A Small Sample Target Detection and Recognition Method Based on Self-Supervised Separated Subspace

Abstract

This application discloses a few-shot target detection and recognition method based on self-supervised subspace separation, belonging to the field of remote sensing image processing technology. The proposed innovative framework completely isolates the training of base classes and novel classes at three levels. First, a low-rank subspace adapter is proposed for structural separation, achieving network optimization for novel classes without compromising the pre-training performance of base classes. This uses fewer parameters to modulate network components, alleviating the overfitting problem caused by fully fine-tuning under few-shot conditions. Second, an orthogonal subspace extractor is proposed for feature decoupling, adaptively learning the corresponding subspace for each class and extracting decoupled representations, enhancing the separability between classes. Third, a balanced classifier is designed for loss balancing, preventing the final prediction result from being overly biased towards the background or base class. The framework proposed in this invention has significant advantages in few-shot target detection and recognition, showing significant improvements in both base and novel classes.

Description

Technical Field

[0001] This application belongs to the field of remote sensing image processing technology, specifically relating to a small sample target detection and recognition method based on self-supervised separation subspace. Background Technology

[0002] High-resolution remote sensing data carries a wealth of ground feature information and has broad application potential. However, to fully utilize this data and realize its value, intelligent target detection and recognition are essential. On the one hand, remote sensing target detection has received increasing attention and development with the advancement of traditional deep learning, achieving significant progress, but there is still a gap in achieving comprehensive intelligent detection and recognition. On the other hand, despite the growing demand for remote sensing target detection and recognition, in real-world scenarios, due to complex backgrounds and concealed targets, data is often limited, potentially restricted to specific areas or times, posing challenges to data annotation and the detection and recognition capabilities of algorithms. Therefore, researching reliable intelligent remote sensing target small-sample detection and recognition technologies is crucial for improving the accuracy, real-time performance, and adaptability of target detection and recognition. This helps adapt to the variability of modern environments and has extremely important practical significance.

[0003] For natural images, mainstream few-shot object detection and recognition methods fall into two categories: meta-learning and transfer learning. Both methods divide the data into a base class with a large sample size and a novel class with a small sample size, with no overlap between them. Specifically, meta-learning-based methods construct additional support and query sets, where support images for each class are used to learn class-aware features to retrieve objects from the query image. Similar to meta-learning work in other tasks, an episode-based training strategy is used to complete the transformation from the base class to the novel class. However, such methods suffer from complex data organization and processes, low computational efficiency, and are not suitable for practical applications. In contrast, transfer learning-based methods retain the traditional deep learning object detector structure and achieve simultaneous detection and recognition of both the base and novel classes through a two-stage strategy of pre-training and fine-tuning. Their goal is to pre-train on a data-rich base class and then fine-tune some layers of the resulting pre-trained model. During the fine-tuning stage, balanced data containing both a small number of samples from the base and novel classes is used to maintain the model's detection capability for the base class and transfer knowledge to the novel class.

[0004] In remote sensing scenarios, researchers often combine the characteristics of remote sensing images with the two methods mentioned above, such as introducing spatial attention mechanisms, considering the arbitrary orientation of remote sensing targets, or annotating multimodal text information based on expert knowledge. Compared with meta-learning, many existing remote sensing few-shot target detection and recognition works have widely chosen transfer learning as the basic architecture because it has advantages such as more flexible training, better model performance, and higher efficiency. However, this transfer learning actually has an inherent problem: during the fine-tuning stage, there is a conflict in the optimization gradients between the base class and the novel class. Existing methods ignore this problem, which greatly affects the few-shot target detection and recognition in remote sensing images. By visualizing the direction angle of the base class and novel class losses with respect to the gradients near the output layer parameters during the fine-tuning stage, the gradient direction angles in the weight matrices of each layer all exceed 90 degrees, and most are around 140 degrees. This indicates that they have a mutually exclusive effect on network updates, which makes parameter optimization difficult. The shallower the network layer, the more obvious this conflict becomes. In other words, it is difficult for the model to obtain features that are suitable for both the base class and the novel class in the feature extraction layer before entering the classifier. Therefore, for detectors that are sufficiently pre-trained on base classes, maintaining their base class detection capabilities and generalizing to novel classes from small samples are two contradictory tasks. When fine-tuning the coupled detector to detect all classes, the network will update in an unfavorable direction in the face of the above problem. This will not only severely degrade the model's base class target detection performance but also affect the learning of novel classes. In most remote sensing scenarios that include base class targets such as vehicles and ships, the pre-trained model can detect them completely, but after fine-tuning, a large number of detections will be missed, and training for new classes will become more difficult. Summary of the Invention

