A SAR image incremental small sample target detection system and method based on prototype contrast

Abstract

The application discloses a SAR image incremental small sample target detection system and method based on prototype contrast, and the system comprises a class prototype representation generation module, a mixed class prototype contrast coding module, a class prototype calibration module and a target detection module; the class prototype representation generation module is used for extracting a basic class RoI feature mean value of a historical SAR image as a class prototype based on a pre-training model; the mixed class prototype contrast coding module is used for designing a loss function in combination with a new class sample and the class prototype, forcibly aggregating the same class and separating different classes, and optimizing a feature space; the class prototype calibration module is used for measuring and minimizing distribution differences between the new class sample and the class prototype through a Gaussian kernel function, constraining representation deviation in an incremental stage, and obtaining an InFSAR model; and the target detection module is used for acquiring a SAR image and performing small sample target detection based on the InFSAR model.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing image processing technology, specifically relating to a SAR image incremental small sample target detection system and method based on prototype comparison. Background Technology

[0002] Existing SAR few-sample target detection methods require re-accessing old class data to maintain base class performance when incrementally learning new classes, resulting in high storage costs and privacy risks. Furthermore, the scarcity of new class data easily leads to overfitting, while the inaccessibility of base class data causes catastrophic forgetting. Currently, there is a lack of solutions that can maintain base class performance while efficiently adapting to new classes during incremental learning. For example, traditional methods such as TFA (Two-stage Fine-tuning Approach) can alleviate forgetting by freezing some parameters, but they cannot effectively utilize base class knowledge to optimize the representation of new classes, leading to classification confusion and performance degradation. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0004] A prototype-based incremental small-sample target detection system for SAR images includes: a prototype-based representation generation module, a hybrid prototype-based comparison encoding module, a prototype-based calibration module, and a target detection module.

[0005] The class prototype representation generation module is used to extract the mean value of basic class RoI features of historical SAR images as class prototypes based on the pre-trained model.

[0006] The hybrid class prototype comparison encoding module is used to combine the newly added class samples and the class prototype to design a loss function, force the aggregation of similar classes and the separation of dissimilar classes, and optimize the feature space.

[0007] The class prototype calibration module is used to measure and minimize the distribution difference between the newly added class samples and the class prototype through a Gaussian kernel function, constrain the representation offset in the incremental stage, and obtain the InFSAR model.

[0008] The target detection module is used to acquire SAR images and perform small-sample target detection based on the InFSAR model.

[0009] Preferably, the workflow of the class prototype representation generation module includes:

[0010] The Faster R-CNN object detection framework is adopted, with the backbone network being ResNet-101 combined with a feature pyramid network to construct the feature extraction model;

[0011] The historical SAR images of the ship are acquired, and the feature extraction model is trained using the historical SAR images to obtain the pre-trained model;

[0012] The pre-trained model is used to extract the basic RoI features of the historical SAR image, and the mean representation of the basic RoI features is calculated to obtain the class prototype:

[0013]

[0014] Where, p c Represents the class prototype, x c x represents the set of all samples in historical SAR images. i y represents the input feature vector. i Indicates category label, f θ This represents the feature extractor of the pre-trained model.

[0015] Preferably, the workflow of the hybrid prototype comparison and encoding module includes:

[0016] Input the K-shot samples of the newly added class, and obtain the RoI feature vector φ through the frozen feature extractor. i ;

[0017] Calculate the eigenvector φ i With the class prototype set P = {p1, ..., p c Cosine similarity between}

[0018]

[0019] Where sim represents the cosine similarity, φ j This represents the vector embedded in the hybrid prototype comparison encoding module;

[0020] Based on the cosine similarity, a loss function is designed by combining the newly added class samples and the class prototype.

[0021]

[0022]

[0023] Where M represents the number of labeled sample examples in the dataset. M represents the IoU score that matches its truth box, ε represents the threshold for the Intersection over Union (IoU), and M represents the IoU score. yi This indicates the number of samples of the same type in the current batch. Indicates the indicator function, y j Let p represent the feature of the j-th RoI, C represent the total number of categories, c represent the number of categories, and p represent the feature of the j-th RoI. kThis represents the prototype of class k, and τ represents the temperature coefficient.

