Micro-target detection method and device based on feature alignment and interaction, and server

Abstract

The application discloses a kind of based on feature alignment and interaction micro target detection method, device and server, belong to target detection technical field.It includes: image is input two layers W-LDS module and mamba perception module processing, obtains shallow layer physical fidelity feature;Shallow layer physical fidelity feature is input mamba perception module, obtains middle layer semantic enhancement feature;Middle layer semantic enhancement feature is input mamba perception module, obtains high layer semantic abstract feature;High layer semantic abstract feature is input dynamic interaction module and generates enhanced high layer feature;Feature is input EAF-Net, and shallow layer alignment fusion feature, middle layer alignment fusion feature and enhanced alignment fusion feature are generated;Shallow layer alignment fusion feature, middle layer alignment fusion feature and enhanced alignment fusion feature are input detection head, and micro target detection result is obtained.By lossless down-sampling and the double-flow collaborative mechanism of mamba perception, combined with feature interaction and alignment fusion technology, the accurate detection of micro target is realized.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method, apparatus and server for detecting tiny targets based on feature alignment and interaction. Background Technology

[0002] In the current fields of low-altitude economy and intelligent reconnaissance, target detection in UAV aerial images faces severe challenges. Unlike conventional natural scene images, targets in aerial remote sensing images, such as ground vehicles, pedestrians, or small facilities, generally exhibit the dual characteristics of extreme scale degradation and high-frequency coupling with background noise. They often occupy only a very small number of physical pixels in a wide field of view, lacking texture details and easily being submerged by complex ground backgrounds.

[0003] In the development of target detection technology, classic architectures represented by convolutional neural networks and emerging architectures represented by Transformers constitute the current technological mainstream. However, when these general-purpose models are directly transferred to the task of detecting small targets on UAVs, their inherent structural defects are gradually exposed.

[0004] Specifically, firstly, in the feature extraction stage, downsampling is commonly used, which easily filters out key high-frequency details of small targets, resulting in severe attenuation of target information in deep features. Secondly, in the multi-scale feature fusion stage, it mainly relies on simple interpolation and element-wise addition, lacking effective correction for spatial positional deviations between features of different scales, ultimately affecting the detection accuracy and positioning accuracy of small targets. Summary of the Invention

[0005] This invention provides a method, apparatus, and server for detecting small targets based on feature alignment and interaction, in order to solve the technical problems in the prior art where information attenuation is caused by the filtering out of high-frequency details in small targets and the impact of cross-scale spatial position deviation on the accuracy of small target detection.

[0006] In a first aspect, embodiments of the present invention provide a method for detecting small targets based on feature alignment and interaction, comprising:

[0007] The image is input into a two-layer W-LDS module and a Mamba sensing module for processing, which is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, thus obtaining shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling.

[0008] The shallow physical fidelity features are processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features.

[0009] The mid-level semantic enhancement features are processed by the Mamba perception module to further expand the receptive field, suppress background noise, and extract deep semantic context to obtain high-level semantic abstract features.

[0010] The Mamba sensing module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain through a fast Fourier transform and a frequency domain state space model module. The spatial flow branch refines the extraction of local texture and geometric features of small targets through a hybrid extended selection kernel unit.

[0011] The high-level semantic abstract features are processed by a dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit.

[0012] The shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features are respectively input into EAF-Net for multi-scale feature alignment and fusion, generating shallow aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features.

[0013] The shallow alignment fusion feature, the middle alignment fusion feature, and the enhanced alignment fusion feature are input into the decoupled convolutional detection head to obtain the small target detection result.

[0014] Secondly, embodiments of the present invention also provide a small target detection device based on feature alignment and interaction, comprising:

[0015] The shallow physical fidelity feature acquisition module is used to input the image into the two-layer W-LDS module and the Mamba sensing module for processing. It is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, so as to obtain shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling.