[0005] To address the aforementioned technical challenges and completely isolate the training of basic and new classes, this invention proposes a small-sample target detection and recognition method based on self-supervised subspace separation. It employs an innovative two-stage detection architecture, namely the Balanced Orthogonal Subspace Separator (BOSS), which includes three layers of innovative components: a low-rank subspace adapter for structural separation, an orthogonal subspace extractor for feature decoupling, and a balanced classifier for loss balancing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A small-sample target detection and recognition method based on self-supervised subspace separation includes the following steps:

[0008] The first step is to pre-train the two-stage detection network using sufficient base class data, and then further learn the orthogonal subspace of the base classes by extracting the classification features of the candidate boxes, and extract the separability features between the base classes for classification.

[0009] The second step involves adding a low-rank subspace adapter to the linear and convolutional layers of the pre-trained network to separate the structure. Candidate box features of the base class and novel class are obtained on the original branches and the additional branches, respectively. Only the branches of the novel class are used for fine-tuning, while the parameters of the original branches are frozen.

[0010] The third step is to adaptively learn the corresponding subspace for each class based on the orthogonal subspace extractor to decouple features. These subspaces are spanned by orthogonal basis. Only the subspace of the novel class is learned, and the parameters of the subspace of the base class are frozen. The candidate box features output by the original branch and the additional branch in the second step are projected to the base class subspace and the novel class subspace, respectively, to extract the separability features of each class.

[0011] The fourth step is to feed the separability features obtained in the third step into the base class classifier to obtain the base class detection and recognition results. Based on the base class detection and recognition results, the candidate boxes of the novel class are sampled. The training of the novel class classifier is further optimized using a balanced loss function to prevent the final prediction results from being too biased towards the background or the base class.

[0012] The fifth step involves training the small sample target detection and recognition model through steps one through four. During the testing phase, the original branch and the additional branch are used to detect base class and novel class targets respectively, and the results are integrated.

[0013] Further, in the first step, the extracted candidate box features are used as input and projected into a learnable orthogonal subspace to extract separability features between base classes for classification, including:

[0014] For each base class, a set of orthogonal vectors is learned to represent the span vector of its corresponding orthogonal subspace. The basis vectors of the subspaces of different classes are also orthogonal to each other.

[0015] Project the candidate box features onto the subspace corresponding to each category to obtain the features corresponding to each category;

[0016] For background features, the sum of all category features is obtained by subtracting the candidate box features from the candidate box features.

[0017] Furthermore, in the second step, a low-rank subspace adapter is added to the neck region, region proposal network, and candidate box feature extraction component in the two-stage detection network. The low-rank subspace adapter is used for each convolutional and linear layer contained in the component, where the parameters of the original network layers are frozen, and the parameters contained in the added low-rank subspace adapter are used for fine-tuning, including:

[0018] For each network layer, the input is simultaneously fed into the original network layer and the low-rank subspace adapter. The output of the original network layer represents the output of the base class, and the output of the low-rank subspace adapter represents the output of the novel class. The outputs of these two parts are further fed into the subsequent original network layer and low-rank subspace adapter.

[0019] The parameters inside the low-rank subspace adapter are multiple low-rank matrices. By reconstructing them, a matrix with the same size as the original network weights is obtained, and then added to the original network weights to obtain new network weights.

[0020] During training, the parameters inside the low-rank subspace adapter are optimized, and the weights of the original network layers are frozen.

[0021] Furthermore, in the third step, the base class and novel class candidate box features extracted in the second step are used as input, and projected onto the orthogonal subspaces of the base class and novel class respectively. Separability features between each category are then extracted for classification, including:

[0022] For each novel class, a set of orthogonal vectors is learned to represent the span vector of its corresponding orthogonal subspace. The basis vectors of the subspaces of the novel class and the base class are also orthogonal to each other.

