning of remote sensing images. Specifically, it includes receiving a remote sensing image and a text description of the image; extracting a first feature vector from the remote sensing image and a second feature vector from the text description; inputting the first feature vector into an existing session of a classification model to obtain a first classification of the remote sensing image; inputting the first feature vector into a predetermined number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; wherein the incremental learning sessions are used for refined few-sample classification of the remote sensing image; each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training; and fusing the first feature vector, the first classification, and the second classification to obtain a classification prediction result for the image. By freezing the backbone network and isolating the classification heads for sessions, catastrophic forgetting is completely eliminated at the architectural level, eliminating the need for replaying historical data. The inter-class decision boundaries are continuously expanded in cosine space through the angle spacing decision boundary refinement mechanism, effectively resisting boundary degradation under conditions of few samples. The gradient imbalance in multi-label sparse scenarios is systematically corrected through an asymmetric gradient rebalancing strategy, ensuring that newly introduced categories receive sufficient learning signals. The synergistic effect of each module forms a complete technical system for open-world remote sensing scenarios, with good adaptability to engineering deployment.

[0028] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0029] Exemplary methods In the exemplary method of this application, the support set can be a small-shot incremental learning core training sample set, containing only a small number of labeled remote sensing image samples, used to initialize new category prototypes and train incremental session classification heads. The query set can be an incremental task undecided sample set, much larger than the support set, used to verify the category features learned by the model from the support set. The validation set can be an independent, gradient-free update dataset, which does not participate in weight training throughout the process, and is only used for parameter tuning, threshold search, monitoring overfitting, and selecting the optimal model.

[0030] specifically refer to Figure 1 In some embodiments, this disclosure includes a classification method for performing multi-label incremental learning on remote sensing images, comprising: Step 101: Receive the remote sensing image and its text description. Specifically, the remote sensing image refers to a digital image formed by remote sensing platforms such as satellites, drones, and aircraft, equipped with optical, radar, and multispectral sensors, capturing electromagnetic wave reflection and radiation information from the Earth's surface at a distance without direct contact with ground targets, and then processing it through photoelectric conversion and transmission. The text description of the remote sensing image can be achieved through manual annotation or large-scale model annotation.

[0031] Step 102: Extract the first feature vector of the remote sensing image and the second feature vector of the text description. Specifically, this step can be achieved by loading a CLIP visual encoder pre-trained with large-scale image-text pairs. Specifically, all encoder parameters are set to an untrainable state and frozen throughout the entire incremental process. Visual features are extracted from each remote sensing image in the batch and L2 normalized to obtain the first feature vector, referred to by those skilled in the art as the normalized visual feature vector. The second feature vector, referred to by those skilled in the art as the categorical text prototype vector, can be obtained by encoding with the CLIP text encoder and then performing L2 normalization. The CLIP visual encoder can employ a CLIP visual encoder with a ViT-L / 14 architecture.

[0032] It is worth noting that those skilled in the art will understand that the method for extracting the first feature vector of a remote sensing image and the second feature vector of the text description of the remote sensing image is not limited to using a CLIP visual encoder or a CLIP text encoder. In fact, those skilled in the art can use any suitable encoder to implement the above technical solution. Further details are omitted here.

[0033] Step 103: Input the first feature vector into the existing session of the classification model to obtain the first classification of the remote sensing image. Here, all parameters of the existing session are frozen. Specifically, all trainable parameters of the existing session are frozen. These may include the attention weights, feedforward network weights and biases of the session backbone encoder; or the context aggregation and temporal fusion network parameters of the session layer; or the linear layer weights, biases, and normalized scaling offset parameters of the classification head / task head.

[0034] In step 103, the classification head refers to the output layer at the end of the model responsible for the classification task. In one possible embodiment, the classification head receives the global session features output by the CLIP visual encoder, maps the features to a preset category space through linear mapping, and then outputs the probability of each category through an activation function. It is used for session intent recognition, scene classification, and topic categorization, only performing feature discrimination and category division, and does not participate in backbone feature extraction. It is a dedicated output structure adapted to downstream classification tasks.