[0016] The mid-level semantic enhancement feature acquisition module is used to process the shallow physical fidelity features through the Mamba perception module, and to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism to obtain mid-level semantic enhancement features.

[0017] The high-level semantic abstraction feature acquisition module is used to process the mid-level semantic enhancement features through the Mamba perception module to further expand the receptive field, suppress background noise, and extract deep semantic context to obtain high-level semantic abstraction features.

[0018] The enhanced high-level feature acquisition module is used to process the high-level semantic abstract features through the dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit.

[0019] The multi-scale alignment and fusion module is used to input the shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features into EAF-Net for multi-scale feature alignment and fusion, generating shallow aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. The EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features.

[0020] The target detection module is used to input the shallow alignment fusion feature, the middle alignment fusion feature and the enhanced alignment fusion feature into the decoupled convolutional detection head to obtain the small target detection result.

[0021] Thirdly, embodiments of the present invention also provide a server, comprising:

[0022] One or more processors;

[0023] Storage device for storing one or more programs;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors implement the small target detection method based on feature alignment and interaction as provided in the above embodiments.

[0025] The present invention provides a method, apparatus, and server for detecting small targets based on feature alignment and interaction. The image is input into a two-layer W-LDS module and a Mamba perception module for processing. This process preserves high-frequency details and physical pixel information of small targets while reducing resolution, resulting in shallow physical fidelity features. The W-LDS module utilizes discrete wavelet transform decomposition to achieve lossless downsampling. The shallow physical fidelity features are then processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features. These mid-level semantic enhancement features are further processed by the Mamba perception module to further expand the receptive field, suppress background noise, and refine deep semantic context, resulting in high-level semantic abstraction features. The Mamba perception module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain through fast Fourier transform and a frequency domain state-space model module. The spatial flow branch uses hybrid extension selection... The kernel unit refines the local texture and geometric features of the micro-target. The high-level semantic abstract features are then processed by a dynamic interaction module to generate enhanced high-level features. This dynamic interaction module includes a spectrum-aware branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. The shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features are input into EAF-Net for multi-scale feature alignment and fusion, generating shallow-aligned fusion features, mid-level aligned fusion features, and enhanced aligned fusion features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module performs secondary semantic enhancement on the features in the frequency domain. The feature alignment module guides the features for spatial position correction. The context fusion module achieves progressive complementary fusion of shallow and deep features. The shallow-aligned fusion features, mid-level aligned fusion features, and enhanced aligned fusion features are input into a decoupled convolutional detection head to obtain the micro-target detection result. By employing lossless downsampling of the W-LDS module, dual-stream collaboration of the Mamba sensing module, holographic frequency-space interaction of the dynamic interaction module, and explicit alignment and fusion of EAF-Net, the high-frequency details of small targets are preserved, global context information is perceived, and cross-scale features are spatially aligned, thereby improving the detection accuracy and precision of small targets. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart of the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0028] Figure 2This is a schematic diagram of the W-LDS module in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0029] Figure 3 This is a schematic diagram of the structure of the Mamba sensing module in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0030] Figure 4 This is a schematic diagram of the feature expansion branch in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0031] Figure 5 This is a schematic diagram of the channel fusion unit in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0032] Figure 6 This is a schematic diagram of the dynamic interaction module in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0033] Figure 7 This is a schematic diagram of the spectrum sensing branch in the micro-target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0034] Figure 8 This is a schematic diagram of the module length-aware routing operator in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0035] Figure 9 This is a schematic diagram of the structure of the local aggregation branch in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0036] Figure 10 This is a schematic diagram of the global frequency domain modulation module in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0037] Figure 11 This is a schematic diagram of the feature alignment module in the small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention;

[0038] Figure 12 This is a schematic diagram of the micro-target detection device based on feature alignment and interaction provided in Embodiment 2 of the present invention;

[0039] Figure 13 This is a structural diagram of the server provided in Embodiment 3 of the present invention. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0041] Example 1

[0042] Figure 1 This is a flowchart of a small target detection method based on feature alignment and interaction provided in Embodiment 1 of the present invention. This embodiment is applicable to small target detection scenarios in UAV aerial images or remote sensing images, and specifically includes the following steps:

[0043] Step 110: The image is input into a two-layer W-LDS module and a Mamba sensing module for processing, which is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, and obtain shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling.