[0023] The base class's subspace is frozen, and only the novel class's subspace base is fine-tuned.

[0024] Furthermore, in the fourth step, the base class separability features obtained in the third step are used as input to the base class classifier to obtain the detection and recognition results of the base class targets, and to guide the sampling for novel class training. When calculating the novel class classification loss to train the novel class classifier, a balanced loss function is used to prevent bias towards the background or base class targets, including:

[0025] In the second step, after fine-tuning by the low-rank subspace adapter, if the candidate boxes of the base class and the novel class are the same, then based on the classification results of the base class candidate boxes, only the background class candidate boxes are selected, the novel class recognition is performed, and the classification loss is calculated. Otherwise, the candidate box features of the novel class are passed through the base class orthogonal subspace again to extract the base class classification features, and the above operations are performed.

[0026] For candidate box classification results that require loss calculation, increase the loss weight of candidate boxes labeled as foreground to indirectly suppress the penalty for prediction errors when the label is background.

[0027] Beneficial effects:

[0028] Thanks to the constructed separation framework, this invention can be fine-tuned specifically for novel classes, ensuring that their optimization gradients are not affected and preventing base class degradation. At the same time, it greatly improves knowledge learning for small sample targets, and the method significantly enhances the target detection performance of all categories. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the overall process of a small-sample target detection and recognition method based on self-supervised subspace separation according to the present invention.

[0030] Figure 2 This is a diagram of the network framework proposed in this invention;

[0031] Figure 3 This is a schematic diagram of a low-rank subspace adapter; where (a) represents linear layer modulation and (b) represents convolutional layer modulation.

[0032] Figure 4 This is a schematic diagram of an orthogonal subspace extractor; where (a) is the orthogonal subspace extractor and (b) is the subspace relation.

[0033] Figure 5 This is a schematic diagram illustrating how the method of the present invention avoids gradient conflicts compared to typical methods; where (a) represents the typical method and (b) represents the method of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0035] like Figure 1 As shown, a small-sample target detection and recognition method based on self-supervised subspace separation according to the present invention includes the following steps:

[0036] The first step is to pre-train the two-stage detection network using rich base class data, and then further learn the orthogonal subspace of the base classes by extracting the classification features of the candidate boxes, and extract the separability features between the base classes for classification.

[0037] The second step involves adding a low-rank subspace adapter to the linear and convolutional layers of the pre-trained network to separate the structure. Candidate box features of the base class and novel class are obtained on the original branches and the additional branches, respectively. Only the branches of the novel class are used for fine-tuning, while the parameters of the original branches are frozen.

[0038] The third step involves adaptively learning the corresponding subspace for each class using an orthogonal subspace extractor to decouple features. These subspaces are spanned by orthogonal bases. Only the subspace for the novel class is learned, while the parameters of the base class subspaces are frozen. The candidate box features output from the original branches and the additional branches in the second step are projected onto the base class subspace and the novel class subspace, respectively, to extract the separability features for each class.

[0039] The fourth step involves feeding the base class separability features obtained in the third step into the base class classifier to obtain the base class detection and recognition results. Based on these results, candidate boxes for the novel class are sampled. Furthermore, a balanced loss function is used to optimize the training of the novel class classifier, preventing the final prediction results from being too biased towards the background or the base class.

[0040] Specifically, the overall network framework constructed by this invention is as follows: Figure 2 As shown, N represents the number of candidate boxes, D represents the dimension of the features, N×D is the size of the candidate box features, and C base Represents the number of base classes, N×(C base +1)×D represents the size of the separability feature of the base class, N×C base That is, the size of the classification probability output by the base class classifier, C novel Represents the number of novel classes, N×(C novel +1)×D represents the size of the novel class separability feature, N×C novel That is, the size of the classification probability output by the novel classifier. This represents the base class pre-training loss function. This represents the novelty class fine-tuning loss function. The framework is based on the classic two-stage detector Faster R-CNN and includes the four steps mentioned above. The first stage, base class pre-training, corresponds to the first step, where the base class features in the diagram represent the process of learning the orthogonal subspace of the base classes to extract classification features. The second stage, novelty class fine-tuning, encompasses the low-rank subspace adapter, the orthogonal subspace extractor for novelty classes, and the balanced classifier involved in the second, third, and fourth steps, as shown below. Figure 2 The lower half of the diagram is shown. During training, the original model interacts with these modules, transferring information stored in pre-training to them to achieve generalization to new classes. Finally, during inference, predictions are made separately by the base class classifier and the novel class classifier, and then combined to obtain the complete detection result. The following will elaborate on these three components.

