Small sample target detection method and device based on multi-angle enhanced feature pyramid

Abstract

The application provides a kind of small sample target detection method and device based on multi-angle enhanced feature pyramid, it is related to remote sensing image processing technical field.The method comprises: inputting small sample remote sensing image into the double-angle enhanced feature pyramid module based on self-supervised visual large model, dynamic scale reconstruction module and feature pyramid network to obtain double-angle enhanced feature pyramid and attention map;Input multi-angle enhanced feature pyramid module based on double-angle enhanced feature pyramid module and residual weighted adjustment mechanism to obtain multi-angle enhanced feature pyramid;Input region candidate network and target detection head to obtain target positioning and classification result of remote sensing image.The multi-angle enhanced feature pyramid constructed by the application provides high-quality feature support for the candidate region recommendation network, realizes end-to-end small sample remote sensing image target detection, can effectively improve the generalization and robustness of target detection network, significantly improves the detection precision of model under a small amount of labeled samples.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing image processing technology, and in particular to a method and apparatus for small-sample target detection based on multi-angle enhanced feature pyramids. Background Technology

[0002] Remote sensing image target detection technology has immense practical value in modern life. For example, in the defense and military field, the all-round, wide-spectrum, and multi-scale characteristics of remote sensing technology make the battlefield more transparent. During the construction of national defense projects, by comparing visible light, infrared, and radar remote sensing images acquired at different times in the project area, the differences in terrain and landforms between the camouflaged project environment and the initial environment can be examined, providing a basis for more scientific camouflage. In the field of resource exploration, remote sensing images capture information such as surface mineralization, vegetation, and geological structures, enabling the monitoring and identification of mineral and water resources, and improving the accuracy and efficiency of mineral area location.

[0003] Traditional object detection algorithms have achieved good results. When applying these algorithms in the natural world, sufficient high-quality labeled data is typically needed to train the object detection model. This helps the model better learn the discriminative features of objects, thereby improving the model's detection accuracy and generalization ability. High-quality labeled data not only requires accurate object category labeling and precise bounding box positions, but also needs to cover diverse scene conditions and object morphological variations. However, labeling a large number of high-quality training samples is time-consuming and labor-intensive. Moreover, in the field of remote sensing, optical remote sensing images often exhibit high object feature complexity and large scale and shape variations. Therefore, labeling high-quality remote sensing samples is time-consuming and labor-intensive. Furthermore, in military and defense settings, obtaining sufficient training samples is impossible due to confidentiality requirements. Therefore, it is necessary to study object detection in remote sensing images under conditions of limited labeled samples.

[0004] While numerous methods have been developed for target detection in small-sample remote sensing images, they still fall short in mitigating base class feature bias to improve the semantic generalization and discriminative power of detection models, making it difficult to meet the needs of target detection in small-sample remote sensing images. Therefore, a novel feature pyramid with strong semantic generalization and discriminative power is urgently needed to improve the detection accuracy of the model. Summary of the Invention

[0005] To address the technical problem that existing research methods are insufficient to meet the high generalization requirements of target detection in small-sample remote sensing images, this invention provides a method and apparatus for small-sample target detection based on a multi-angle enhanced feature pyramid. The technical solution is as follows: On the one hand, a small-sample target detection method based on multi-angle enhanced feature pyramids is provided. This method is implemented using a small-sample remote sensing image target detection device and includes: S1. Obtain small sample remote sensing images. Construct a dual-angle enhanced feature pyramid module based on a self-supervised visual large model, a dynamic scale reconstruction module, and a feature pyramid network. Input the small sample remote sensing images into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map.

[0006] S2. Construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. Input the dual-angle enhanced feature pyramid and the attention map into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid.

[0007] S3. Input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing images. Fine-tune the region candidate network and the target detection head according to the corresponding annotations of the small sample remote sensing images. Obtain the target detection network of the small sample remote sensing images based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head.

[0008] S4. Acquire the remote sensing image to be detected, input the remote sensing image into the small sample remote sensing image target detection network, and obtain the target detection result.

[0009] Optionally, in S1, the small sample remote sensing images are input into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid, including: By inputting small sample remote sensing images into a self-supervised visual large model, visual feature maps of encoders at different levels in the self-supervised visual large model are obtained.

[0010] Visual feature maps from different encoder levels and remote sensing images are input into the dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps.

[0011] Multi-scale feature maps are input into a feature pyramid network for multi-scale feature fusion to obtain a dual-angle enhanced feature pyramid.

[0012] Optionally, visual feature maps from different encoder levels and remote sensing images are input into a dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps, including: Shallow features from the visual feature maps of different encoder levels are input into the plug-and-play feature upsampling submodule to obtain the upsampled feature map.

[0013] Deep features from the visual feature maps of different encoder levels are input into the downsampling submodule to obtain downsampled feature maps.

[0014] A multi-scale feature map is obtained based on the upsampled feature map, the visual feature map, and the downsampled feature map.

[0015] Optionally, the process of constructing the attention map includes: By utilizing the self-attention mechanism of a self-supervised large visual model, a coarse attention map is constructed by calculating the attention between category terms and image patch embedding features. The coarse attention map is then denoised using a linear piecewise function to obtain the final attention map.

[0016] Alternatively, the process of constructing the coarse attention map is shown in equations (1)-(2): (1) (2) In the formula, Represents a sequence of lexical terms. Indicates category terms, Indicates the first Image block words, Indicates the first Image blocks, Indicates shared ownership Image blocks, Represents the attention matrix. This indicates the self-attention mechanism. Represents the query matrix. Represents the key matrix, Representative value matrix, Indicates matrix transpose. Represents the dimension of the key matrix vector. Indicates category attention, Indicates the first Image block word attention.