[0024] Based on the loss function It forces similar samples to aggregate in the feature space, while separating the new class from the prototype of the base class, thus optimizing the feature space.

[0025] Preferably, the workflow of the prototype calibration module includes:

[0026] Given the distribution of class prototypes E={d j The distribution F = {f |j=1,…,C} and the newly added class samples i |i=1,…,Q}, where d j and f i These represent the probability distributions of the class prototype and the newly added class samples, respectively.

[0027] The distance between two distributions is quantified by calculating the maximum mean difference using a Gaussian kernel function.

[0028]

[0029] Where k(x,y) represents the Gaussian kernel function, σ represents the hyperparameter, MMD represents the maximum mean difference, m represents the number of samples in set F, and n represents the number of samples in set E;

[0030] By incorporating MMD as a regularization term into the total loss function, the new class prototype distribution is made to approximate the class prototype distribution, thereby constraining the representation shift during the incremental learning process, and thus obtaining the InFSAR model.

[0031] This invention also provides a method for incremental small-sample target detection in SAR images based on prototype comparison. The method, applied to any of the systems described above, includes the following steps:

[0032] S1. Extract the mean value of the basic class RoI features from historical SAR images based on the pre-trained model as the class prototype;

[0033] S2. Combine the newly added class samples and the class prototype to design a loss function, force the aggregation of similar classes and the separation of dissimilar classes, and optimize the feature space;

[0034] S3. By measuring and minimizing the distribution difference between the newly added class samples and the class prototype through the Gaussian kernel function, the representation offset in the incremental stage is constrained to obtain the InFSAR model;

[0035] S4. Acquire SAR images and perform small-sample target detection based on the InFSAR model.

[0036] Preferably, S1 includes:

[0037] The Faster R-CNN object detection framework is adopted, with the backbone network being ResNet-101 combined with a feature pyramid network to construct the feature extraction model;

[0038] The historical SAR images of the ship are acquired, and the feature extraction model is trained using the historical SAR images to obtain the pre-trained model;

[0039] The pre-trained model is used to extract the basic RoI features of the historical SAR image, and the mean representation of the basic RoI features is calculated to obtain the class prototype:

[0040]

[0041] Where, p c Represents the class prototype, x c x represents the set of all samples in historical SAR images. i y represents the input feature vector. i Indicates category label, f θ This represents the feature extractor of the pre-trained model.

[0042] Preferably, S2 includes:

[0043] Input the K-shot samples of the newly added class, and obtain the RoI feature vector φ through the frozen feature extractor. i ;

[0044] Calculate the eigenvector φ i With the class prototype set P = {p1, ..., p c Cosine similarity between}

[0045]

[0046] Where sim represents the cosine similarity, φ j This represents the vector embedded in the hybrid prototype comparison encoding module;

[0047] Based on the cosine similarity, a loss function is designed by combining the newly added class samples and the class prototype.

[0048] Where M represents the number of labeled sample examples in the dataset. M represents the IoU score that matches its truth box, ε represents the threshold for the Intersection over Union (IoU), and M represents the IoU score. yi This indicates the number of samples of the same type in the current batch. Indicates the indicator function, y j Let p represent the feature of the j-th RoI, C represent the total number of categories, c represent the number of categories, and p represent the feature of the j-th RoI. kThis represents the prototype of class k, and τ represents the temperature coefficient.

[0049] Based on the loss function It forces similar samples to aggregate in the feature space, while separating the new class from the prototype of the base class, thus optimizing the feature space.

[0050] Preferably, S3 includes:

[0051] Given the distribution of class prototypes E={d j The distribution F = {f |j=1,…,C} and the newly added class samples i |i=1,…,Q}, where d j and f i These represent the probability distributions of the class prototype and the newly added class samples, respectively.

[0052] The distance between two distributions is quantified by calculating the maximum mean difference using a Gaussian kernel function.