[0035] Step 104: Input the first feature vector of the remote sensing image into a preset number of incremental learning sessions of the classification model to obtain the second classification of the remote sensing image; wherein, each incremental learning session is used for refined few-sample classification of the remote sensing image; each incremental learning session uses the angle spacing decision boundary refinement function and the asymmetric gradient rebalancing loss function for parameter iterative training.

[0036] In step 105, the first feature vector, the first classification, and the second classification are fused to obtain the image classification prediction result. Specifically, the fusion method can be optimal threshold comparison by category, and outputting the final multi-label prediction label set based on the category comparison. For any input remote sensing image, after the first feature vector is extracted by the CLIP visual encoder with parameters frozen in step 102, it can be sequentially passed through the classification heads of all sessions, and the classification data (which may be referred to as logical values ​​by those skilled in the art) output by each classification head can be concatenated to obtain the joint prediction vector. After applying sigmoid activation to the vector, it is compared category-by-category with the optimal threshold calculated based on the prior frequency of the categories, and the final classification prediction result of the remote sensing image is output. The computational cost of the inference process included in this disclosed technical solution is only proportional to the total number of seen categories, thus exhibiting good scalability.

[0037] In step 105, the activation of the sigmoid function can use a non-linear activation function to compress and map any real-number input to the 0-1 interval. Its curve is smooth and monotonically increasing, with a large gradient in the middle and tending to saturate at both ends. It is commonly used for binary classification probability output and gating threshold control, and can introduce non-linear fitting of complex relationships. It can also be applied to the normalization of model classification head probabilities.

[0038] Specifically, regarding the incremental learning sessions included in this disclosure, when all existing sessions in the classification model fail to accurately classify remote sensing images, an incremental learning session can be added to the classification model and iteratively trained. Specifically, when the first feature vector is input into an existing session of the classification model to obtain the first classification of the remote sensing image, if the confidence scores of all first classifications are below a preset third threshold, it indicates that the existing sessions cannot accurately classify the remote sensing image. In response to this instruction, the classification model establishes an incremental learning session. Here, the confidence score is a quantification value of the reliability of the classification model's own prediction results, ranging from 0 to 1. The classification model of this disclosure can output the probabilities of its respective categories using the Softmax function and the Sigmoid function; the higher the probability, the stronger the confidence score. The Softmax function is a multi-classification activation function that can convert a set of arbitrary real-valued scores into a probability distribution with a sum of 1, where each output value represents the confidence score of the corresponding category. This function can highlight the optimal category and can be used in conjunction with ordinary cross-entropy loss, making it suitable for label classification tasks. The Sigmoid function is an activation function that compresses and maps any real-valued input to a range between 0 and 1. Its curve is smooth and monotonically increasing, and the output can be interpreted as class probabilities. It can be used for label classification tasks, with each class output independently.

[0039] The predetermined number of incremental learning sessions included in this disclosure can be iteratively trained to train their parameters to meet the accuracy requirements needed for classification during subsequent inference. See specifically... Figure 2 In one possible implementation, the parameter iterative training of a predetermined number of Nth incremental learning sessions includes: Step 201 initializes the training parameters for the incremental learning session. Specifically, the training parameters in step 201 can be trainable parameters in the classification head of the incremental learning session. In a specific embodiment, the second feature vector can be used as the initial value of the new classification head weight matrix for the initial incremental learning session, thereby injecting CLIP text-visual alignment semantic information into the classification head. Specifically, when starting the new category learning process for the multi-label incremental classification task of remote sensing images, a category support set corresponding to the incremental task is first constructed. This set consists of a small number of remote sensing images with labeled surface features and scene types, along with matching text descriptions. Throughout the process of initializing category-specific bias terms and initializing the classification head weight matrix with text prototypes, the support set is used to complete category prior frequency statistics, similar feature aggregation, and new category semantic anchoring. All initial parameter calibrations and small-sample feature learning for the classification head in the incremental learning sessions use the support set as the sole source of reference samples.