[0044] An image is a digital image composed of an array of pixels, where each pixel records the light intensity or color information of a corresponding location in a scene. Images can be grayscale or color, and their spatial dimensions and color dimensions are typically represented by width, height, and the number of channels. In computer vision tasks, images serve as raw input data, carrying visual information about the target to be detected. For example, an image could be an aerial photograph captured by a camera mounted on a drone, or a remote sensing image obtained by a satellite.

[0045] For example, such as Figure 2As shown, the W-LDS module utilizes discrete wavelet transform decomposition to achieve lossless downsampling. Optionally, the module can first use a Haar wavelet basis as the transform kernel to perform a two-dimensional discrete wavelet transform on the input features, orthogonally decomposing them into four non-overlapping sub-bands containing information in different frequency bands: a low-frequency sub-band to preserve the basic illumination and contour information of the image; a horizontal sub-band to preserve the edge gradient in the horizontal direction; a vertical sub-band to preserve the edge gradient in the vertical direction; and a diagonal sub-band to preserve the texture features in the diagonal direction. After this transformation, the spatial resolution of each sub-band is reduced to half of the original input, while the number of channels remains unchanged, thus completely and losslessly distributing the image information across the four frequency components. Next, the W-LDS module performs a channel recombination operation, splicing the above four sub-bands along the channel dimension to form a stacked feature tensor with the number of channels expanded to four times that of the original input. This stacked feature tensor simultaneously contains low-frequency approximation information and high-frequency detail information in three directions, achieving explicit preservation of key high-frequency features such as small target edges and corners. Finally, in order to control the number of channels and fuse features from different frequency bands, the module uses a 1×1 convolutional layer to perform linear feature fusion and channel dimensionality reduction on the stacked features, adjusting the number of channels to the preset target dimension.

[0046] For example, such as Figure 3 As shown, the Mamba Sensing module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain through Fast Fourier Transform (FFT) and a frequency domain state-space model module. The spatial flow branch refines the extraction of local texture and geometric features of small targets through a hybrid extended selection kernel unit. Specifically, the frequency flow branch includes a FFT unit, a frequency domain state-space model module, and an inverse FFT unit. The frequency domain state-space model module includes a two-dimensional FFT unit, a two-dimensional selective scanning mechanism unit, and a two-dimensional inverse FFT. The two-dimensional FFT unit transforms the input features from the spatial domain to the complex frequency domain. The two-dimensional selective scanning mechanism unit performs state-space sequence modeling in the frequency domain to capture long-distance dependencies, thereby obtaining global context information. The two-dimensional inverse FFT restores the processed frequency domain features to the spatial domain. Optionally, the processing flow of the frequency flow branch is as follows: first, the input features are transformed to the frequency domain by a Fast Fourier Transform (FFT) unit, then fed into the frequency domain state space model module, and finally restored to the spatial domain by an Inverse Fast Fourier Transform (IFFT) unit. The frequency domain state space model module takes a complex spectral feature map as input and expands it into a one-dimensional sequence along multiple directions (such as from top left to bottom right, bottom right to top left, etc.) using a two-dimensional selective scanning mechanism. It then uses the state space model to perform long-distance dependency modeling on the sequence, and subsequently reassembles the processed sequence into a two-dimensional spectral feature map, thereby achieving low-cost capture of the global context in the frequency domain.