[0017] Optionally, in S2, the dual-angle augmented feature pyramid and attention map are input into the multi-angle augmented feature pyramid module to obtain a multi-angle augmented feature pyramid, including: Interpolation is used to transform the attention map into a multi-scale attention map. The multi-scale attention map and the dual-angle enhanced feature pyramid are then effectively combined through a residual weighted adjustment mechanism to obtain the multi-angle enhanced feature pyramid.

[0018] Optionally, the process of obtaining the multi-angle enhanced feature pyramid is shown in equations (3)-(4) below: (3) (4) In the formula, This represents the multi-angle enhanced feature pyramid modulated by the attention map. Representation layer normalization, The first part represents the dual-angle enhanced feature pyramid. Layer feature map, Represents the attention modulation coefficient. Indicates the first Attention maps corresponding to layer feature maps Indicates linear interpolation. This represents the original attention map.

[0019] On the other hand, a small-sample target detection device based on a multi-angle enhanced feature pyramid is provided. This device is applied to a small-sample target detection method based on a multi-angle enhanced feature pyramid. The device includes: The dual-angle enhanced feature pyramid module is used to acquire small sample remote sensing images. It is constructed based on a self-supervised visual large model, a dynamic scale reconstruction module, and a feature pyramid network. The small sample remote sensing images are input into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map.

[0020] The multi-angle enhanced feature pyramid module is used to construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. The dual-angle enhanced feature pyramid and attention map are input into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid.

[0021] The fine-tuning module is used to input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing images. Based on the annotations corresponding to the small sample remote sensing images, the region candidate network and the target detection head are fine-tuned. Based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head, the target detection network of the small sample remote sensing images is obtained.

[0022] The target detection module is used to acquire remote sensing images of the targets to be detected, input the remote sensing images into a small sample remote sensing image target detection network, and obtain the target detection results.

[0023] Optionally, the dual-angle enhanced feature pyramid module is further used for: By inputting small sample remote sensing images into a self-supervised visual large model, visual feature maps of encoders at different levels in the self-supervised visual large model are obtained.

[0024] Visual feature maps from different encoder levels and remote sensing images are input into the dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps.

[0025] Multi-scale feature maps are input into a feature pyramid network for multi-scale feature fusion to obtain a dual-angle enhanced feature pyramid.

[0026] Optionally, the dual-angle enhanced feature pyramid module is further used for: Shallow features from the visual feature maps of different encoder levels are input into the plug-and-play feature upsampling submodule to obtain the upsampled feature map.

[0027] Deep features from the visual feature maps of different encoder levels are input into the downsampling submodule to obtain downsampled feature maps.

[0028] A multi-scale feature map is obtained based on the upsampled feature map, the visual feature map, and the downsampled feature map.

[0029] Optionally, the dual-angle enhanced feature pyramid module is further used for: By utilizing the self-attention mechanism of a self-supervised large visual model, a coarse attention map is constructed by calculating the attention between category terms and image patch embedding features. The coarse attention map is then denoised using a linear piecewise function to obtain the final attention map.

[0030] Alternatively, the process of constructing the coarse attention map is shown in equations (1)-(2): (1) (2) In the formula, Represents a sequence of lexical terms. Indicates category terms, Indicates the first Image block words, Indicates the first Image blocks, Indicates shared ownership Image blocks, Represents the attention matrix. This indicates the self-attention mechanism. Represents the query matrix. Represents the key matrix, Representative value matrix, Indicates matrix transpose. Represents the dimension of the key matrix vector. Indicates category attention, Indicates the first Image block word attention.

[0031] Optionally, the multi-angle enhanced feature pyramid module is further used for: Interpolation is used to transform the attention map into a multi-scale attention map. The multi-scale attention map and the dual-angle enhanced feature pyramid are then effectively combined through a residual weighted adjustment mechanism to obtain the multi-angle enhanced feature pyramid.

[0032] Optionally, the process of obtaining the multi-angle enhanced feature pyramid is shown in equations (3)-(4) below: (3) (4) In the formula, This represents the multi-angle enhanced feature pyramid modulated by the attention map. Representation layer normalization, The first part represents the dual-angle enhanced feature pyramid. Layer feature map, Represents the attention modulation coefficient. Indicates the first Attention maps corresponding to layer feature maps Indicates linear interpolation. This represents the original attention map.

[0033] On the other hand, a small sample remote sensing image target detection device is provided, the small sample remote sensing image target detection device comprising: a processor; a memory, the memory storing computer-readable instructions, which, when executed by the processor, implement any of the methods described above for small sample target detection based on multi-angle enhanced feature pyramids.

[0034] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods for few-sample target detection based on multi-angle enhanced feature pyramids.

[0035] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, the general feature output capability of a self-supervised large visual model is utilized to solve the base class feature shift problem in traditional small-sample remote sensing image target detection methods based on ResNet backbone networks.

[0036] By introducing a dynamic scale reconstruction module, the problem of single-scale output features in large visual models is addressed. This module is lightweight and class-independent, providing multi-scale feature support for feature fusion, enabling the detection of targets of different sizes.

[0037] Based on the self-attention mechanism in the large visual model, an attention map reflecting the foreground target is constructed. Through the residual weighted modulation mechanism, spatial prior information is provided for the dual-angle enhanced feature pyramid, thereby improving feature discriminability.

