A small target detection countermeasure method based on three-branch dual-domain refinement processing

Abstract

The application discloses a small target detection anti-attack method based on three-branch double-domain fine processing, and belongs to the technical field of computer vision security. In view of the problems that small target features are scarce and easy to be annihilated under the attack, the application proposes a RESO network: a three-branch parallel topology of details, contexts and semantics is constructed in a spatial domain, multi-granularity features are extracted by using different receptive fields, and non-robust noise is recalibrated at a pixel level by using a soft gating mechanism based on Sigmoid; in a frequency domain, a frequency spectrum is divided into three bands of high, medium and low frequencies according to a frequency radius, and differential enhancement is performed on the small target band distribution characteristics; finally, a cross-domain intersection fusion mechanism is used to realize dynamic aggregation of spatial details and frequency structures by using adaptive attention weights; the application breaks through the limitation of traditional defense methods on rough small target feature extraction, and significantly improves the detection accuracy and anti-attack robustness of small targets on challenging data sets such as VisDrone and AI-TOD.

Description

Technical Field

[0001] This invention relates to the field of computer vision security technology, and in particular to a small target detection adversarial defense method based on three-branch dual-domain refined processing. Background Technology

[0002] Adversarial defense for small object detection refers to defensive techniques and strategies for object detection models when facing adversarial attacks, especially in scenarios involving the detection of small objects. Adversarial attacks introduce perturbations into the input image that are imperceptible to the human eye, causing deep learning models to make incorrect predictions. Small objects themselves lack feature information, making the impact of adversarial perturbations more significant. Defense mechanisms need to balance sensitivity to small objects with robustness against adversarial attacks.

[0003] Traditional defense methods are primarily designed for large targets or image classification tasks, and their effectiveness is limited in scenarios with small targets. Traditional methods suffer from at least the following drawbacks: (1) Limitations of space domain defense Traditional defense methods primarily operate in the spatial domain, which presents the following problems: Sensitive to high-frequency disturbances: Adversarial attacks often use high-frequency components for disturbances, which are difficult to remove effectively using traditional spatial domain filtering; Small target feature loss: While suppressing adversarial perturbations, it is easy to accidentally damage the effective features of small targets;

[0004] High computational overhead: Adversarial training and other methods require a large amount of computational resources.

[0005] (2) Insufficient frequency domain processing Most existing frequency domain defense methods use simple low-pass / high-pass filtering, lack differentiated processing strategies for different frequency bands, fail to consider the complementarity of spatial and frequency domain features, and are difficult to accurately distinguish between countermeasures and real small target features.

[0006] (3) Lack of cross-domain information fusion Existing methods do not fully utilize the correlation information between spatial domain details and frequency domain structure, resulting in limited defense effectiveness.

[0007] The shortcomings of existing methods are listed in Table 1 below: Table 1. Main shortcomings of traditional defense methods Defense type Representative method Main defects Input preprocessing Gaussian blur, JPEG compression Simultaneously suppresses small target features, resulting in a low defense rate. Combat training PGD ​​competitive training High computational cost, effective only against specific attacks Patch detection PAD, SAC, Jedi Poor detection of natural appearance patches; FID distance analysis shows that defense is difficult. Authentication defense ObjectSeeker Strict threat model assumptions are required; computational costs increase exponentially with the number of patches. In addition, the detection challenges for the VisDrone and AI-TOD datasets include: (1) VisDrone Scale extremism: Many targets occupy only a few pixels, and the information on deep feature maps approaches zero; High target density: A single image can contain 200+ targets, with severe overlap of bounding boxes; Complex background: numerous interfering factors such as buildings, shadows, and vegetation; Category imbalance: The number of pedestrian and vehicle samples far exceeds that of other categories.

[0008] (2) AI-TOD Extremely lacking in features: the target is represented by only a few pixels, lacking texture and structural information; The positioning accuracy requirements are stringent: a deviation of just a few pixels means detection failure; IoU metric fails: Traditional IoU is extremely sensitive to the positional deviation of small targets; Label assignment is difficult: there is a severe imbalance between positive and negative samples. Summary of the Invention

[0009] The purpose of this invention is to provide a small target detection adversarial defense method based on three-branch dual-domain refined processing to solve the above-mentioned problems.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A small target detection adversarial defense method based on three-branch dual-domain refined processing, the method being as follows: A three-branch parallel topology of detail branch, context branch and semantic branch is constructed in the spatial domain. Multi-granularity features are extracted using different receptive fields, and non-robust noise is recalibrated at the pixel level through a soft gating mechanism based on Sigmoid. In the frequency domain, the spectrum is divided into three independent sub-bands—high frequency, mid frequency, and low frequency—based on the frequency radius, and differentiated enhancement is performed according to the frequency band distribution characteristics of small targets. Finally, a cross-domain intersection and fusion mechanism is used to achieve dynamic aggregation of spatial details and frequency structure through adaptive attention weights.