[0053]

[0054] Where k(x,y) represents the Gaussian kernel function, σ represents the hyperparameter, MMD represents the maximum mean difference, m represents the number of samples in set F, and n represents the number of samples in set E;

[0055] By incorporating MMD as a regularization term into the total loss function, the new class prototype distribution is made to approximate the class prototype distribution, thereby constraining the representation shift during the incremental learning process, and thus obtaining the InFSAR model.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] This invention addresses the problems of high storage costs and significant privacy risks caused by re-accessing base class data when incrementally learning new categories in existing technologies, as well as overfitting and performance degradation (catastrophic forgetting) of base classes due to the scarcity of new category data. By designing a class prototype representation generation module (CPRG), a hybrid class prototype contrastive encoding module (MCPCE), and a prototype calibration strategy, this invention effectively preserves the detection capabilities of base categories without accessing old data, while simultaneously enhancing the discriminative power of new category representations. This enables efficient, flexible, and secure incremental target detection in complex SAR scenarios, meeting the needs of dynamically added categories and improving the overall robustness and scalability of the model. Attached Figure Description

[0058] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the system framework of an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the class prototype representation generation module structure according to an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the hybrid prototype comparison encoding module structure according to an embodiment of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Example 1

[0066] In this embodiment, as Figure 1 , Figure 2 As shown, a SAR image incremental small sample target detection system based on prototype comparison includes: a prototype representation generation module, a hybrid prototype comparison encoding module, a prototype calibration module, and a target detection module.

[0067] The class prototype representation generation module is used to extract the mean of basic class RoI features from historical SAR images as class prototypes based on a pre-trained model.

[0068] The workflow of the class prototype representation generation module includes: using the Faster R-CNN target detection framework, with the backbone network being ResNet-101 combined with a feature pyramid network to construct a feature extraction model; acquiring historical SAR images of ships, and using the historical SAR images to train the feature extraction model to obtain a pre-trained model; using the pre-trained model to extract basic RoI features from the historical SAR images, and calculating the mean representation of the basic RoI features to obtain the class prototype.

[0069] In this embodiment, the class prototype representation generation module (CPRG) functions to learn robust prototype representations containing high-dimensional semantic information of each base class. An example diagram of the module is shown below. Figure 3 As shown, this module can effectively improve the model's generalization ability when processing new types of data and reduce the risk of catastrophic forgetting by capturing and representing the core features of the base class.

[0070] In the basic training phase, the Faster R-CNN object detection framework was adopted, with a ResNet-101 backbone network combined with a Feature Pyramid Network (FPN) to extract multi-scale SAR image features. The region of interest (Rol) feature dimension of the network output was 1024. The training data was the basic class dataset C. base (For example, the 6 types of ships in SRSDD-v1.0), by optimizing the target detection loss function Complete model training, among which, For classifying losses, For bounding box regression loss, The region proposal network loss is used. Training parameters are set as follows: initial learning rate 0.002, batch size 2, stochastic gradient descent (SGD) optimizer, momentum 0.9, weight decay coefficient 0.00001, training epochs 40k (SRSDD-v1.0) or 80k (SARDet-100K), and the learning rate is reduced to 0.0005 during the mid-training phase for stable convergence. After model pre-training, for each base class c∈C... base Extract the RoI features of all its samples and calculate the class prototype p. c This serves as the mean representation of the feature space for that category. Specifically, given the base class sample set D... M ={(x1, y1), ..., (x m y m )},in Given the input feature vector, y i For category labels, the formula for calculating class prototypes is:

[0071]

[0072] Where, p c Represents the class prototype, xc x represents the set of all samples in historical SAR images. i y represents the input feature vector. i Indicates category label, f θ This represents the feature extractor of the pre-trained model. The generated class prototype p... c Its corresponding Softmax classification layer weight w i They are stored together in the prototype library, with the storage format being a key-value pair set {(p1, w1), ..., (p C w C This is so that it can be called in the subsequent incremental learning phase.

[0073] The hybrid class prototype contrast encoding module is used to combine newly added class samples and class prototypes to design a loss function, forcing the aggregation of similar classes and the separation of dissimilar classes, thereby optimizing the feature space.