[0038] The detection method is highly flexible and easily expandable, making it applicable to target detection in various small-sample remote sensing images, and can effectively improve the detection accuracy and robustness of new categories. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0040] Figure 1 This is a flowchart of a small sample target detection method based on a multi-angle enhanced feature pyramid provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the dynamic scale reconstruction module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the dual-angle enhanced feature pyramid provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a multi-angle enhanced feature pyramid obtained by attention map weighted modulation according to an embodiment of the present invention; Figure 5 These are attention diagrams before and after noise reduction provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the overall process framework provided in the embodiments of the present invention; Figure 7 This is a block diagram of a small sample target detection device based on a multi-angle enhanced feature pyramid provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a small-sample remote sensing image target detection device provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0042] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0043] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0044] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0045] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0046] This invention provides a small-sample target detection method based on a multi-angle enhanced feature pyramid. This method can be implemented using a small-sample remote sensing image target detection device, which can be a terminal or a server. Figure 1 The flowchart shown is for a few-sample target detection method based on multi-angle enhanced feature pyramids. The processing flow of this method can include the following steps: S1. Obtain small sample remote sensing images. Construct a dual-angle enhanced feature pyramid module based on a self-supervised visual large model, a dynamic scale reconstruction module, and a feature pyramid network. Input the small sample remote sensing images into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map.

[0047] Optionally, the step S1 above, which involves inputting a small sample of remote sensing images into a dual-angle enhanced feature pyramid module to obtain a dual-angle enhanced feature pyramid, includes the following steps S11-S13: S11. Input small sample remote sensing images into the self-supervised visual large model to obtain visual feature maps of encoders at different levels in the self-supervised visual large model.

[0048] In one feasible implementation, a self-supervised visual large model is used to output visual features of encoders at different levels. Thanks to the use of a large-scale dataset, the visual large model is pre-trained through self-supervised learning. Under the premise of freezing parameters, it can output general features with strong expressiveness.

[0049] Specifically, the large visual model outputs features from different levels of encoders (11, 13, 19, 23), providing semantic information at different levels. (Image) First, it is divided into multiple image blocks of the same size, among which and These represent the height and width of the image, respectively. The output configurations of different encoder levels are shown in Table 1. Table 1 Output configuration of encoders at different levels in the self-supervised large-scale vision model

[0050] S12. Input the visual feature maps of different encoder levels and remote sensing images into the dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps.

[0051] Optionally, step S12 above may include the following steps S121-S123: S121. Input the shallow features in the visual feature maps of different encoder levels into the plug-and-play feature upsampling submodule to obtain the upsampled feature map.

[0052] In one feasible implementation, a dynamic scale reconstruction module is used to address the problem of the single scale of the output features of the large visual model. The plug-and-play feature upsampling submodule is used to dynamically upsample the shallow features of the self-supervised large visual model output. This module does not contain any parameters that need to be trained, so it is class-independent and retains the general features of the self-supervised large visual model output to the greatest extent. This enables the effective reconstruction of the output features of the large visual model under different size and appearance change conditions.

[0053] S122. Input the deep features from the visual feature maps of different encoder levels into the downsampling submodule to obtain the downsampled feature map.

[0054] In one feasible implementation, max pooling is used to downsample the deep features output by the self-supervised visual large model, providing the model with the ability to process targets of different scales.

[0055] S123. Obtain a multi-scale feature map based on the upsampled feature map, the visual feature map, and the downsampled feature map.

[0056] In one feasible implementation, the scale reconstruction of the output features is completed by combining the above-described upsampling and downsampling operations with the features output by the self-supervised visual large model.

[0057] S13. Input the multi-scale feature map into the feature pyramid network to perform multi-scale feature fusion and obtain the dual-angle enhanced feature pyramid.

[0058] In one feasible implementation, a feature pyramid network is used to effectively fuse the multi-scale features output by the dynamic scale reconstruction module.

[0059] In this embodiment, a large visual model is used to extract remote sensing image feature outputs from encoders of different levels. For the feature outputs of the 11th and 13th layer encoders, a plug-and-play upsampling model is used to upsample the resolution feature maps to a specified resolution. The feature output of the 19th layer encoder remains unchanged, while the feature output of the 23rd layer encoder is downsampled. Subsequently, a feature pyramid network is used for feature fusion. Figure 2 As shown, the specific steps may include: B1: For the input image The system uses a large visual model to extract features at different levels. After the image is divided into image patches of the same size, it undergoes convolutional operations for feature encoding. The encoded features are then input into multiple identical visual encoders for processing. Each encoder includes self-attention and fully connected operations. (1) In the formula, These represent encoders at different levels. This represents the output characteristics of encoders at different levels. Indicates the first Self-attention computation in layer encoders Indicates the first Fully connected operations in layer encoders Indicates the first Self-attention computation in layer encoders Indicates the first Fully connected operations in layer encoders This represents the convolution operation. This indicates that the size after splitting is Image blocks.

[0060] B2: Due to feature map The scale size is Therefore, the dynamic scaling reconstruction module is used to reconstruct the features: (2) (3) (4) (5) In the formula, This represents the max pooling operation, used for downsampling features. It means to remain unchanged. and These represent using a plug-and-play upsampling module to upsample the features to... and Feature map of size, This represents the multi-scale features reconstructed by the dynamic scale reconstruction module, with resolutions of [missing information]. , , , The dynamic scale reconstruction module and the plug-and-play upsampling module, such as... Figure 3 As shown.