[0047] For example, such as Figure 3 As shown, the spatial flow branch includes a hybrid extended selection kernel unit, which comprises a four-branch parallel unit and an SSG mechanism unit. The four-branch parallel unit includes a 3×3 convolution, an 11×1 convolution, a 1×11 convolution, and a feature expansion branch. The feature expansion branch includes a parallel dilated convolution branch, a 3×3 convolution branch, and an identity mapping branch, used to extract rich gradient flows through multi-scale features to expand the receptive field of small targets. The SSG mechanism unit is used to: concatenate the output features of the four-branch parallel unit to obtain a first concatenated feature; perform average pooling and max pooling operations on the first concatenated feature to obtain average pooling features and max pooling features; concatenate the average pooling features and max pooling features along the channel dimension to aggregate spatial information to obtain concatenated spatial features; pass the concatenated spatial features through a 7×7 convolutional layer and an activation function to obtain a spatial attention map; and multiply the spatial attention map element-wise with the first concatenated feature to obtain the output features of the spatial flow branch. Optionally, the spatial flow branch processing flow involves the input features being processed in parallel through 3×3 convolutions, 11×1 convolutions, 1×11 convolutions, and a feature expansion branch. The output features of each branch are then input into the SSG mechanism unit. First, the output features of each branch are concatenated. Then, parallel operations of average pooling and max pooling, channel-dimension concatenation, a 7×7 convolutional layer, and a sigmoid activation function are performed to generate a spatial attention map. Finally, the spatial attention map is element-wise multiplied with the concatenated output features of each branch to obtain the output features of the spatial flow branch. For example, ... Figure 4 As shown, the feature expansion branch includes three parallel sub-branches during the training phase: a 3×3 dilated convolution with an expansion rate of 2, a standard 3×3 convolution, and an identity mapping branch. Each branch is followed by batch normalization, enriching the gradient flow through multi-scale feature extraction. However, during the inference phase, the three branches are fused into a single 3×3 dilated convolution using a structure reparameterization technique, expanding the receptive field without increasing inference latency. For example, as... Figure 5As shown, the channel fusion unit is used to add the output features of the frequency flow branch and the spatial flow branch element-wise to obtain the added features. The added features are then processed sequentially through global average pooling, a multilayer perceptron, and an activation function to obtain the channel weight vector. The channel weight vector is then multiplied channel-wise with the added features to obtain the channel fusion features. In summary, the processing flow of the Mamba perception module is as follows: the input features are fed into the frequency flow branch and the spatial flow branch respectively. The frequency flow branch is used to capture global context information to obtain the output features of the frequency flow branch. The spatial flow branch is used to extract the local texture and geometric features of small targets to obtain the output features of the spatial flow branch. The output features of the frequency flow branch and the output features of the spatial flow branch are fed into the channel fusion unit to obtain the final output features of the Mamba perception module, i.e., the channel fusion features. Step 110 processes the input image through a cascaded two-layer W-LDS module and the Mamba perception module. While gradually reducing the spatial resolution of the feature map, the lossless decomposition characteristics of wavelet transform are used to completely preserve the high-frequency edges and physical pixel information of small targets, thereby generating shallow physical fidelity features rich in detail.

[0048] Step 120: The shallow physical fidelity features are processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features.

[0049] Shallow physical fidelity features refer to features extracted from the shallow layers of a neural network. They are characterized by high resolution and a small number of channels, preserving rich spatial details and original pixel information in the input image, such as high-frequency components like edges, textures, and corners. These features maintain a high degree of physical consistency with the original image, realistically reflecting the geometric structure and local morphology of the target. For example, in step 120, the shallow physical fidelity features are input into the Mamba perception module. Leveraging its dual-stream collaborative mechanism of frequency and spatial streams, it effectively captures global long-distance dependencies while preserving high-frequency details, simultaneously enhancing the semantic representation ability of small targets. Ultimately, this generates mid-level semantically enhanced features, achieving the transition from detail fidelity to semantic perception.