[0041] First, to preserve the detection results of base classes during the fine-tuning phase, this invention freezes some components in the pre-trained network, including the backbone network, neck, region proposal network, and candidate box feature extraction part. Obviously, using these frozen parameters, the base classifier will only obtain base class targets. Therefore, to update the network parameters on this basis to detect novel class targets while maintaining the detection results of base class targets, this invention, inspired by adapter fine-tuning methods for visual base models and large language models, proposes a branch fine-tuning method based on Low-Rank Subspace Adapter (LoSA). This method adds a trainable adapter to the frozen structure for updating, rather than directly updating the original model as in traditional transfer learning. In other words, the added adapter is dedicated to learning to detect novel classes, achieving structural separation from the original framework. LoSA can be used to modulate arbitrary convolutional and linear layers, i.e., linear layer LoSA and convolutional layer LoSA modulation, approximating the learned weight matrix as a low-rank matrix.

[0042] Figure 3 This paper demonstrates the low-rank decomposition method employed in this invention, which reconstructs a learnable weight matrix based on the Hadamard product. Specifically, Figure 3 (a) is linear layer modulation. Figure 3 (b) represents the modulation of the convolutional layer.

[0043] For linear LoSA, the calculation process for input x is as follows:

[0044]

[0045] in, Represents the set of real numbers, C in It is the input feature dimension, C out This refers to the output feature dimension, where r1 and r2 are the dimensions of the low-rank matrix used for decomposition. Here, W represents the input features, W represents the original pre-trained weight parameters, and b represents the pre-trained bias parameters. as well as These are learnable parameters in the low-rank subspace. T represents the matrix transpose operation, ⊙ represents the Hadamard product (element-by-element multiplication), Linear represents the linear layer operation of the network, and Cat represents the matrix concatenation operation. F base , Let represent the candidate bounding box features of the base class and novel class respectively, and y represent the result obtained by concatenating the output base class and novel class features. Similarly, the calculation of LoSA for the convolutional layer is as follows:

[0046]

[0047] in, These are the input features, where H and W are the height and width of the two-dimensional input, respectively. is a learnable parameter in the low-rank subspace, d is the kernel size, and The `reshape` operation represents the size transformation of a matrix, changing the original (C) matrix size. in ×d×d)×C out Transform into C out ×C in ×d×d, Conv2d indicates that when implementing 2D convolutional layer operations in the network, r1 and r2 will be set to r1, r2 << min(C in C out This ensures that the final constructed low-rank matrix... The rank is sufficiently low, thus significantly improving optimization efficiency and alleviating memory consumption. Since pre-trained overparameterized models typically have low intrinsic dimensionality, the aforementioned low-rank decomposition and reconstruction process can be effectively trained. Furthermore, previous full-scale fine-tuning methods are more prone to overfitting under small sample conditions, while the model of this invention trains only a small number of parameters, mitigating this problem and achieving better performance. Additionally, to maintain the model in its original state during initial fine-tuning, i.e., F... base =F novel X2 is initialized to zero, while other low-rank matrices are initialized using normal methods. Then, F for each layer... novel All will be gradually updated to accommodate novel object detection. Due to the high flexibility of this LoSA structure, it can be used for fine-tuning of some or all modules. The shape of the input features needs to be adjusted accordingly for different modules. For example, when only the region proposal network and candidate box feature extraction are fine-tuned, all candidate boxes originate from the feature map output from the neck. The candidate boxes corresponding to the base class branch and the novel class branch both originate from it and are concatenated before being input into the candidate box feature extraction part; in this case, the number of candidate boxes is twice that of the pre-training stage. For x input into the candidate box feature extraction, the following segmentation is required, and the corresponding F is obtained through LoSA. base and F novel :

[0048]

[0049] in, This represents the features obtained from the base class candidate boxes output by the base class branch of the region proposal network. This represents the features obtained from the novel class candidate boxes output by the novel class branch of the region proposal network. f represents a linear or convolutional operation.