[0061] B3: Multiscale It contains output information from different encoders, and the output features at each level contain information with different focuses. Shallow features generally contain information such as the texture and edges of the object, while deep features contain semantic information. Therefore, it is necessary to fuse information from multiple scale features. (6) In the formula, The output of the dynamic scale reconstruction module represents the first... Layer features, This indicates the characteristics of the pyramid after merging. express convolution, This indicates a 2x upsampling. This represents the ReLU activation function, used for nonlinear transformations.

[0062] Through the above operations, a dual-angle enhanced feature pyramid is obtained.

[0063] This invention enhances the semantic generalization and scale robustness of feature pyramids using a self-supervised large-scale visual model and a plug-and-play upsampling module. Specifically, to enhance the semantic generalization of feature pyramids, the self-supervised large-scale visual model, relying on large-scale data and employing self-supervised training, leverages its general feature representation capabilities to address the base class feature shift problem present in traditional methods. To enhance the scale robustness of feature pyramids, a plug-and-play upsampling module is used to construct a dynamic scale reconstruction module without any training parameters, addressing the issue of the single output scale of the large-scale visual model. Combining these two points, a dual-angle enhanced feature pyramid is constructed.

[0064] Optionally, the process of constructing the attention map in step S1 above includes: By utilizing the self-attention mechanism of a self-supervised large visual model, a coarse attention map is constructed by calculating the attention between category terms and image patch embedding features. The coarse attention map is then denoised using a linear piecewise function to obtain the final attention map.

[0065] In one feasible implementation, in using attention maps to enhance the discriminative power of the dual-angle enhanced feature pyramid, an offline attention map storage unit is constructed. Utilizing the self-attention mechanism of a self-supervised large visual model, a category-level attention map is constructed by calculating the attention between category terms and image patch embedding features. Simultaneously, to accelerate the training speed of the detection model, the attention map is pre-constructed and saved as a .pt file.

[0066] Specifically, attention maps are used to enhance the discriminative power of the dual-angle enhanced feature pyramid, such as... Figure 4 As shown, the specific steps may include: C1: Extracted from the multi-head self-attention in the last encoder of the supervised visual large model: (7) (8) In the formula, , , These represent the query, key, and value matrices obtained by linear mapping of the input features, respectively. Indicates the first The linear mapping parameters corresponding to each attention map are used to project the input features into different subspaces. This represents the feature dimension of each attention map. Indicates the number of attention heads. The function is used to normalize the attention score, representing the importance of different positions. The outputs of multiple attentions are concatenated along the feature dimension.

[0067] C2: To balance the importance of different attention heads, the attention from multiple attention heads is averaged. (9) The large visual model's self-attention layer contains a total of 12 attention heads. Indicates the first One point of attention, This represents average attention.

[0068] C3: When performing self-attention calculation, the input includes category terms and image patch terms. Only the attention between category terms and image patch terms is extracted to construct a coarse attention map. (10) (11) In the formula, Represents a sequence of lexical terms. Indicates category terms, Indicates the first Image block words, Indicates the first Image blocks, Indicates shared ownership Image blocks, Represents the attention matrix. This indicates the self-attention mechanism. Represents the query matrix. Represents the key matrix, Representative value matrix, Indicates matrix transpose. Represents the dimension of the key matrix vector. This represents category-based attention, reflecting the interaction between category-based terms and all image-based terms. Indicates the first Image block word attention.

[0069] C4: Denoising coarse attention maps using a linear piecewise function: (12) In the formula, This represents the attention map after denoising. This represents a coarse attention map. The attention maps before and after denoising are shown below. Figure 5 As shown.

[0070] C5: Pre-save the attention map as a .pt file.

[0071] S2. Construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. Input the dual-angle enhanced feature pyramid and the attention map into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid.

[0072] Optionally, in step S2 above, inputting the dual-angle enhanced feature pyramid and attention map into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid includes: Interpolation is used to transform the attention map into a multi-scale attention map. The multi-scale attention map and the dual-angle enhanced feature pyramid are then effectively combined through a residual weighted adjustment mechanism to obtain the multi-angle enhanced feature pyramid.

[0073] In one feasible implementation, since the dual-angle augmentation feature pyramid is multi-scale, it is necessary to align the attention map and feature map in the scale space, and choose to use interpolation operations to transform the attention map into a multi-scale attention map.

[0074] Furthermore, the residual weighted modulation mechanism, which uses attention maps to reweight the features of the dual-angle enhanced feature pyramid, can enhance the feature discriminativeness of the dual-angle enhanced feature pyramid and highlight the foreground features of the remote sensing target: (13) (14) In the formula, This represents the multi-angle enhanced feature pyramid modulated by the attention map. Representation layer normalization, The first part represents the dual-angle enhanced feature pyramid. Layer feature map, These represent attention modulation coefficients, used to control the modulation intensity of the attention map on the feature map. Indicates and the Attention maps corresponding to layer feature maps Indicates linear interpolation. This represents the original attention map.

[0075] This invention, based on the dual-angle enhanced feature pyramid, leverages the advantage of the attention mechanism in self-supervised visual large models for modeling the correlation of image regions. By combining attention maps with a residual weighted adjustment mechanism, it strengthens the feature regions in the dual-angle enhanced feature pyramid that are highly correlated with the target semantics, thereby improving the discriminativeness of the feature pyramid and ultimately constructing a multi-angle enhanced feature pyramid.