[0050] Step 130: The mid-level semantic enhancement features are processed by the Mamba perception module to further expand the receptive field, suppress background noise, and refine the deep semantic context to obtain high-level semantic abstract features.

[0051] Mid-level semantic enhancement features refer to features extracted from the middle layers of a neural network. Their resolution falls between that of shallow and deep layers, and the number of channels is increased compared to shallow layers. While preserving some spatial details, these features begin to aggregate local semantic information, reflecting the object's contours, component structures, and medium-scale contextual relationships. Compared to shallow physical fidelity features, mid-level semantic enhancement features possess stronger semantic discriminative capabilities. For example, in step 130, the mid-level semantic enhancement features are input into the Mamba perception module. By deepening the network layers, the receptive field is further expanded, a dual-stream collaborative mechanism effectively suppresses background noise, and higher-level semantic contextual information is extracted, ultimately generating high-level semantic abstract features.

[0052] Step 140: The high-level semantic abstract features are processed by the dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit.

[0053] High-level semantic abstract features refer to features extracted from deep layers of neural networks. They have lower spatial resolution, more channels, and each feature point corresponds to a larger receptive field. These features have strong semantic discriminative ability, but retain less of the precise location and edge details of the target.

[0054] For example, such as Figure 6 As shown, the dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. Specifically, as... Figure 7As shown, the spectrum-aware branch processes the high-level semantic abstract features sequentially through a two-dimensional fast Fourier transform unit, a global spectrum self-attention unit, and a two-dimensional inverse fast Fourier transform unit to obtain spectral features. The global spectrum self-attention unit performs frequency domain transformation and self-attention calculation on the output features of the two-dimensional fast Fourier transform unit to achieve efficient capture of global context information and enhancement of frequency domain features. Optionally, the processing flow of the spectrum-aware branch is to orthogonally map the input high-level semantic abstract features to the complex frequency domain through a two-dimensional fast Fourier transform. This transformation decomposes the pixel information in the spatial domain into combinations of different frequency components, generating a spectral feature map containing amplitude and phase spectra. Since each point in the frequency domain corresponds to the superposition of all pixels in the spatial domain, this spectral feature map explicitly encodes the global structural information of the entire image, rather than being limited to the local neighborhood. Secondly, in the frequency domain space, the spectrum-aware branch deploys a global spectrum self-attention unit. The global spectrum self-attention unit first projects the spectral feature map through three independent linear layers into a spectral query vector, a spectral key vector, and a spectral value vector, respectively. Next, spectral self-attention computation is performed, calculating the dot product between the spectral query vector and the spectral key vector to obtain the correlation matrix between the frequency components. This correlation matrix is ​​then used as weights to perform weighted aggregation of the spectral value vectors. Finally, the complex spectral features after self-attention interaction are restored back to the real space domain through a two-dimensional inverse fast Fourier transform, outputting the spectral features.