[0011] As a preferred technical solution, the method specifically includes the following steps: S1, Extract multi-scale feature maps from the input image, and select a specific feature layer containing information about small targets. (e.g., P3 layer) Input RESO module; the "RESO module" described in this invention is a three-branch module; S2, perform three-branch fine-grained processing of the spatial domain: construct detail branch, context branch and semantic branch; the detail branch uses small receptive field convolution to capture local texture, the context branch uses dilated convolution to capture the background around the target, and the semantic branch uses global average pooling to extract category priors; The problem of insufficient feature information for small targets is addressed through three dimensions, as shown in Table 2 below.

[0012] Table 2. The role of spatial domain three-branch refinement processing Dimension Corresponding branch Problems to be solved Detail Dimensions Small receptive field convolution branch Small target boundary features are easily lost. Context Dimension dilated convolution branch The small goal lacks information about the surrounding environment. Semantic dimension Global pooling branches Weak ability to classify small targets S3, Perform robust feature selection in the spatial domain: In each of the three branches of the spatial domain, the channel importance score is evaluated using convolution, and a three-branch mask is generated using the Sigmoid activation function. Pixel-level calibration is performed on intermediate features to obtain spatially refined features. ; S4, Perform frequency domain three-branch refinement processing: Transform the features to the frequency domain using discrete Fourier transform, and refine the features according to the preset low-frequency radius. and high frequency radius The spectrum is divided into three independent sub-bands: low frequency, mid frequency, and high frequency. Frequency domain characteristics are used to achieve anti-disturbance separation and small target feature enhancement. Specifically, its effects are shown in Table 3.

[0013] Table 3. The role of frequency domain three-branch refinement processing frequency band Processing strategy Purpose low frequency Retain or moderately enhance Provide background context information Intermediate frequency Key enhancement Small target core feature area high frequency Suppression or filtering Noise Removal and Disturbance Combat S5, Perform frequency domain differentiation enhancement: For different frequency band characteristics, a robust screening mechanism is used to extract stable components within each frequency band, which are then concatenated into refined frequency domain features after inverse transformation. ; S6, Perform bi-domain feature intersection fusion: Utilize a cross-domain attention mechanism to calculate the correlation matrix between spatial domain details and frequency domain structure. : , In the formula, Represents a spatial domain query vector. Definition: From the perspective of spatial refinement features The query representation extracted from; By using 1×1 convolutional or fully connected layers, refined features from the spatial domain are projected onto the query space; Dimensions: [N, C'], where N = H×W (number of pixel positions), and C' is the query feature dimension; Semantic meaning: It represents "what information you want to query" for each pixel position in the spatial domain, that is, the query request sent from the spatial domain to the frequency domain.

[0014] This represents the transpose of the frequency domain key vector. Definition: Refining features from the frequency domain The extracted key representation is then transposed. Dimensions: [C', N] (after transpose); Semantic meaning: Represents what information each position in the frequency domain "possesses", that is, the response index of the frequency domain to the spatial domain query; The purpose of transpose: to make × It can calculate the similarity matrix (attention score) between each location in the spatial domain and each location in the frequency domain. Represents the dimension of the key vector. Definition: Key vector K freq The feature dimension (number of channels C'); when When the value is large, the value of the dot product Q×K^T will become very large; Excessively large values ​​can cause Softmax to enter the saturation region, and the gradient will approach zero. Divide by Normalize the dot product to a reasonable range to prevent gradient vanishing; S7, Perform adaptive feature integration: through learnable fusion weight factors A weighted integration of interactive features and original frequency domain features is achieved to generate the final refined features of small targets. : , In the formula, Representing refined features from the frequency domain The extracted value vector is generated through a linear transformation. = Linear_V( The frequency domain features are projected onto the value space through 1×1 convolutional or fully connected layers. This represents the "actual information content to be conveyed" at each location in the frequency domain, that is, the substantive response of the frequency domain to a query in the spatial domain.

[0015] The spatial domain initiates a query: "What information do I need?"