[0040] Specifically, the above embodiments can be implemented by introducing a new set of categories in the current session, constructing a text description for each category in the current session using a prompt template, encoding it with a CLIP text encoder and then performing L2 normalization to obtain a category text prototype vector, which serves as the initial value of the weight matrix of the classification head of the initial incremental learning session, thereby injecting CLIP text-visual aligned semantic information into the classification head.

[0041] The category-specific bias term is initialized by statistically analyzing the measured positive sample frequency πj for each category based on the support set of the current session. A positive sample is a sample that matches the target category and belongs to the expected matching object. In classification tasks, a sample with a label of 1 can be defined as a positive sample. The classification model in this disclosure can learn feature representations based on positive samples during training, narrowing the distance between similar classes. This is the core foundation for constructing various loss functions and the basis for classification head discrimination. The initialized category-specific bias term can be expressed as formula (1): (1) in, The category-specific bias term representing the j-th category. This represents the prior frequency of positive samples of the j-th class in the training support set. In the technical solution disclosed herein, the class-specific bias term can refer to the class-by-class independently adjustable bias parameter within the output layer of the classification head, and the positive sample prior frequency can refer to the proportion of positive samples that exist in the entire sample set in the training data, which belongs to the model's preset prior information. This represents the probability of positive events naturally occurring in the real-world scenario, used to balance the bias in the distribution of positive and negative samples.

[0042] Initializing the category-specific biases ensures that the sigmoid function output of the classification head is close to the prior frequencies of each class before training begins in the initial incremental learning session, providing a calibrated starting state. Simultaneously, all parameters of existing session classification heads are frozen, allowing only the weights and biases of the incremental learning session's classification head to participate in subsequent gradient updates. The support set serves as a small set of samples for training the incremental learning session's classification head, acting as a reference and template. The model first learns the features in the support set, treating them as class prototypes; then, it compares the query set samples with the support set for classification. The category-specific biases are unique bias parameters for each class in the classification head, used to fine-tune the class's base score, compensating for class prior differences and sample imbalance. This improves the classification head's ability to distinguish rare classes and difficult-to-classify samples.

[0043] Step 202: Input the first feature vector into the Nth incremental learning session to obtain the encoded first feature vector.

[0044] Step 203 involves inputting the encoded first feature vector into the angular margin decision boundary refinement function to obtain the adaptive cosine margin loss value for classification prediction. Specifically, after completing the incremental learning session based on the support set prototype learning, a query set is introduced as the main training input sample for incremental model training. A batch of remote sensing images from the query set are input into the frozen backbone network to extract the first feature vector. Cosine similarity calculation, adaptive angular margin solution, and classification logic value deduction are performed based on the query set samples. The model achieves the training objective of increasing the angular margin between dissimilar classes and clustering similar classes by comparing the class feature prototypes learned from the support set with the feature distribution of the query set samples. All gradient iteration updates use the query set as the training driving sample. In this disclosed technical solution, the angular margin decision boundary refinement is an incremental classification head based on cosine similarity, used to dynamically adjust the decision boundary position of each positive sample. The angular margin decision boundary refinement function dynamically adjusts the decision boundary position of each positive sample during training through adaptive angular margin. See details below. Figure 3 In one possible implementation, the adaptive cosine margin loss value for classification prediction output by the classification prediction input angular spacing decision boundary refinement function includes: Step 301: Calculate the cosine similarity between the first feature vector and the parameter weights of the incremental learning session. Specifically, the cosine similarity between the first feature vector and the few-sample classification vector can be calculated by dividing the vector dot product by the product of the magnitudes of the two vectors.