[0055] For example, amplitude-guided sparse branching is used to divide the high-level semantic abstract features into S×S local grids. An importance map reflecting the importance of each local grid is generated using a modulus-aware routing operator. High-response regions are selected based on the importance map and clustered to obtain clustered region features. The clustered region features are then refined using a sliding window hole attention operator to obtain enhanced region features. Finally, the enhanced region features are filled back into the original positions of the high-response regions in the high-level semantic abstract features using an inverse mapping feature restoration operation, resulting in sparse spatial features. Figure 8 As shown, the modulus-aware routing operator is used to linearly map the features within each local grid after partitioning, generating query vectors and key vectors; the L2 norm of the query vector is calculated by the modulus calculation module, and the L2 norm is input into the dynamic factor generator to generate a dynamic scaling factor. and dynamic offset factor The query vector and the key vector are projected into a high-dimensional space through kernel function mapping, and linear attention is calculated to obtain the linear attention calculation result; the linear attention calculation result is then compared with the dynamic scaling factor. Perform multiplicative modulation to generate a multiplicative modulation result; then combine the multiplicative modulation result with the dynamic offset factor. Additive modulation is performed to generate an importance map reflecting each local grid. Optionally, the amplitude-guided sparse branching process can proceed as follows: First, the input high-level semantic abstract features are spatially divided into S×S local grids. For each grid, a modulus-aware routing operator is used to linearly map the features within that grid to generate a query vector and a key vector. Using the L2 norm of the query vector, a dynamic factor generator performs nonlinear mapping based on feature intensity to generate a dynamic scaling factor. and dynamic offset factor Simultaneously, the query and key are mapped using a kernel function, and then linear attention is calculated. The result is then compared with a dynamic scaling factor. Multiplication, and dynamic offset factor The features of each grid cell are summed to generate an importance map reflecting the importance of each grid cell. Then, based on the importance maps of all grid cells, high-response regions with the highest response values ​​are selected, and the features of these regions are clustered. Next, the clustered region features are fed into a sliding window hole attention operator for refined feature extraction, resulting in enhanced region features. Finally, through an inverse mapping feature restoration operation, the enhanced region features are filled back into the feature map according to their original grid positions. Unselected regions retain their original features, thus outputting sparse spatial features.

[0056] For example, such as Figure 9As shown, the local aggregation branch is used to decouple the high-level semantic abstract features into a first feature subgroup and a second feature subgroup through channel grouping. The first feature subgroup is then subjected to a 3×3 convolution to obtain multi-scale spatial features. The second feature subgroup is then subjected to a 1×1 convolution to obtain point-by-point channel features. Through residual connections, the spatial weight map generated from the multi-scale spatial features is injected into the point-by-point channel features to obtain a first interactive enhancement feature. The channel weights extracted from the point-by-point channel features are then injected back into the multi-scale spatial features to obtain a second interactive enhancement feature. The first and second interactive enhancement features are then concatenated and fused to obtain the local aggregation feature. Optionally, the processing flow of the local aggregation branch is as follows: First, the input high-level semantic abstract features are grouped along the channel dimension and decoupled into a first feature subgroup and a second feature subgroup to achieve information complementarity after feature decoupling. The first feature subgroup extracts multi-scale spatial features through a 3×3 convolution, focusing on the geometric topology of the target and generating a spatial weight map. The second feature subgroup is processed by 1×1 convolution to extract point-by-point channel features, maintaining semantic purity in the channel dimension, identifying target categories and background noise, and generating channel weights. Then, two asymmetric interactive fusion processes are performed. First, the spatial weight map generated from the multi-scale spatial features is injected into the point-by-point channel features via residual connections in an element-wise modulation manner, enabling the channel features, which originally lacked spatial awareness, to acquire localization information, resulting in the first interactive enhancement feature. Second, the channel weights extracted from the point-by-point channel features are injected back into the multi-scale spatial features, enabling the spatial filter to distinguish between the true target signal and clutter interference during large receptive field scanning, correcting potential spatial position drift in the deep network, resulting in the second interactive enhancement feature. Finally, the two interactively enhanced feature subgroups are concatenated along the channel dimension to obtain the local aggregated feature.

[0057] For example, the gated fusion unit is used to process the spectral features through an activation function to generate gate weights; perform element-wise multiplication of the gate weights with the sparse spatial features to obtain a first intermediate feature; perform element-wise multiplication of the gate weights with the local aggregated features to obtain a second intermediate feature; and perform addition, fusion, and normalization processing on the spectral features, the first intermediate feature, and the second intermediate feature to obtain enhanced high-level features. Optionally, the activation function is the Sigmoid function, so that the generated gate weights have values ​​between 0 and 1, and the normalization operation is LayerNorm. In summary, optionally, the processing flow of the dynamic interaction module is as follows: the high-level semantic abstract features obtained in step 130 can be input into the spectrum sensing branch, the amplitude-guided sparse branch, and the local aggregation branch respectively to obtain spectrum features, sparse spatial features, and local aggregation features. The spectrum sensing branch is used to capture global context information and perform frequency domain feature enhancement, the amplitude-guided sparse branch is used to focus on salient regions and generate sparse spatial features, and the local aggregation branch is used to extract local geometric details and generate local aggregation features. Then, the spectrum features, sparse spatial features, and local aggregation features are sent together to the gate control fusion unit. After gate weight modulation and feature fusion, enhanced high-level features are obtained.