[0016] The frequency domain provides an index: "What information do I have?"

[0017] The frequency domain provides the following information: "This is the specific information content." S8, Input the detection head, and it will output the category label and bounding box regression results of the small target.

[0018] As a preferred technical solution In step S2, the dilation rate of the detail branch convolution is set to 1, and the dilation rate of the context branch dilated convolution is set to 2, so as to achieve complementary modeling of small target features with different spatial spans.

[0019] As a further preferred technical solution In step S3, the spatial domain three-branch mask We obtain the following formula: , In the formula, These are the intermediate features of the corresponding branches after convolution transformations through different receptive fields. This is a soft-gated activation function used to generate functions in the range of... A continuous weight distribution is used to achieve nonlinear preservation of weak signals from small targets.

[0020] As a further preferred technical solution In step S4, the frequency band allocation strategy is aimed at Feature map of resolution, setting low-frequency radius High frequency radius This is to deconstruct the target's contour, structure, and edge information.

[0021] As a further preferred technical solution In step S4, the spectrum-independent subband division is performed using an indicator function. The implementation and division logic are as follows: , The frequency radius in the spectrum is... and These are the preset low-frequency and high-frequency cutoff thresholds, respectively. For a feature map spectrum of size N×N, the spectral subband division is determined by an indicator function. accomplish.

[0022] Define frequency distance function Let the center of the spectrum be (N / 2, N / 2), and the distance from the frequency point (u, v) to the center be defined as: d(u,v)=(u−2N) 2 +(v−2N) 2 For example, for an 80×80 feature map (P3 layer): the spectral center is (40, 40).

[0023] Table 4 Different Subband Types Subband type Indicator function definition meaning Low-frequency subband <![CDATA[I low ={d(u,v)≤ rlow }]]> Low-frequency region closer to the center Intermediate frequency subband <![CDATA[I mid ={r low <d(u,v)≤r high }]]> Mid-frequency range, between low and high frequencies High-frequency subband <![CDATA[I high ={d(u,v)>r high }]]> High-frequency regions far from the center

[0024] Functional deentanglement of global contour, local structure, and edge details is achieved by radially cutting the spectrogram.

[0025] As a further preferred technical solution In step S7, the adaptive feature integration includes the following steps: first, utilizing interaction features through Convolution and the Sigmoid function calculate the fusion weight factor along the channel dimension. : , Then based on the fusion weight factor The final refined features are obtained by weighted summation of spatial interaction information and original frequency domain features. : , In the formula, The Hadamard product is a mechanism used to dynamically balance spatial localization accuracy with frequency structure robustness based on the feature distribution of the input samples.

[0026] Compared with existing technologies, the advantages of this invention are as follows: The RESO method of this invention, through refined recalibration of two domains and three branches, not only improves the APS index under clean samples, but also shows a lower mAP decrease under various perturbation intensities; This invention overcomes the limitation of traditional defense methods in coarse feature extraction of small targets, and significantly improves the detection accuracy and adversarial robustness of small targets on challenging datasets such as VisDrone and AI-TOD. On the AI-TOD dataset, it not only improves the detection performance of clean samples, but also enhances the robustness of adversarial samples, and significantly reduces the mAP decrease. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a flowchart of a small target detection technique based on a three-branch structure. Detailed Implementation

[0028] To explain the technical content, objectives, and effects of the present invention in detail, the following specific embodiments are provided to further illustrate the content of the present invention. However, the content of the present invention is far more than the following examples.

[0029] Example 1 A small target detection adversarial defense method based on three-branch dual-domain refined processing, see [link to relevant documentation]. Figure 1 and Figure 2 It includes the following steps: S1, Input: Input the image to be processed and extract the P3 layer feature map through the YOLOv8 backbone network; S2, spatial domain three-branch parallel processing (left branch), includes: Detail branch: Extracts detailed information such as texture and edges. Contextual branching: Extracting contextual information such as the surrounding environment and scene layout. Semantic branching: Extracting high-level semantic information such as category attributes and spatial relationships. The three branches extract robust features using the Sigmoid function and are then concatenated and fused. The main steps for each branch are as follows: (1) Feature transformation First, the spatial domain features of the input Perform feature transformation to generate intermediate features The feature transformation methods differ slightly between different branches: Detailed branch intermediate features Employing convolutional kernels with small receptive fields to accurately capture detailed features: , Intermediate features of context branches Employing dilated convolutions with moderate receptive fields to capture contextual features: , Semantic branch intermediate features Semantic features are extracted using global receptive field operations. , in, This is a global average pooling operation.