[0045] And in step 302, based on the absolute value of the difference between the cosine similarity and the preset third threshold, the decision boundary refinement function calculates the adaptive cosine margin loss value for classification prediction based on the basic distance strength of the training samples and the adaptive exponent. After obtaining the cosine similarity between the first feature vector of the training samples and the small sample classification vector and the adaptive cosine margin loss value for classification prediction, the magnitude of the adaptive cosine margin loss value for classification prediction output by the angle-based distance decision boundary refinement function is determined based on whether the cosine similarity meets the preset threshold.

[0046] In one specific embodiment, during forward propagation training, the normalized visual features f of the batch samples are multiplied by the L2-normalized class weight vector of the class prototype to obtain the cosine similarity score of the positive class label within the batch. For the cosine similarity scores of positive class labels within a batch, the adaptive angular spacing of the samples is calculated; the formula for calculating the adaptive angular spacing can be expressed as formula (2): (2) in, Representing the The adaptive angle interval margin corresponding to each sample. This represents the cosine similarity between the normalized visual features and the corresponding category weight vector. This represents the adaptive interval decay index. This represents the hyperparameter of the foundation spacing strength.

[0047] For negative tags, the spacing penalty is zero.

[0048] The final classification logic value after spacing adjustment is , where s is a learnable temperature parameter that is updated via gradient descent.

[0049] In the inference phase after the classification model has been trained, directly... The classification logical value is calculated without imposing a margin penalty, preserving the standard cosine metric structure of the embedding space. Here, the inner product can refer to the method of multiplying the corresponding values ​​of two vectors of the same dimension and then summing them. This method can be used to measure the similarity and angular relationship between two vectors: that is, the larger the inner product, the closer the vector directions are; the inner product is 0 when they are orthogonal. This is a fundamental core operation for semantic retrieval and vector comparison. The specific implementation method of this step is detailed below.

[0050] Step 204 involves inputting the adaptive cosine margin loss value into the asymmetric gradient rebalancing loss function to obtain the scalar loss value for classification prediction. Specifically, the classification prediction output from step 204 is activated by a sigmoid function to obtain the prediction probability of historical data. .in, The activation function is Sigmoid. In the technical solution disclosed herein, the asymmetric gradient rebalancing loss function is a composite loss function designed for multi-label incremental learning scenarios. It integrates three independent mechanisms for scoring each training iteration of the incremental learning session, each addressing a specific difficulty in multi-label incremental learning. The three mechanisms are calculated simultaneously and act element-wise in a single forward propagation, jointly producing the final scalar loss value. See details... Figure 4 In one possible implementation, inputting the classification prediction into the asymmetric gradient rebalancing loss function to obtain the scalar loss value for the classification prediction includes: Step 401: Calculate the sigmoid probability of the predicted classification using the sigmoid function and split the sigmoid probability into the positive class prediction probability and the negative class prediction probability.

[0051] Step 402: Calculate the probability offset term for the predicted classification using the asymmetric gradient rebalancing loss function. The formula for calculating the probability offset can be Equation (3): (3) in, This represents the negative class probability truncation offset threshold.

[0052] Step 403: Calculate the asymmetric focus term for the predicted classification using the asymmetric gradient rebalancing loss function. In the asymmetric focus term, a first focus weight is set for the positive class label, and a second focus weight is set for the negative class label. The calculation formula for the asymmetric focus term can be Equation (4): (4) The focus weight of the positive class samples can be expressed as: The focus weight of the negative class sample can be expressed as pt = 1 - p Focus Index Values ​​are taken from positive class samples. Negative class sample values Among them, the relationship can be satisfied .

[0053] Step 404: Calculate the class prior reweighting term for the predicted classification based on the negative class prediction probability using the asymmetric gradient rebalancing loss function. The formula for calculating the prior reweighting coefficient can be Equation (5): (5) in, This is a temperature parameter. Used for positive class samples. Negative class samples use .