[0058] Step 150: Input the shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features into EAF-Net for multi-scale feature alignment and fusion to generate shallow aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features.

[0059] For example, such as Figure 10As shown, the global frequency domain modulation module processes the enhanced high-level features through a lightweight MLP branch and a learnable weight branch, then fuses the output features of the lightweight MLP branch and the learnable weight branch to obtain dual-branch output features. These dual-branch output features are then sequentially processed through a fast Fourier transform, a frequency-selective scanning mechanism, and an inverse fast Fourier transform to achieve secondary global semantic enhancement, resulting in enhanced global features. The lightweight MLP branch includes global average pooling, a fully connected layer, and an activation function to extract global semantic information from the enhanced high-level features and capture global dependency information of the image. The learnable weight branch includes features aggregation for specific regions and learnable center weight encoding operations to aggregate local corner information of the enhanced high-level features and capture salient topological structures in the image. The frequency-selective scanning mechanism flattens the two-dimensional spectrum into a one-dimensional sequence, establishes long-range dependencies between frequency components using the Mamba operator, and selectively enhances target periodic frequency components while suppressing noise components, achieving secondary refinement of the overall image semantics. In summary, the processing flow of the global frequency domain modulation module is as follows: First, the input enhanced high-level features are fed into a lightweight MLP branch and a learnable weight branch for parallel processing. The lightweight MLP branch extracts global semantic vectors through global average pooling and a two-layer fully connected network, capturing global dependency information of the image. The learnable weight branch uses specific region feature aggregation and learnable center weight encoding to divide the feature map into local regions. Within each region, it extracts salient responses through pooling or statistical aggregation. Simultaneously, it uses the learnable center vector to perform similarity matching with the input features, encoding the input signal into a weight distribution corresponding to the visual template, thereby aggregating local corner information and capturing salient topological structures in the image. The output features of the two branches are fused to form a dual-branch output feature. Subsequently, the dual-branch output features are sequentially processed through Fast Fourier Transform (FFT), frequency-domain selective scanning mechanism, and inverse FFT: First, the features are transformed to the frequency domain using FFT, then the two-dimensional spectrum is flattened into a one-dimensional sequence using the frequency-domain selective scanning mechanism, and the long-range dependencies between frequency components are established using the linear complexity characteristics of the Mamba operator. The frequency components representing the target period are selectively enhanced while noise components are suppressed, achieving secondary refinement of the semantics of the entire graph. Finally, the inverse FFT is used to restore the features to the spatial domain, outputting enhanced global features.

[0060] For example, such as Figure 11As shown, the feature alignment module includes a wavelet-guided downsampling unit, which decomposes the first input feature into low-frequency subbands and high-frequency subbands in the horizontal, vertical, and diagonal directions using discrete wavelet transform, and concatenates the four subbands in the channel dimension to generate a downsampled guided feature with resolution matching the deep feature; a spatial offset prediction unit, which concatenates the downsampled guided feature with the second input feature and predicts the spatial offset field and modulation mask through a convolutional layer; and a deformable convolutional unit, which performs pixel-level adaptive resampling of the second input feature according to the spatial offset field and modulates the feature using the modulation mask to output aligned features. Optionally, the processing flow of the feature alignment module is as follows: first, the first input feature is used as a guide map input to the wavelet-guided downsampling unit, which decomposes it into low-frequency subbands and high-frequency subbands in the horizontal, vertical, and diagonal directions using discrete wavelet transform, and concatenates the four subbands in the channel dimension. This achieves physical resolution reduction to match the size of the deep feature while completely preserving the edges and texture gradients in the guide map, generating the downsampled guided feature. Then, the downsampling guiding feature is concatenated with the second input feature and fed into the spatial offset prediction unit. A set of spatial offset fields and modulation masks are predicted through convolutional layers, where the offset field represents the geometric deformation of the sampling points, and the modulation mask represents the importance weighting of the sampling points. Finally, the predicted spatial offset field and modulation mask are input together with the second input feature into a deformable convolutional unit. Based on the offset, the sampling point positions of the second input feature are adaptively adjusted and resampled at the pixel level, and the modulation mask is used for feature modulation to output aligned features. This achieves precise pixel alignment of deep semantic features to shallow geometric features at the physical level.