[0030] (2) Importance assessment The importance of each feature channel is evaluated using a 1×1 convolutional layer to generate an importance score. : , in, .

[0031] (3) Robust feature selection The importance score is normalized using the Sigmoid function to generate a robust feature mask. : , The Sigmoid function is defined as follows: , (4) Feature recalibration Generate robust feature mask With intermediate features By performing channel-by-channel multiplication, robust features are obtained. : , (5) Residual connection To preserve information from the original features, residual connections are introduced: , in, This is the final output of the corresponding branch.

[0032] S3, frequency domain three-branch parallel processing (right branch), includes: High-frequency branch: processes edge and texture information. Intermediate frequency branch: processes local structural information. Low-frequency branch: processes global contour information. The three branches are combined and fused through DFT transformation, frequency band extraction, robust feature selection, and finally splicing. The main steps for each branch are as follows: (1) DFT transformation First, the spatial domain features of the input Perform a Discrete Fourier Transform (DFT) to obtain the frequency domain features. : , (2) Frequency band extraction Based on the frequency band division strategy, the spectra of the high-frequency band, mid-frequency band, and low-frequency band are extracted respectively: , , , in, This is an indicator function; it returns 1 if the condition is met, and 0 otherwise. , , These are the frequency spectra for the high-frequency band, mid-frequency band, and low-frequency band, respectively.

[0033] (3) Frequency domain feature processing The spectra of different frequency bands are processed to generate intermediate frequency domain features: High-frequency branching: Enhances edge and texture information. , Mid-frequency branch: Enhances local structural information: , Low-frequency branch: Enhances global contour information. , in, , , This represents the amplitude spectrum of the corresponding frequency band; , , This represents the phase spectrum of the corresponding frequency band.

[0034] (4) Importance assessment The importance of each frequency band component is evaluated using a 1×1 convolutional layer to generate an importance score. : , in, .

[0035] (5) Robust frequency domain feature screening The importance score is normalized using the Sigmoid function to generate a robust frequency domain feature mask. : , (6) Frequency domain feature recalibration The generated robust frequency domain feature mask With intermediate frequency domain characteristics By performing channel-by-channel multiplication, robust frequency domain features are obtained. : , (7) Inverse Transformation The processed frequency domain features are then transformed back to the spatial domain using the inverse discrete Fourier transform (IDFT): , (8) Residual connection To preserve information from the original features, residual connections are introduced: , in, This is the final output of the corresponding branch.

[0036] S4, Dual-domain feature intersection and fusion: The refined features of the spatial domain and frequency domain are interactively fused. Adaptive fusion is achieved through channel attention mechanisms and learnable weights; After obtaining the robust frequency domain features extracted from the three branches, the RESO method generates refined frequency domain features through concatenation and fusion. The mathematical expression for concatenation and fusion is: , in, This is for refining the frequency domain features after splicing; For channel splicing operations; , , These are robust frequency domain features extracted from the three branches.

[0037] For input size The features, the size of the spliced ​​features are .

[0038] S5, Output: Input the fused features into the YOLOv8 detection head and output the final detection result.

[0039] The fine-grained enhancement phase of the RESO module corresponds to steps S2-S3 above. Specifically... Figure 3 In the first step, the feature map input to the P3 layer is preprocessed, including defining the convolution kernel parameters for the three parallel branches in the spatial domain (3×3 for the detail branch, dilated convolution r=2 for the context branch, and global average pooling for the semantic branch).

[0040] Figure 3 This demonstrates the complete processing pipeline from raw image input to detection output, encompassing four core stages: multi-scale feature extraction, spatial domain three-branch processing, frequency domain three-branch processing, and cross-domain fusion. The top light gray module performs feature transformation and convolution processing; the middle light blue module completes multi-scale feature extraction and frequency domain analysis; and the bottom light orange module performs multi-task classification and regression, outputting the final detection result.

[0041] The feature deconstruction process is as follows: the features are fed into three branches in the spatial domain to obtain an intermediate feature map, and then a branch-specific mask is obtained by a soft gating mechanism based on the Sigmoid function. Each element in the mask corresponds to the probability that the feature map at that location is a "robust discriminative signal". The higher the mask value, the more stable the detailed information (such as edges and contours) of the small target in that region, and the less affected it is by adversarial perturbations.