[0076] S3. Input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing images. Fine-tune the region candidate network and the target detection head according to the corresponding annotations of the small sample remote sensing images. Obtain the target detection network of the small sample remote sensing images based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head.

[0077] In one feasible implementation, a small-sample remote sensing image target detection network based on a multi-angle enhanced feature pyramid is constructed, such as... Figure 6 As shown, the multi-angle enhanced feature pyramid target detection network for small-sample remote sensing images includes: a self-supervised visual large model, a multi-angle enhanced feature pyramid, an attention map-enhanced feature pyramid discriminative module, a region candidate network, and a target detection head. Among them, the multi-angle enhanced feature pyramid network includes a dual-angle enhanced feature pyramid with semantic generalization and scale robustness, as well as a feature discriminative module that uses attention maps to enhance the dual-angle enhanced feature pyramid. The dual-angle enhanced feature pyramid includes semantic generalization enhancement based on a self-supervised visual large model and scale robustness enhancement based on a dynamic scale reconstruction module.

[0078] Among them, the self-supervised visual large model is used to output general visual features and attention maps of remote sensing images.

[0079] A multi-angle enhanced feature pyramid network is used to provide high-quality feature support for the region recommendation network, generate more accurate candidate recommendation regions, solve the base class shift problem in traditional methods, and effectively highlight remote sensing foreground target regions.

[0080] Specifically, the dynamic scale reconstruction module is used to reconstruct the single-scale features output by the large-scale visual model to generate multi-scale features when the target category in the output image changes in size, shape, and category. The feature pyramid network is used to fuse multi-scale features, including fusing the texture and edge information of shallow features and the semantic information of deep features, and fusing the small target information of large-scale feature maps and the large target information of small-scale feature maps. The self-supervised large-scale visual model, the dynamic scale reconstruction module, and the feature pyramid network are combined to form a dual-angle enhanced feature pyramid.

[0081] Furthermore, a residual weighted modulation mechanism is used to fuse the attention map and the feature pyramid. The attention map-enhanced feature pyramid discriminative module is used to reweight the features of the feature pyramid, highlighting foreground image features of remote sensing targets and suppressing background information features.

[0082] The region recommendation network (region candidate network) is used to process the multi-angle enhanced feature pyramid, initially generate candidate recommendation regions, and provide region of interest features for the detection head.

[0083] The target detection head is used to process features of the region of interest and achieves accurate localization and recognition of targets in remote sensing images through classification and localization loss optimization.

[0084] Multi-angle enhanced feature pyramids provide high-quality feature support for the region candidate network. End-to-end target detection network pre-training is achieved through classification and regression loss optimization of the detection head. Fine-tuning of the network using a small number of labeled samples enables target detection in small-sample remote sensing images.

[0085] Specifically, the feature pyramid is enhanced from multiple angles and input into subsequent detection model components, and optimized using the cross-entropy loss function and the Smooth L1 loss function, including: Multi-angle enhanced feature pyramids are input into the candidate region network to perform preliminary foreground-background separation of detection boxes and coarse localization of the detection boxes: (15) (16) (17) (18) (19) in, This represents a multi-angle enhanced feature pyramid. Indicates the first Layer feature map, and These represent the height and width of the feature map, respectively. Indicates the number of channels. Indicates through sharing Intermediate features extracted by convolution Represents the ReLU activation function. Indicates the first The foreground / background classification probabilities of all anchor boxes on the layer are determined by... The convolutional layer is followed by the Softmax function to obtain the result. This represents the classification branch of the region recommendation network. This represents the regression branch of the region recommendation network. This represents the total loss of the region recommendation network. Indicates the first Each anchor box predicts the probability that it contains a foreground target. Indicates the first Layer feature map, Indicates the first Each anchor box contains the ground truth value of the foreground target. Indicates the first The predicted coordinates of each anchor frame. Indicates the first The true coordinates of each anchor frame. This represents the corresponding bounding box regression offset, determined by another... The convolutional layer predicts that, Represents the classification loss function. Represents the regression loss function. and These represent the normalization factors for classification and regression, respectively. The weighting coefficients are used to balance the classification loss and regression loss.

[0086] Furthermore, the candidate region network outputs 2000 candidate regions, and features are extracted using region-of-interest pooling. These features are then fed into the detection system for fine-grained classification and regression. (20) in, This represents the total number of samples involved in the classification loss calculation. Indicates the number of positive samples. Indicates the first Predicted category probabilities for each candidate region This indicates its corresponding real category label. This represents the regression parameters of the predicted bounding box. This represents the corresponding true regression target. Represents the classification loss function. Represents the regression loss function. The weighting coefficients are used to balance the classification loss and regression loss.

[0087] The above process completes the pre-training of the target detection network. A small number of labeled samples are used to fine-tune the region recommendation network and detection head of the detection model, thus realizing a target detection network for small-sample remote sensing images.

[0088] The experimental setup of this invention includes: In the training phase of the small-sample remote sensing image target detection method based on multi-angle enhanced feature pyramids, the two-stage target detection model Faster-RCNN was selected as the basic detection framework. Throughout the training process, the parameters of the large visual model remained fixed. In the training configuration, stochastic gradient descent was chosen as the optimizer, with an initial learning rate of 0.01, momentum of 0.9, and weight decay coefficient of 0.0001. The learning rate scheduling strategy combined warm-up and step-by-step decay: warm-up was performed for the first 200 iterations, followed by step-by-step decay at iterations of 24,000, 36,000, and 48,000, for a total of 63,000 iterations. For data preprocessing, image scaling to 800×800, random horizontal flipping with a probability of 0.5, normalization, and padding were used for data augmentation. Finally, the detection model was fine-tuned on new class datasets, with the number of small samples used for fine-tuning set to 3, 10, and 20, respectively.