[0050] The low-rank subspace adapter proposed in this invention achieves structural separation based on Faster R-CNN. It can independently extract candidate box features from the base class and novel class parts, aiding in feature decoupling and loss calculation in subsequent steps. Furthermore, its flexibility allows this separation design to be easily transferred to other base detectors, such as DETR or the YOLO series.

[0051] After obtaining candidate box features from the pre-trained base class branch and the novel class branch fine-tuned by LoSA, this invention further proposes an orthogonal subspace extractor (OSE), which decouples features from different categories, enhances their discriminability, and improves the model's ability to transfer to new categories. The specific implementation is as follows... Figure 4 As shown. For each base class and new class, a set of orthogonal bases is learned to span the hyperplane, i.e., a low-dimensional orthogonal subspace. The candidate box features F of the original base classes. base And the candidate box features F of the original novel class novel The features will be projected onto these hyperplanes to obtain features for classification. For simplicity, this invention uses a model containing only one base class A and one novel class B. Figure 4 For example, (a):

[0052]

[0053] in Let F represent an orthogonal basis, which is a set of learnable parameters, where i and j represent indices of the orthogonal basis. A , These correspond to the classification representations of class A and class B, respectively. A n B Let represent the starting indices of the orthogonal basis vectors spanning subspaces A and B, respectively, and let n represent the number of orthogonal bases spanning each subspace, whose value is preset. In the case shown above, n A =0, n B =2, n=2, but in actual implementation, this invention is applicable to cases containing multiple base classes and novel classes, and their calculation methods are the same as F. A and F B Similarly, n will be set larger to extract more expressive feature representations. Finally, based on F... base For each base class A 1 A 2 ... Able to obtain feature representation Based on F novel For each novel class B 1 B 2 ... Able to obtain feature representation ..., Ultimately, this invention can obtain the background class representation F in the two stages of base class pre-training and novel class fine-tuning. bbg and F nbg :

[0054]

[0055] During the fine-tuning of novel classes, the subspaces belonging to the base classes are fixed, i.e., the relevant orthogonal bases are frozen. Finally, all feature representations are input into the corresponding classifiers for class prediction. Since the subspaces learned in this invention are mutually orthogonal, the representations extracted from these hyperplanes are also orthogonal. Similarly, using F... A and F B For example, Figure 4 (b) illustrates this phenomenon. For ease of visualization, only a portion of the orthogonal basis P1 in the class A subspace is shown in the figure. The theoretical derivation is as follows:

[0056]

[0057] Therefore, by using the OSE (Orthogonal Subspace Extractor) of this invention, decorrelation transformation is achieved, which enhances the independence between features. Furthermore, unlike other strategies for learning orthogonality, which extend to higher-dimensional subspaces and limit the expressive power of the obtained decoupled representations, the proposed OSE demonstrates its effectiveness through integration into the fully isolated framework BOSS (Balanced Orthogonal Subspace Separator) of this invention. It emphasizes learning separate representations of base classes and novel classes, promoting transfer learning for detecting novel classes, and self-supervised learning of feature-separable orthogonal subspaces.

[0058] Finally, to identify the decoupled features of each category extracted from OSE, this invention designs a balanced classifier (BC) module, comprising a base classifier and a novel classifier trained in two separate stages, wherein the base classifier is completely frozen in the second stage. The training loss in the first stage... as follows:

[0059]

[0060] in, This represents the classification loss using cross-entropy loss, which is based on the probability distribution C output by the base class classifier. b (F′ base It is calculated using ). This directly uses the feature F obtained from the candidate box feature extraction part. base To predict the detection box, calculate the L1 loss and then obtain the result. This indicates the calculation of classification loss and detection loss for the candidate bounding boxes output by the region proposal network. The classification loss uses the binary cross-entropy loss of foreground and background, and the detection loss uses the L1 loss. The orthogonality loss of the base class subspace constructed in OSE is:

[0061]

[0062] Where, n base N represents the number of base classes. b =n·n base , where represents the number of all bases that constitute all base class subspaces, and these vectors are constrained to be mutually orthogonal.