[0042] Dual-domain cross-domain interaction and fusion stage: corresponding to step S4 above. By presetting the frequency band radii r_low and r_high (set to 4 and 12 respectively for 80×80 feature maps), the spectrum is divided into high, medium and low frequency parts.

[0043] Fusion Logic: A cross-domain intersection fusion method is adopted. First, spatial domain detail features are used as the query vector, and frequency domain structural features are used as the key and value vectors, respectively, to calculate their correlation matrix. Subsequently, an adaptive weight α (automatically generated through learnable 1×1 convolution) is introduced to dynamically adjust the fusion ratio of spatial details and frequency structure. This method can effectively uncover weak features of small targets that are submerged under adversarial interference, ultimately generating a highly robust feature representation for small targets, which is then input into the detection head to output the results.

[0044] Example 2 Comparison and verification of the effects of different methods The models used in this experiment were all YOLOv8 base models, and two publicly available small object detection datasets were selected: the VisDrone2019 drone aerial photography dataset and the AI-TOD minimal object detection dataset. VisDrone2019 contains 6471 training images, with small objects accounting for over 60%; AI-TOD contains over 700,000 object instances, with an average object size of only [missing information]. Pixels are currently an extremely challenging evaluation dataset.

[0045] The comparative defense methods employ two mainstream target detection enhancement and defense techniques: FocalSpotter and UAV-DETR. FocalSpotter improves the detection performance of small targets through multi-scale feature focusing technology, but its defense effect is limited under strong adversarial attacks (such as PGD and FGSM). UAV-DETR uses frequency domain fusion technology to improve the detection accuracy of UAV targets, but lacks a targeted functional deentanglement mechanism when dealing with adversarial noise. In contrast, the RESO method of this invention, through refined recalibration in a dual-domain, three-branch manner, not only improves the APS index under clean samples but also exhibits a lower mAP decrease under various perturbation intensities. The comparison results are shown in Table 5.

[0046] Table 5 Performance comparison of different methods on the VisDrone 2019 drone aerial photography dataset , As shown in Table 5, on the FGSM adversarial sample with epsilon=0.01, the RESO method achieved the best performance on the VisDrone2019 dataset. On clean samples, the RESO method achieved an mAP of 33.4%, which is 4.9 percentage points higher than the original YOLOv8, and an APS of 25.5%. The mAP decreased from 9.9% to 8.2%, fully demonstrating the robustness of the RESO method under adversarial attacks.

[0047] Table 6 Performance comparison of different methods on the AI-TOD minimal object detection dataset , Table 6 shows the detection performance of different methods on the AI-TOD dataset for clean samples and FGSM adversarial samples (epsilon=0.01, denoted as Adv), covering three metrics: mAP, mAP50, and APS, as well as the decrease in mAP. YOLOv8 achieves a clean sample mAP of 32.6%, mAP50 of 53.8%, and APS of 21.5%, while under adversarial samples, the mAP is 22.1% and the APS is 13.2%, representing a 10.5% decrease in mAP. Introducing structures such as FPN, PANet, and BiFPN (see Table 6 for details) and the proposed method RESO (Ours), the clean sample mAP, mAP50, and APS are generally improved. For example, RESO achieves a clean mAP of 38.3%, mAP50 of 62.7%, and APS of 29.4%, while under adversarial samples, the advanced mAP is 30.1% and the APS is 21.3%, representing a -8.2% decrease in mAP. Overall, the improved methods (especially RESO) on the AI ​​TOD dataset not only improve the detection performance of clean samples but also enhance robustness against adversarial samples, significantly reducing the decrease in mAP.

[0048] Table 7 Explanation of different prior art Structure name Year of proposal Source paper Citation count public knowledge FPN 2017 Lin et al., CVPR 10000+ Publicly known infrastructure PANet 2018 Liu et al., CVPR 5000+ Public knowledge improves structure BiFPN 2020 Tan et al., CVPR 3000+ Publicly known efficient structure Among them, FPN (Feature Pyramid Network) Key innovations: Top-down path: High semantic feature upsampling and propagation; Lateral connectivity: fusing high semantic and high-resolution features; Multi-scale prediction: Detection is performed independently at different levels;

[0049] The improvements to YOLOv8 are significant because they enhance multi-scale feature representation capabilities and improve small target detection performance.

[0050] PANet (Path Aggregation Network) Core idea: Add a bottom-up path on top of FPN to shorten the information propagation distance of low-level features.