[0089] In this embodiment, the multi-angle enhanced feature pyramid small sample remote sensing image target detection network, which has been trained, is suitable for scenarios where it is impossible to obtain a large amount of labeled data.

[0090] To verify the effectiveness of the few-sample remote sensing image target detection method based on multi-angle enhanced feature pyramids provided in this invention, this embodiment uses the DIOR and NWPU VHR datasets to test the performance of the method. To improve the reliability of the test, four different base / new class partitions were performed on the DIOR dataset (as shown in Table 2), and two different base / new class partitions were performed on the NWPUVHR dataset (as shown in Table 3). The final performance was evaluated using the mean accuracy (mAP) metric: mAP reflects the comprehensive detection capability of the few-sample remote sensing image target detection network based on multi-angle enhanced feature pyramids across various categories.

[0091] Table 2. Four different new / base class partitioning methods for the DIOR dataset.

[0092] Table 3 Two different new / base class partitioning methods for the NWPU VHR dataset

[0093] To further verify the performance of the method described in this embodiment, a representative small-sample remote sensing image target detection method was selected for comparison. The experimental results are shown in Tables 4 and 5, respectively.

[0094] Table 4. Experimental results of the present invention and comparative experimental methods on the DIOR dataset.

[0095] Table 5. Experimental results of the present invention and comparative experimental methods on the NWPU VHR dataset.

[0096] Tables 4 and 5 show that this embodiment achieves the best experimental results under different small-sample dataset partitioning conditions. For example, on the DIOR dataset, under the 20-shot condition setting of partitioning method 1, the present invention achieves 65.70% on all categories, significantly outperforming the second-best method DI-FSdet's 62.31%. Under the 3-shot condition setting of partitioning method 4, the present invention achieves 23.70% for the Novel class, far exceeding the highest of 16.74% among the comparative methods. On the NWPUVHR dataset, under the 20-shot condition setting of partitioning method 2, the present invention achieves 86.78% on all categories, significantly outperforming the second-best method DI-FSdet's 78.63%. Furthermore, as the number of labeled samples increases from 3-shot to 20-shot, the performance of the present invention shows a stable and significant upward trend. Notably, in all experimental settings, the experimental results for the new class are significantly improved, which fully verifies that the proposed method has strong generalization ability in small-sample scenarios.

[0097] Self-supervised large-scale visual models can learn feature representations that evolve layer by layer from local structure to global semantics. However, the semantic and spatial structural information contained in the output features of different encoder layers varies significantly. In target detection of small-sample remote sensing images, feature pyramids fuse multiple layers of features to balance semantic discrimination and spatial localization capabilities, and their performance is highly dependent on the selection of the input feature hierarchy. Therefore, using the DIOR dataset, a series of ablation experiments were designed around different encoder output selections under the first class division condition. The system uses shallow, medium, and deep features to construct feature pyramids and their impact on target detection performance. The experimental results are shown in Table 6.

[0098] Table 6 Performance indicators of different encoder output layers under partitioning method 1

[0099] As shown in Table 6, in the few-shot object detection task of the DIOR dataset, the model performance is significantly correlated with the selection strategy of the encoder output layer. When the partitioning method is 1, as the number of encoder output layers increases, the overall detection accuracy of the model shows a steady upward trend under different few-shot settings such as 3-shot, 5-shot, 10-shot, and 20-shot. For example, when the encoder output layer gradually transitions from a shallow combination [7,9,11,13] to a deeper combination [20,21,22,23], the model's accuracy under the 20-shot setting increases from 62.91% to 65.70%. It is worth noting that when the encoder output layer selection strategy adopts a combination of shallow and deep features (such as [11,13,19,23]), the model achieves the best or second-best performance under all few-shot settings, with accuracies of 65.74% and 66.96% under the 10-shot and 20-shot settings, respectively, which are significantly better than other single-depth feature combinations. A deeper analysis reveals that this phenomenon closely aligns with the characteristics of deep convolutional neural networks such as ResNet101. In these networks, shallow features often contain low-level visual information such as image edges and textures, while deep features correspond to higher-level semantic information with greater semantic abstraction. By effectively fusing features from different levels, the model can simultaneously utilize both low-level detail information and high-level semantic information, thereby more robustly capturing the rich features of the target object under limited sample conditions, ultimately achieving a significant improvement in detection performance.

[0100] As can be seen from the above description, the method of the present invention greatly reduces the number of parameters that need to be trained and achieves a significant performance improvement in the small sample fine-tuning stage. Therefore, the present invention further explores the use of the method proposed in the present invention for cross-domain small sample remote sensing image target detection.

[0101] Table 7. Cross-domain small sample target detection performance of the present invention

[0102] As can be seen from Table 7, the method of this embodiment achieves the best results in cross-domain small sample target detection, further illustrating the effectiveness of the method of this embodiment.

[0103] S4. Acquire the remote sensing image to be detected, input the remote sensing image into the small sample remote sensing image target detection network, and obtain the target detection result.