[0054] Step 405 involves calculating the scalar loss value by combining the probability shift term, the asymmetric focus term, and the prior reweighting term. Specifically, these three mechanisms can be combined using element-wise multiplication. (Scalar loss value) It can be expressed by formula (6): (6) in, For prior weighting of the class, This represents the logarithmic loss term of the basic binary cross-entropy.

[0055] The above three mechanisms are combined to form an asymmetric gradient rebalancing loss function. Specifically, in step 205, the parameters of the current session classification head can be updated using the AdamW optimizer.

[0056] The scalar loss indirectly affects the quality of parameter updates in the following ways: the three mechanisms used to calculate the scalar loss ensure that backpropagation in each training batch produces meaningful, non-zero gradients. If a standard binary cross-entropy loss is used, the weak gradients generated by easily separable negative samples will be overwhelmed by the average effect of a massive number of negative samples. Furthermore, adjusting the probability offset term can prevent negative classes from being properly classified. The absolute value of the term increases sharply, stabilizing the loss curve during training and avoiding gradient oscillations. Furthermore, the tuning of the asymmetric focus term can eliminate noisy gradients from a large number of easily separable negative samples, allowing the optimizer to focus more on difficult samples and accelerating effective convergence.

[0057] There is a crucial synergy between the adaptive cosine margin loss and the scalar loss. When using the scalar loss, because the sigmoid function is close to saturation within most of the cosine similarity range, the asymmetric focus term and cross-entropy gradient simultaneously disappear. Therefore, the scaling factor of the margin decision boundary refinement function solution is forcibly overridden to a smaller initial value for cosine similarity scaling, rather than the default initial value for scaling the abstract semantic representation task. The three mechanisms of the asymmetric gradient rebalancing loss function become ineffective: reducing the scaling factor to a value that keeps the sigmoid function in a mild region, allowing the gradient of the asymmetric gradient rebalancing loss function to flow normally, and ensuring that the margin mechanism of the asymmetric gradient rebalancing loss function operates in the cosine space before scaling, unaffected by the reduction in the scaling factor. The "mild region" can refer to the linear fitting interval of the sigmoid function, the numerical range within the input domain where the gradient is not saturated and can backpropagate normally.

[0058] And in step 205, when the scalar loss value is not less than a preset first threshold, the training parameters of the Nth incremental learning session are adjusted to obtain an intermediate incremental learning session. The intermediate incremental learning session is used to replace the Nth incremental learning session for iterative training until the number of training iterations reaches a preset second threshold, and the intermediate incremental learning session is output as the Nth incremental learning session. Specifically, the incremental learning session in the classification model of this disclosure does not dynamically determine whether to continue iterating based on the loss function output. The number of training rounds for each incremental task is fixedly determined by the preset second threshold set in the configuration file.

[0059] Where N is a natural number greater than or equal to 1 and less than a preset threshold.

[0060] In one possible implementation, the effectiveness of incremental learning in widening the angular distance between different categories can be quantitatively evaluated by calculating the angular distance matrix between the weight vectors of each category. A larger distance indicates a clearer decision boundary between categories. Specifically, this implementation involves comparing the weight parameters of the trained classifier head with preset reference weights to determine the angular difference between the outlier vectors, thus completing a quantitative evaluation of the incremental classification performance.

[0061] In one possible implementation, when the Nth incremental learning session is completed, the training of the N+1th incremental learning session begins; until all preset number of incremental learning sessions are trained.

[0062] In one possible implementation, after a preset number of incremental learning dialogues have been trained, a decision threshold search is performed on the second classification, including: Within a preset search range, the sigmoid probability of the second category output for each incremental learning session is calculated based on a preset discrete difference.

[0063] The sigmoid probability is binarized and predicted using each decision threshold obtained during the traversal training process, and the evaluation index of the corresponding binarized prediction result is calculated.

[0064] Select the optimal decision threshold based on the evaluation indicators.

[0065] Output the optimal threshold and freeze all trainable parameters for a preset number of incremental learning sessions.