[0074] In this embodiment, the Hybrid Prototype Contrastive Encoding Module (MCPCE) optimizes the feature space of the new class through contrastive learning, making similar samples clustered tightly and dissimilar samples (including the new class and the base class) significantly separated, thereby reducing classification confusion. Introducing the base class prototype into the comparison enhances the discriminativeness of the new class representation and avoids overfitting due to data scarcity. An example diagram is shown below. Figure 4 As shown.

[0075] The workflow of the hybrid class prototype comparison coding module includes: inputting the newly added class sample C novel K-shot samples (K = 3, 5, 10, 30) were used to obtain RoI feature vectors φ using a frozen feature extractor. i To improve feature quality, only RoI features with an Intersection over Union (IoU) ≥ ε (ε = 0.7) with the ground truth bounding boxes are retained, while low-quality proposals are filtered out to ensure the reliability of feature encoding. The feature vector φ is calculated. i With the class prototype set P = {p1, ..., p c Cosine similarity between}

[0076]

[0077] Where sim represents the cosine similarity, φ j This represents the vector embedded in the hybrid class prototype contrast encoding module; a loss function is designed based on cosine similarity, combining newly added class samples and class prototypes.

[0078]

[0079] Where M represents the number of labeled sample examples in the dataset. M represents the IoU score that matches its truth box, ε represents the threshold for the Intersection over Union (IoU), and M represents the IoU score. yiThis indicates the number of samples of the same type in the current batch. Indicates the indicator function, y j Let p represent the feature of the j-th RoI, C represent the total number of categories, c represent the number of categories, and p represent the feature of the j-th RoI. k This represents the class prototype of category k, where τ represents the temperature coefficient; based on the loss function. It forces similar samples to aggregate in the feature space, while separating the new class from the prototype of the base class, thus optimizing the feature space.

[0080] The class prototype calibration module is used to measure and minimize the distribution difference between the newly added class samples and the class prototype through a Gaussian kernel function, constrain the representation offset in the incremental stage, and obtain the InFSAR model.

[0081] In this embodiment, a distribution alignment strategy effectively reduces the encroachment of new class prototypes on the feature space of the base classes, maintains the detection performance of the base classes, and mitigates catastrophic forgetting. MMD loss ensures that new and old class prototypes maintain consistency in the latent space, preventing the model from deviating from the initial optimization direction due to incremental updates.

[0082] The workflow of the class prototype calibration module includes: given the distribution of class prototypes E = {d} j The distribution F = {f |j=1,…,C} and the newly added class samples i |i=1,…,Q}, where d j and f i Let represent the probability distributions of the class prototype and the newly added class samples, respectively; the distance between the two distributions is quantified by calculating the maximum mean difference using a Gaussian kernel function.

[0083]

[0084]

[0085] Where k(x,y) represents the Gaussian kernel function, σ represents the hyperparameter, MMD represents the maximum mean difference, m represents the number of samples in set F, and n represents the number of samples in set E; by adding MMD as a regularization term to the total loss function, the new class prototype distribution is approximated to the class prototype distribution, thereby constraining the representation bias in the incremental learning process. The calibration loss is defined as:

[0086] L MMD =MMD 2 (E,F);

[0087] Overall loss function design: Integrating the target detection task loss, contrast loss, and distribution calibration loss to form a multi-task joint optimization objective: L = L TFA +αL MCFCF +β LMMDHere, α = 0.5 and β = 0.5 are hyperparameters used to balance the weights of detection accuracy, contrastive learning, and distribution alignment.

[0088] Incremental fine-tuning process: (1) Input preparation: Load the basic category prototype library generated by the training model, and the K-shot sample data of the newly added categories. (2) Parameter freezing: Fix the parameters of the feature extractor (ResNet-101+FPN), and only optimize the weights of the classifier and regressor to reduce computational overhead and prevent loss of basic category knowledge. (3) Forward propagation: Calculate RoI features and detection loss L. TFA , contrast loss L MCFCF and MMD loss L MMD (4) Backpropagation: Retrainable parameters are retrained based on the total loss L until the model converges (usually requiring 10-20 rounds of retraining) to obtain the InFSAR model.

[0089] The target detection module is used to acquire SAR images and perform small-sample target detection based on the InFSAR model.