[0104] This invention enhances the semantic generalization of the feature pyramid based on the output features of a self-supervised large-scale visual model, improves the scale robustness of the feature pyramid based on a plug-and-play upsampling module, and enhances the discriminativeness of the feature pyramid based on attention maps. Specifically, in enhancing semantic generalization, it fully utilizes the general feature representation capabilities obtained through self-supervised training of a large-scale visual model on a massive dataset, addressing the base class feature shift problem in traditional methods. In enhancing scale robustness, it employs a plug-and-play upsampling module to construct a dynamic scale reconstruction module that requires no training, addressing the issue of single-scale output features from the large-scale visual model. In enhancing discriminativeness, it leverages the advantage of the self-supervised large-scale visual model's attention mechanism in characterizing the correlation between different regions of the image, assigning greater weights to regions in the feature pyramid that are highly semantically related to the target through attention maps. Through these designs, this invention constructs a multi-angle enhanced feature pyramid, providing high-quality feature support for candidate region recommendation networks. Combined with the classification and localization loss functions of the detection head, it achieves end-to-end target detection in small-sample remote sensing images. This invention effectively improves the generalization and robustness of the target detection network, thereby significantly improving the detection accuracy of the model with a small number of labeled samples.

[0105] In this embodiment of the invention, the problem of base class feature shift in traditional small-sample remote sensing image target detection methods based on ResNet backbone networks is solved by utilizing the general feature output capability of self-supervised large visual models.

[0106] By introducing a dynamic scale reconstruction module, the problem of single-scale output features in large visual models is addressed. This module is lightweight and class-independent, providing multi-scale feature support for feature fusion, enabling the detection of targets of different sizes.

[0107] Based on the self-attention mechanism in the large visual model, an attention map reflecting the foreground target is constructed. Through the residual weighted modulation mechanism, spatial prior information is provided for the dual-angle enhanced feature pyramid, thereby improving feature discriminability.

[0108] The detection method is highly flexible and easily expandable, making it applicable to target detection in various small-sample remote sensing images, and can effectively improve the detection accuracy and robustness of new categories.

[0109] Figure 7 This is a block diagram illustrating a small-sample target detection device based on a multi-angle enhanced feature pyramid, according to an exemplary embodiment. The device is used in a small-sample target detection method based on a multi-angle enhanced feature pyramid. (Refer to...) Figure 7 The device includes a dual-angle enhanced feature pyramid module 310, a multi-angle enhanced feature pyramid module 320, a fine-tuning module 330, and a target detection module 340. Among them: The dual-angle enhanced feature pyramid module 310 is used to acquire small sample remote sensing images. The dual-angle enhanced feature pyramid module is constructed based on a self-supervised visual large model, a dynamic scale reconstruction module, and a feature pyramid network. The small sample remote sensing images are input into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map.

[0110] The multi-angle enhanced feature pyramid module 320 is used to construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. The dual-angle enhanced feature pyramid and the attention map are input into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid.

[0111] The fine-tuning module 330 is used to input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing image. The region candidate network and the target detection head are fine-tuned according to the annotations corresponding to the small sample remote sensing image. The target detection network of the small sample remote sensing image is obtained based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head.

[0112] The target detection module 340 is used to acquire the remote sensing image to be detected, input the remote sensing image into the small sample remote sensing image target detection network, and obtain the target detection result.

[0113] In this embodiment of the invention, the problem of base class feature shift in traditional small-sample remote sensing image target detection methods based on ResNet backbone networks is solved by utilizing the general feature output capability of self-supervised large visual models.

[0114] By introducing a dynamic scale reconstruction module, the problem of single-scale output features in large visual models is addressed. This module is lightweight and class-independent, providing multi-scale feature support for feature fusion, enabling the detection of targets of different sizes.

[0115] Based on the self-attention mechanism in the large visual model, an attention map reflecting the foreground target is constructed. Through the residual weighted modulation mechanism, spatial prior information is provided for the dual-angle enhanced feature pyramid, thereby improving feature discriminability.

[0116] The detection method is highly flexible and easily expandable, making it applicable to target detection in various small-sample remote sensing images, and can effectively improve the detection accuracy and robustness of new categories.

[0117] Figure 8 This is a schematic diagram of the structure of a small-sample remote sensing image target detection device provided in an embodiment of the present invention, as shown below. Figure 8 As shown, a small-sample remote sensing image target detection device may include the above-mentioned Figure 7 The illustrated small-sample target detection device is based on a multi-angle enhanced feature pyramid. Optionally, the small-sample remote sensing image target detection device 410 may include a first processor 2001.

[0118] Optionally, the small sample remote sensing image target detection device 410 may also include a memory 2002 and a transceiver 2003.

[0119] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0120] The following is combined Figure 8 A detailed description of each component of the small-sample remote sensing image target detection device 410 is provided below: The first processor 2001 is the control center of the small-sample remote sensing image target detection device 410. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0121] Optionally, the first processor 2001 can perform various functions of the small sample remote sensing image target detection device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0122] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 8 CPU0 and CPU1 are shown in the diagram.

[0123] In a specific implementation, as one example, the small-sample remote sensing image target detection device 410 may also include multiple processors, for example... Figure 8 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0124] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0125] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or exist independently, and may be connected via the interface circuit of the small sample remote sensing image target detection device 410. Figure 8 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0126] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0127] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 8 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0128] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected to the interface circuit of the small sample remote sensing image target detection device 410. Figure 8 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0129] It should be noted that, Figure 8 The structure of the small sample remote sensing image target detection device 410 shown does not constitute a limitation on this router. Actual knowledge structure recognition devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0130] Furthermore, the technical effect of the small sample remote sensing image target detection device 410 can be referred to the technical effect of the small sample target detection method based on multi-angle enhanced feature pyramid described in the above method embodiments, and will not be repeated here.