[0066] Specifically, after each training session, a threshold grid search is performed on the sigmoid prediction probabilities of all seen categories, using either the training or validation set as a benchmark. Candidate thresholds are traversed within a preset range with a fixed step size, and the threshold maximizing the F1-Micro score is selected as the inference threshold. This disclosure's technical solution can employ a globally unified threshold mode (applying a single threshold to all categories) or a category-independent threshold mode (searching independently for each category to optimize the classification boundary). The globally unified threshold mode has lower computational overhead, while the category-independent threshold mode is more adaptable to imbalanced class scenarios. In this disclosure, the training set is the dataset used for learning parameters during the incremental learning session, and the validation set is an independent dataset that does not participate in parameter updates during the incremental learning session training. The validation set is used to evaluate the model's current performance, tune hyperparameters, select the optimal model, monitor overfitting, and determine whether to stop training early and adjust the learning rate. In the technical solution disclosed herein, the F1-Micro score can refer to a multi-class / multi-label evaluation metric. It first aggregates TP (True Positives), FP (False Positives), and FN (False Negatives) for all categories, then calculates the global precision and recall, and finally obtains the F1 score (the combined score of precision and recall). It emphasizes overall performance, is robust to class imbalance, and its value ranges from 0 to 1, with higher values ​​being better.

[0067] In one specific embodiment, in cosine space, the decision boundary without spacing adjustment is ,Right now After applying the external spacing, positive samples need to satisfy... Only samples that meet the criteria can be classified as positive. This is equivalent to changing the decision boundary for positive samples from... Tighten. The key to adaptive spacing lies in... Item: When That is, when the sample features are highly aligned with the class center, The effective value of the outer spacing is almost zero, imposing almost no additional constraints on the samples. When That is, when aligned in the middle, The effective value of the external spacing is approximately 0.028, slightly tightening the boundaries of the external spacing. In the above calculations, the parameters... This represents the adaptive interval dynamic adjustment coefficient. The larger the value, the slower the decay of the outer spacing of the low cosine region, and the stronger the push of the sample separation mechanism on the hard-to-separate samples. The smaller the value, the more concentrated the external spacing is in the extremely low cosine region, and the weaker the push of the sample separation mechanism on difficult-to-separate samples.

[0068] The relationship between threshold search and scalar loss / adaptive cosine interval loss is as follows: the spacing mechanism of adaptive cosine interval loss adjusts the angular decision boundary in the cosine space during the training phase, while threshold search adjusts the numerical decision boundary in the sigmoid probability space during the inference phase. They operate at different stages and in different spaces, complementing each other.

[0069] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0070] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0071] Exemplary device Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a classification device for performing multi-label incremental learning on remote sensing images.

[0072] refer to Figure 5 The classification device for performing multi-label incremental learning on remote sensing images includes: A receiving module configured to receive remote sensing images and text descriptions of the remote sensing images; a vector extraction module configured to extract a first feature vector of the remote sensing images and a second feature vector of the text descriptions; a first classification module configured to input the first feature vectors into an existing session of a classification model to obtain a first classification of the remote sensing images; a second classification module configured to input the first feature vectors of the remote sensing images into a predetermined number of incremental learning sessions of a classification model to obtain a second classification of the remote sensing images; wherein the incremental learning sessions are used for few-sample classification of the remote sensing images; each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training; and a result output module configured to fuse the first feature vectors, the first classification, and the second classification to obtain a classification prediction result for the image.

[0073] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0074] The apparatus of the above embodiments is used to implement the corresponding classification method for multi-label incremental learning of remote sensing images in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0075] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the classification method for multi-label incremental learning of remote sensing images as described in any of the above embodiments.