[0090] Example 2

[0091] In this embodiment, a method for incremental small-sample target detection in SAR images based on prototype comparison includes the following steps:

[0092] S1. Extract the mean value of the base class RoI features from historical SAR images based on the pre-trained model as the class prototype.

[0093] S1 includes: employing the Faster R-CNN object detection framework, with a ResNet-101 backbone network combined with a feature pyramid network to construct a feature extraction model; acquiring historical SAR images of ships and using these historical SAR images to train the feature extraction model, obtaining a pre-trained model; using the pre-trained model to extract basic RoI features from the historical SAR images, and calculating the mean representation of the basic RoI features to obtain the class prototype:

[0094]

[0095] Where, p c Represents the class prototype, x c x represents the set of all samples in historical SAR images. i y represents the input feature vector. i Indicates category label, f θ This represents the feature extractor of the pre-trained model.

[0096] S2. Combine the newly added class samples and class prototypes to design a loss function, force the aggregation of similar classes and the separation of dissimilar classes, and optimize the feature space.

[0097] S2 includes: inputting K-shot samples of newly added class samples, and obtaining the RoI feature vector φ through a frozen feature extractor. i ; Calculate the eigenvector φ i With the class prototype set P = {p1, ..., p c Cosine similarity between}

[0098]

[0099] Where sim represents the cosine similarity, φ j This represents the vector embedded in the hybrid class prototype contrast encoding module; a loss function is designed based on cosine similarity, combining newly added class samples and class prototypes.

[0100]

[0101] Where M represents the number of labeled sample examples in the dataset. M represents the IoU score that matches its truth box, ε represents the threshold for the Intersection over Union (IoU), and M represents the IoU score. yi This indicates the number of samples of the same type in the current batch. Indicates the indicator function, y j Let p represent the feature of the j-th RoI, C represent the total number of categories, c represent the number of categories, and p represent the feature of the j-th RoI. k This represents the class prototype of category k, where τ represents the temperature coefficient; based on the loss function. It forces similar samples to aggregate in the feature space, while separating the new class from the prototype of the base class, thus optimizing the feature space.

[0102] S3. By measuring and minimizing the distribution difference between newly added class samples and class prototypes using a Gaussian kernel function, the representational offset in the incremental stage is constrained, resulting in the InFSAR model.

[0103] S3 includes: the distribution E = {d} of a given class prototype. j The distribution F = {f |j=1,…,C} and the newly added class samples i |i=1,…,Q}, where d j and f i Let represent the probability distributions of the class prototype and the newly added class samples, respectively; the distance between the two distributions is quantified by calculating the maximum mean difference using a Gaussian kernel function.

[0104]

[0105] Where k(x,y) represents the Gaussian kernel function, σ represents the hyperparameter, MMD represents the maximum mean difference, m represents the number of samples in set F, and n represents the number of samples in set E; by adding MMD as a regularization term to the total loss function, the new class prototype distribution is approximated to the class prototype distribution, thereby constraining the representation shift in the incremental learning process, and the InFSAR model is obtained.

[0106] S4. Acquire SAR images and perform small-sample target detection based on the InFSAR model.

[0107] Example 3

[0108] The proposed Incremental FSAR (InFSAR) method for SAR image incremental target detection, based on prototype comparison, significantly improves the overall performance of the model in incremental learning through the synergistic effect of prototype representation generation, hybrid contrastive learning, and distributed calibration strategies. Regarding the retention of basic class detection performance, the model achieves a basic class mean accuracy (bAP) of 77.44% on the 30-shot incremental task of the SRSDD-v1.0 dataset, a 13.14% improvement over the traditional baseline method TFA. On the SARDet-100K dataset, the basic class bAP retention rate exceeds 95%, effectively mitigating the catastrophic forgetting problem caused by the introduction of new class data. For scenarios where new class data is scarce (3-shot to 30-shot), the new class mean accuracy (nAP) is improved by 7.24%-14.05% and 6.14%-14.05% on the SRSDD-v1.0 and SARDet-100K datasets, respectively, significantly outperforming comparative methods such as Meta-RCNN and FPD, while reducing the overfitting rate by more than 30%. Furthermore, the system supports multi-stage incremental tasks (such as 3+2+1 class partitioning), with an overall detection accuracy (All) fluctuation of less than 1%. In the SARDet-100K 30-shot task, the All index reached 69.91%, and in complex SAR scenarios (such as large-scale images and dense targets), it achieved a 18.7% reduction in false positive rate and a 12.3% reduction in false negative rate. By replacing the original data storage with a prototype library, storage costs are reduced by more than 90%, and incremental fine-tuning training time is shortened by 40% (requiring only 10-20 iterations), meeting the need for efficient adaptation to dynamically added classes in practical applications. Moreover, after transferring the model to the MSTAR dataset that was not used in training, the new class detection accuracy (nAP) still remained at 62.3%, validating the method's generalization ability and universality.