[0131] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0132] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0133] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0134] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0135] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0136] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0137] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0138] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0139] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0140] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0141] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0142] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A small-sample target detection method based on multi-angle enhanced feature pyramids, characterized in that, The method includes: S1. Obtain small sample remote sensing images. Construct a dual-angle enhanced feature pyramid module based on a self-supervised visual large model, dynamic scale reconstruction module and feature pyramid network. Input the small sample remote sensing images into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map. S2. Construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. Input the dual-angle enhanced feature pyramid and the attention map into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid. S3. Input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing images. Fine-tune the region candidate network and the target detection head according to the annotations corresponding to the small sample remote sensing images. Obtain the target detection network of the small sample remote sensing images based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head. S4. Acquire the remote sensing image to be detected, input the remote sensing image into the small sample remote sensing image target detection network, and obtain the target detection result.

2. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 1, characterized in that, The step S1 involves inputting small sample remote sensing images into a dual-angle enhanced feature pyramid module to obtain a dual-angle enhanced feature pyramid, including: By inputting small sample remote sensing images into a self-supervised visual large model, visual feature maps of encoders at different levels in the self-supervised visual large model are obtained. Visual feature maps from encoders at different levels and remote sensing images are input into the dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps. Multi-scale feature maps are input into a feature pyramid network for multi-scale feature fusion to obtain a dual-angle enhanced feature pyramid.

3. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 2, characterized in that, The step of inputting visual feature maps from different levels of encoders and remote sensing images into a dynamic scale reconstruction module for feature reconstruction to obtain multi-scale feature maps includes: Shallow features from the visual feature maps of different encoder levels are input into the plug-and-play feature upsampling submodule to obtain upsampled feature maps. Deep features from the visual feature maps of different encoder levels are input into the downsampling submodule to obtain the downsampled feature map; A multi-scale feature map is obtained based on the upsampled feature map, the visual feature map, and the downsampled feature map.

4. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 1, characterized in that, The process of constructing the attention map includes: By utilizing the self-attention mechanism of a self-supervised large visual model, a coarse attention map is constructed by calculating the attention between category terms and image patch embedding features. The coarse attention map is then denoised using a linear piecewise function to obtain the final attention map.

5. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 4, characterized in that, The process of constructing the coarse attention map is shown in equations (1)-(2): （1） （2） In the formula, Represents a sequence of lexical terms. Indicates category terms, Indicates the first Image block words, Indicates the first Image blocks, Indicates shared ownership Image blocks, Represents the attention matrix. This indicates the self-attention mechanism. Represents the query matrix. Represents the key matrix, Representative value matrix, Indicates matrix transpose. Represents the dimension of the key matrix vector. Indicates category attention, Indicates the first Image block word attention.

6. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 1, characterized in that, In step S2, the dual-angle enhanced feature pyramid and attention map are input into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid, including: Interpolation is used to transform the attention map into a multi-scale attention map. The multi-scale attention map and the dual-angle enhanced feature pyramid are then effectively combined through a residual weighted adjustment mechanism to obtain the multi-angle enhanced feature pyramid.

7. The small-sample target detection method based on multi-angle enhanced feature pyramid according to claim 6, characterized in that, The process of obtaining the multi-angle enhanced feature pyramid is shown in equations (3)-(4): （3） （4） In the formula, This represents the multi-angle enhanced feature pyramid modulated by the attention map. Representation layer normalization, The first part represents the dual-angle enhanced feature pyramid. Layer feature map, Represents the attention modulation coefficient. Indicates the first Attention maps corresponding to layer feature maps Indicates linear interpolation. This represents the original attention map.

8. A small-sample target detection device based on a multi-angle enhanced feature pyramid, wherein the small-sample target detection device based on a multi-angle enhanced feature pyramid is used to implement the small-sample target detection method based on a multi-angle enhanced feature pyramid as described in any one of claims 1-7, characterized in that, The device includes: The dual-angle enhanced feature pyramid module is used to acquire small sample remote sensing images. The dual-angle enhanced feature pyramid module is constructed based on a self-supervised visual large model, a dynamic scale reconstruction module, and a feature pyramid network. The small sample remote sensing images are input into the dual-angle enhanced feature pyramid module to obtain the dual-angle enhanced feature pyramid and attention map. The multi-angle enhanced feature pyramid module is used to construct a multi-angle enhanced feature pyramid module based on the dual-angle enhanced feature pyramid module and the residual weighted adjustment mechanism. The dual-angle enhanced feature pyramid and the attention map are input into the multi-angle enhanced feature pyramid module to obtain the multi-angle enhanced feature pyramid. The fine-tuning module is used to input the multi-angle enhanced feature pyramid into the region candidate network and the target detection head to obtain the target localization and classification results of the small sample remote sensing images. The region candidate network and the target detection head are fine-tuned according to the corresponding annotations of the small sample remote sensing images. The target detection network of the small sample remote sensing images is obtained based on the multi-angle enhanced feature pyramid module and the fine-tuned region candidate network and target detection head. The target detection module is used to acquire remote sensing images of the targets to be detected, input the remote sensing images into a small sample remote sensing image target detection network, and obtain the target detection results.

9. A target detection device for small sample remote sensing images, characterized in that, The small-sample remote sensing image target detection device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images