[0076] Figure 6 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0077] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0078] The memory 1020 can be implemented in the form of ROM (Read-Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0079] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0080] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0081] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0082] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0083] The electronic devices described in the above embodiments are used to implement the corresponding classification methods for multi-label incremental learning of remote sensing images in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0084] The memory 1020 stores machine-readable instructions executable by the processor 1010. When the electronic device is running, the processor 1010 communicates with the memory 1020 via the bus 1050, causing the processor 1010 to execute the following instructions during operation: Receive remote sensing images and their text descriptions; Extract the first feature vector from the remote sensing image and the second feature vector from the text description; Input the first feature vector into the existing session of the classification model to obtain the first classification of the remote sensing image; The first feature vector of the remote sensing image is input into a predetermined number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; wherein, each incremental learning session is used for refined few-sample classification of the remote sensing image; and each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training; and The first feature vector, the first classification, and the second classification are fused to obtain the classification prediction result of the image.

[0085] Through the above methods, the technical solution provided in this disclosure completely eliminates catastrophic forgetting at the architectural level by freezing the backbone network and the session isolation classification head, without the need for historical data replay; it continuously expands the inter-class decision boundary in cosine space through the angle spacing decision boundary refinement mechanism, effectively resisting boundary degradation under conditions of few samples; it systematically corrects the gradient imbalance in multi-label sparse scenarios through an asymmetric gradient rebalancing strategy, ensuring that newly introduced categories obtain sufficient learning signals; and the synergistic effect of each module forms a complete technical system for open-world remote sensing scenarios, with good engineering deployment adaptability.

[0086] Exemplary program product Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the classification method for multi-label incremental learning of remote sensing images as described in any of the above embodiments.

[0087] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0088] The aforementioned non-transitory computer-readable storage media can be any available medium or data storage device that a computer can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0089] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the classification method for multi-label incremental learning of remote sensing images as described in any of the embodiments in the exemplary method section above, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0090] Based on the same inventive concept, corresponding to the classification method for multi-label incremental learning of remote sensing images described in any of the above embodiments, this disclosure also provides a computer program product, which includes computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer to cause the computer and / or the processor to perform the classification method for multi-label incremental learning of remote sensing images. Corresponding to the execution entity for each step in each embodiment of the classification method for multi-label incremental learning of remote sensing images, the processor executing the corresponding step can belong to the corresponding execution entity.

[0091] The computer program product of the above embodiments is used to cause the computer and / or the processor to execute the classification method for multi-label incremental learning of remote sensing images as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0092] Those skilled in the art will recognize that embodiments of this disclosure can be implemented as a system, method, or computer program product. Therefore, this disclosure can be implemented as entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this disclosure can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0093] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (not exhaustive) of a computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0094] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0095] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0096] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0097] It should be understood that each block of a flowchart and / or block diagram, as well as combinations of blocks in a flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine that, when executed by a computer or other programmable data processing device, creates means for implementing the functions / operations specified in the blocks of the flowchart and / or block diagram.

[0098] These computer program instructions may also be stored in a computer-readable medium that enables a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce a product comprising an instruction apparatus that implements the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0099] Computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, such that the instructions that execute on the computer or other programmable apparatus can provide a process for implementing the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0100] Furthermore, although the operations of the methods of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be executed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0101] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0102] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0103] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0104] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0105] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0106] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

[0107] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. The scope of the appended claims is to be interpreted in the broadest sense, thereby encompassing all such modifications and equivalent structures and functions.

Claims

1. A classification method for multi-label incremental learning of remote sensing images, comprising: Receive remote sensing images and text descriptions of the remote sensing images; Extract the first feature vector of the remote sensing image and the second feature vector of the text description; The first feature vector is input into an existing session of the classification model to obtain a first classification of the remote sensing image; The first feature vector of the remote sensing image is input into a preset number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; wherein, each incremental learning session is used for refined few-sample classification of the remote sensing image; each incremental learning session uses the angle spacing decision boundary refinement function and the asymmetric gradient rebalancing loss function for parameter iterative training; as well as The first feature vector, the first classification, and the second classification are fused to obtain the classification prediction result of the image.