[0109] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A SAR image incremental small sample target detection system based on prototype contrast, characterized in that, The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. ， where p c represents a class prototype, x c represents a set of all samples of historical SAR images, x i represents an input feature vector, y i represents a class label, f θ represents a feature extractor of a pre-trained model; The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. Input the K-shot sample of the new class sample, and obtain the RoI feature vector φ through the frozen feature extractor i ; computing the feature vector ϕ i cosine similarity between the class prototype set P = {p1,..., p c} and the feature vector ϕ: ， wherein sim denotes the cosine similarity, ϕ j denotes the vector embedded by the hybrid class prototype contrast encoding module; based on the cosine similarity, combining the newly added class sample and the class prototype design loss function : ， ， wherein M represents the number of data sets of individual labeled sample examples, and i is represented as an IoU score matched with its true value box, and ε represents a threshold value of the intersection over union IoU, and M yi represents the number of same samples in the current batch, represents an indicator function, and y j represents the jth RoI feature, C represents the total number of classes, c represents the class number, and p k represents a class prototype of the class k, and τ represents a temperature coefficient; based on the loss function , forcing the aggregation of homogeneous samples in the feature space while separating the new class from the base class prototypes, optimizing the feature space; The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. Given the distribution of class prototypes E = {d j |j=1,…,C} and the distribution of new class samples F = {f i |i=1,…,Q}, where d j and f i represent the probability distribution of class prototypes and new class samples, respectively. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. ， ， The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof.

2. A SAR image incremental small sample target detection method based on prototype contrast, the method is applied to the system of claim 1, characterized in that, The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. ， where p c represents a class prototype, x c represents a set of all samples of historical SAR images, x i represents an input feature vector, y i represents a class label, f θ represents a feature extractor of a pre-trained model; The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. Input the K-shot sample of the new class sample, and obtain the RoI feature vector φ through the frozen feature extractor i ; computing the feature vector ϕ i cosine similarity between the class prototype set P = {p1,..., p c} and the feature vector ϕ: ， wherein sim denotes the cosine similarity, ϕ j denotes the vector embedded by the hybrid class prototype contrast encoding module; based on the cosine similarity, combining a new class sample and the class prototype design loss function : ， ， wherein M represents the number of data sets of individual labeled sample examples, φ i denotes the IoU score matched with its true value box, and ε represents the threshold of the intersection over union IoU, M yi denotes the number of same-class samples in the current batch, denotes an indicator function, y j denotes the jth RoI feature, C represents the total number of classes, c represents the class number, and p k denotes the class prototype of the kth class, and τ represents the temperature coefficient; based on the loss function , forcing the aggregation of homogeneous samples in the feature space while separating the new class from the base class prototypes, optimizing the feature space; The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. Given the distribution of class prototypes E = {d j |j = 1, …, C} and the distribution of new class samples F = {f i |i = 1, …, Q}, where d j and f i represent the probability distribution of class prototypes and new class samples, respectively. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR) and a device thereof. The application relates to a small sample target detection method based on incremental feature space alignment (InFSAR ， ， Wherein, k(x, y) represents a Gaussian kernel function, σ represents a hyperparameter, MMD represents a maximum mean difference, m represents a sample number in a set F, and n represents a sample number in a set E; By adding the MMD as a regularization term to a total loss function, approximation of a new-class prototype distribution to a class prototype distribution is realized, so as to constrain representation drift in an incremental learning process, and the InFSAR model is obtained.

Images