[0063] During the second stage of fine-tuning, this invention uses a frozen base class classifier to guide the sampling and training of the novel class classifier. It identifies the base class based on the base class classification features of the extracted N candidate boxes and retains only the N... sub There are 10 candidate boxes, which are identified as background and whose labels are not base class. Furthermore, considering that if the part before ROI feature extraction is modulated using LoSA, the base class branch and novel class branch in the RPN will generate different candidate boxes, therefore, candidate boxes from the novel class branch need to be simultaneously re-input into the subsequent base class branch and novel class branch to guide sampling N. sub The detection boxes are generated and trained. In the actual implementation, the above operations are performed in parallel to minimize the impact on efficiency. Then, the present invention further balances the loss based on the distribution of the results, mitigating the prediction bias of the base class and background of the previous classifier caused by the imbalance of sampling under small sample conditions. Specifically, the loss function in the second stage... as follows:

[0064]

[0065] Similarly, n novel N represents the number of novel classes. n =n·n novel This represents the number of bases for all novel class subspaces of Zhang Cheng. This represents the orthogonality loss of the novel class subspaces constructed in OSE. In the balance loss... In, N c =n novel +1, includes the background class, p i，j and y i，j Let c represent the probability and ground truth value that the predicted i-th detection result belongs to the j-th class, respectively. i，j These are the weighting coefficients that balance the classification loss, and they are determined as follows:

[0066]

[0067] Here, α, β, γ, and θ are four hyperparameters, and P represents the probability of the event occurring. When the four hyperparameters α, β, γ, and θ in the above equation take the same value, the loss under different outcome distributions will be equally emphasized. However, since the number of positive target samples is limited, most candidate targets are actually labeled as background. Therefore, the probability of optimizing the distribution of the second and fourth terms in the above equation is high, which can lead to the target being easily misidentified. Therefore, this invention sets larger α and γ to emphasize the correct prediction of novel classes and relatively reduce the penalty for incorrect prediction of background classes. This adjustment makes the loss more balanced, which helps the model detect and identify more novel targets. In addition, the balanced loss of this invention... It can be viewed as a weighted multi-binary classification loss, which reduces mutual inhibition between different categories compared to the cross-entropy of multi-class classification. Furthermore, the fully separable structure and decoupled representation learning of this invention better facilitate independent learning of features for each category.

[0068] Finally, experiments on publicly available datasets such as DIOR (Digital Optical Remote Sensing Image Target Detection Dataset) demonstrate that BOSS achieves optimal performance under various configurations. Through the combined action of multiple modules proposed in this invention, optimization gradient conflicts between base and novel classes are avoided, effectively improving detection performance across all categories. This fully decoupled framework self-supervises the learning of feature-separable orthogonal subspaces, enabling it to adapt to small-sample target detection and recognition in complex scenes across different categories.

[0069] The fifth step involves training the small sample target detection and recognition model through steps one through four. During the testing phase, the model detects base class and novel class targets through two branches and integrates the results.

[0070] Based on the above description, the present invention effectively avoids Figure 5 The gradient conflict problem in typical transfer learning methods in (a) is a significant issue, where all base class vehicles that the pre-trained model can detect are lost after fine-tuning, and training novel classes becomes more difficult; for example, the model cannot detect novel class targets such as baseball fields. In contrast, Figure 5 (b) illustrates the results of the method of the present invention, which prevents base class degradation and focuses on the optimization of new classes, greatly improving the ability to learn target knowledge from small samples.