2. The method according to claim 1, wherein, The parameter iterative training of the Nth incremental learning session of the preset number includes: Initialize the training parameters for the Nth incremental learning session; The first feature vector is input into the Nth incremental learning session to obtain the encoded first feature vector; The encoded first feature vector is input into the angular spacing decision boundary refinement function to obtain the adaptive cosine interval loss value of the classification prediction. The adaptive cosine margin loss value is input into the asymmetric gradient rebalancing loss function to obtain the scalar loss value for the classification prediction; and When the scalar loss value is not less than a preset first threshold, the training parameters of the Nth incremental learning session are adjusted to obtain an intermediate incremental learning session. The intermediate incremental learning session is used to replace the Nth incremental learning session for iterative training until the number of training sessions reaches a preset second threshold. The intermediate incremental learning session is then output as the Nth incremental learning session. Where N is a natural number greater than or equal to 1 and less than a preset threshold.

3. The method according to claim 2, wherein, The first feature vector is input into the angular spacing decision boundary refinement function, and the adaptive cosine interval loss value for the classification prediction is output, including: Calculate the cosine similarity between the first feature vector and the parameter weights of the incremental learning session; and Based on the absolute value of the difference between the cosine similarity and the preset third threshold, the decision boundary refinement function calculates the adaptive cosine interval loss value of the classification prediction based on the basic distance strength of the training samples and the adaptive exponent.

4. The method according to claim 2, wherein, The adaptive cosine margin loss value is input into the asymmetric gradient rebalancing loss function to obtain the scalar loss value for the classification prediction, including: The binary probability of the adaptive cosine interval loss value is calculated using a binary probability mapping function, and the binary probability is split into positive class prediction probability and negative class prediction probability. Based on the negative class prediction probability, calculate the probability offset term of the predicted classification; Based on the positive class prediction probability and the negative class prediction probability, the asymmetric focus term of the predicted classification is calculated; wherein in the asymmetric focus term, a first focus index is set for the positive class label and a second focus index is set for the negative class label; Based on the ratio of positive to negative samples for each category, a class prior reweighting term for the predicted classification is calculated based on the predicted probability of the negative class; and The scalar loss value is obtained by calculating the probability offset term, the asymmetric focus term, and the class prior reweighting term.

5. The method according to any one of claims 2-4, wherein, Once the Nth incremental learning session is completed, training of the N+1th incremental learning session begins; this continues until all preset number of incremental learning sessions have been trained.

6. The method according to claim 5, wherein, After all the preset number of incremental learning dialogues have been trained, a decision threshold search is performed on the second classification, including: Within a preset search range, based on a preset discrete difference, the sigmoid probability of the second category output for each incremental learning session is calculated. The sigmoid probability is binarized and predicted using each decision threshold obtained during the training process. Calculate the evaluation index of the binarized prediction result; Select the optimal decision threshold based on the evaluation indicators; Output the optimal threshold and freeze all trainable parameters for the preset number of incremental learning sessions.

7. A classification device for performing multi-label incremental learning on remote sensing images, comprising: The receiving module is configured to receive remote sensing images and text descriptions of the remote sensing images; The vector extraction module is configured to extract a first feature vector of the remote sensing image and a second feature vector of the text description; The first classification module is configured to input the first feature vector into an existing session of the classification model to obtain a first classification of the remote sensing image; The second classification module is configured to input the first feature vector of the remote sensing image into a preset number of incremental learning sessions of the classification model to obtain a second classification of the remote sensing image; wherein the incremental learning sessions are used for few-sample classification of the remote sensing image; and each incremental learning session uses an angular spacing decision boundary refinement function and an asymmetric gradient rebalancing loss function for iterative parameter training. as well as The result output module is configured to fuse the second feature vector, the first classification, and the second classification to obtain the classification prediction result of the image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 6.

10. A computer program product comprising computer program instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images