[0051] Key innovations: Two-way information flow: top-down + bottom-up; Adaptive feature pooling: aggregates features from all levels; Shorter information path: low-level features directly reach the top level; The improvement of YOLOv8 is significant because it enhances the propagation efficiency of small target features. YOLOv8 has a built-in bidirectional path structure similar to PANet.

[0052] BiFPN (Bi-directional Feature Pyramid Network) Core idea: Propose weighted feature fusion to achieve more efficient multi-scale feature fusion.

[0053] Key innovations: Bidirectional cross-scale connection: Remove single-input nodes and add extra edges; Weighted feature fusion: Different levels of features are assigned different weights; Fast normalization fusion: avoids expensive softmax calculations; The improvement of YOLOv8 is significant because it offers a more efficient feature fusion method than PANet, which can further enhance detection accuracy.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., 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 target detection adversarial defense method based on three-branch dual-domain refined processing, characterized in that, The method is as follows: A three-branch parallel topology of detail branch, context branch and semantic branch is constructed in the spatial domain. Multi-granularity features are extracted using different receptive fields, and non-robust noise is recalibrated at the pixel level through a soft gating mechanism based on Sigmoid. In the frequency domain, the spectrum is divided into three independent sub-bands—high frequency, mid frequency, and low frequency—based on the frequency radius, and differentiated enhancement is performed according to the frequency band distribution characteristics of small targets. Finally, a cross-domain intersection and fusion mechanism is used to achieve dynamic aggregation of spatial details and frequency structure through adaptive attention weights.

2. As described in claim 1, characterized in that, The method specifically includes the following steps: S1, Extract multi-scale feature maps from the input image, and select a specific feature layer containing information about small targets. Input the RESO module; S2, perform three-branch fine-grained processing of the spatial domain: construct detail branch, context branch and semantic branch; the detail branch uses small receptive field convolution to capture local texture, the context branch uses dilated convolution to capture the background around the target, and the semantic branch uses global average pooling to extract category priors; S3, Perform robust feature selection in the spatial domain: In the three branches of the spatial domain, convolution is used to evaluate the channel importance score, and the three branches of the spatial domain are generated by the Sigmoid activation function. Pixel-level calibration is performed on intermediate features to obtain spatially refined features. ; S4, Perform frequency domain three-branch refinement processing: Transform the features to the frequency domain using discrete Fourier transform, and refine the features according to the preset low-frequency radius. and high frequency radius The spectrum is divided into three independent sub-bands: low frequency, mid frequency, and high frequency. S5, Perform frequency domain differentiation enhancement: For different frequency band characteristics, a robust screening mechanism is used to extract stable components within each frequency band, which are then concatenated into refined frequency domain features after inverse transformation. ; S6, Perform bi-domain feature intersection fusion: Utilize a cross-domain attention mechanism to calculate the correlation matrix between spatial domain details and frequency domain structure. : ， In the formula, Represents a spatial domain query vector. This represents the transpose of the frequency domain key vector. Indicates the dimension of the key vector; S7, Perform adaptive feature integration: through learnable fusion weight factors A weighted integration of interactive features and original frequency domain features is achieved to generate the final refined features of small targets. : ， In the formula, Representing refined features from the frequency domain The extracted value vector is generated through a linear transformation; S8, Input the detection head, and it will output the category label and bounding box regression results of the small target.

3. The method according to claim 2, characterized in that, In step S2, the dilation rate of the detail branch convolution is set to 1, and the dilation rate of the context branch dilated convolution is set to 2.

4. The method according to claim 2, characterized in that, In step S3, the spatial domain three-branch mask We obtain the following formula: ， In the formula, These are the intermediate features of the corresponding branches after convolution transformations through different receptive fields. This is a soft-gated activation function.

5. The method according to claim 2, characterized in that, In step S4, the frequency band allocation strategy is aimed at Feature map of resolution, setting low-frequency radius High frequency radius .

6. The method according to claim 2, characterized in that, In step S4, the spectrum-independent subband division is performed using an indicator function. The implementation and division logic are as follows: ， The frequency radius in the spectrum is... and These are the preset low-frequency and high-frequency cutoff thresholds, respectively.

7. The method according to claim 2, characterized in that, In step S7, the adaptive feature integration includes the following steps: first, utilizing interaction features through Convolution and the Sigmoid function calculate the fusion weight factor along the channel dimension. : ， Then based on the fusion weight factor The final refined features are obtained by weighted summation of spatial interaction information and original frequency domain features. : ， In the formula, For Hadamard products.

Images