[0071] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A small-sample target detection and recognition method based on self-supervised subspace separation, characterized in that, Includes the following steps: The first step involves pre-training the two-stage detection network using sufficient base class data, then extracting the classification features of candidate boxes to further learn the orthogonal subspace of the base classes, and extracting the separability features between base classes for classification; the data is remote sensing image data. The second step involves adding a low-rank subspace adapter to the linear and convolutional layers of the pre-trained network to separate the structure. Candidate box features of the base class and novel class are obtained on the original branches and the additional branches, respectively. Only the branches of the novel class are used for fine-tuning, while the parameters of the original branches are frozen. For the neck region, region proposal network, and candidate box feature extraction components in the two-stage detection network, a low-rank subspace adapter is added. This low-rank subspace adapter is used for each convolutional and linear layer within the component. The parameters of the original network layers are frozen, and the parameters included in the added low-rank subspace adapter are used for fine-tuning, including: For each network layer, the input is simultaneously fed into the original network layer and the low-rank subspace adapter. The output of the original network layer represents the output of the base class, and the output of the low-rank subspace adapter represents the output of the novel class. The outputs of these two parts are further fed into the subsequent original network layer and low-rank subspace adapter. The parameters inside the low-rank subspace adapter are multiple low-rank matrices. By reconstructing them, a matrix with the same size as the original network weights is obtained, and then added to the original network weights to obtain new network weights. During training, the parameters inside the low-rank subspace adapter are optimized, and the weights of the original network layers are frozen. The third step is to adaptively learn the corresponding subspace for each class based on the orthogonal subspace extractor to decouple features. These subspaces are spanned by orthogonal basis. Only the subspace of the novel class is learned, and the parameters of the subspace of the base class are frozen. The candidate box features output by the original branch and the additional branch in the second step are projected to the base class subspace and the novel class subspace, respectively, to extract the separability features of each class. The fourth step is to feed the separability features obtained in the third step into the base class classifier to obtain the base class detection and recognition results. Based on the base class detection and recognition results, the candidate boxes of the novel class are sampled. The training of the novel class classifier is further optimized using a balanced loss function to prevent the final prediction results from being too biased towards the background or the base class. The fifth step involves training the small sample target detection and recognition model through steps one through four. During the testing phase, the original branch and the additional branch are used to detect base class and novel class targets respectively, and the results are integrated.

2. The method for small-sample target detection and recognition based on self-supervised subspace separation according to claim 1, characterized in that: In the first step, the extracted candidate box features are used as input and projected into a learnable orthogonal subspace. Separability features between base classes are extracted for classification, including: For each base class, a set of orthogonal vectors is learned to represent the span vector of its corresponding orthogonal subspace. The basis vectors of the subspaces of different classes are also orthogonal to each other. Project the candidate box features onto the subspace corresponding to each category to obtain the features corresponding to each category; For background features, the sum of all category features is obtained by subtracting the candidate box features from the candidate box features.

3. The method for small-sample target detection and recognition based on self-supervised subspace separation according to claim 1, characterized in that: In the third step, the base class and novel class candidate box features extracted in the second step are used as input, and they are projected into the orthogonal subspaces of the base class and novel class, respectively. Separability features between each category are then extracted for classification, including: For each novel class, a set of orthogonal vectors is learned to represent the span vector of its corresponding orthogonal subspace. The basis vectors of the subspaces of the novel class and the base class are also orthogonal to each other. The base class's subspace is frozen, and only the novel class's subspace base is fine-tuned.

4. The method for small-sample target detection and recognition based on self-supervised subspace separation according to claim 1, characterized in that: In the fourth step, the base class separability features obtained in the third step are used as input to the base class classifier to obtain the detection and recognition results of the base class targets, and to guide the sampling for novel class training. When calculating the novel class classification loss to train the novel class classifier, a balanced loss function is used to prevent bias towards the background or base class targets, including: In the second step, after fine-tuning by the low-rank subspace adapter, if the candidate boxes of the base class and the novel class are the same, then based on the classification results of the base class candidate boxes, only the background class candidate boxes are selected, the novel class recognition is performed, and the classification loss is calculated. Otherwise, the candidate box features of the novel class are passed through the base class orthogonal subspace again to extract the base class classification features, and the above operations are performed. For candidate box classification results that require loss calculation, increase the loss weight of candidate boxes labeled as foreground to indirectly suppress the penalty for prediction errors when the label is background.

Images