[0061] For example, the context fusion module includes a first-level fusion unit and a second-level fusion unit. The first-level fusion unit performs preliminary feature fusion and background noise filtering through concatenation, convolutional layers, and activation function operations. The second-level fusion unit performs secondary purification and spatial location calibration of the primary fused features obtained from the first-level fusion unit through concatenation, two convolutional layers, activation functions, and weighted fusion operations. Optionally, the processing flow of the context fusion module is as follows: First, shallow input features and upsampled deep input features are input into the first-level fusion unit. The two features are concatenated, and a first-level spatial gating weight map is generated through convolutional layers and a sigmoid activation function. This weight map performs weighted fusion on the two features in a complementary manner to achieve preliminary feature fusion and background noise filtering, resulting in primary fused features. Subsequently, the primary fused features and shallow input features are fed into the second-level fusion unit. After concatenation and convolutional layers, a fused feature map is obtained. The fused feature map is then processed through convolutional layers and activation functions to obtain second-level spatial detail weights and second-level semantic context weights. Finally, the secondary spatial detail weights are multiplied element-wise with the shallow input features, and then added to the secondary semantic context weights to obtain the output features of the context fusion module.

[0062] For example, the EAF-Net processing flow in step 150 is as follows: the enhanced high-level features obtained in step 140 are input into the global frequency domain modulation module for processing to obtain enhanced global features. These enhanced global features serve as the first output of EAF-Net, namely, the enhanced alignment fusion feature. Simultaneously, these enhanced global features and the mid-level semantic enhancement features obtained in step 120 are input into the feature alignment module to obtain the first alignment feature. The first alignment feature and the mid-level semantic enhancement features are then fed into the context fusion module to obtain the first fused feature. This first fused feature serves as the second output of EAF-Net, namely, the mid-level alignment fusion feature. Next, the first fused feature and the shallow physical fidelity features obtained in step 110 are input into the feature alignment module to obtain the second alignment feature. The second alignment feature and the shallow physical fidelity features are then fed into the context fusion module to obtain the second fused feature. This second fused feature serves as the third output of EAF-Net, namely, the shallow alignment fusion feature. This completes the EAF-Net processing flow, achieving the alignment and fusion of multi-scale features.

[0063] Step 160: Input the shallow alignment fusion feature, the middle alignment fusion feature and the enhanced alignment fusion feature into the decoupled convolutional detection head to obtain the small target detection result.

[0064] The detection head is the module in the object detection model responsible for outputting the final detection result. Located at the end of the model, it receives the processed feature map and, through a series of calculations, transforms the semantic information contained in the feature map into specific detection boxes and class labels. Typically, the detection head contains two parallel branches: one for predicting the class confidence of the target, and the other for regressing the position coordinates of the target bounding box. For example, the detection head in this embodiment adopts a decoupled convolutional architecture, containing three parallel decoupled convolutional branches, corresponding to the shallow alignment fusion features, mid-level alignment fusion features, and enhanced alignment fusion features output by EAF-Net, respectively. Each branch independently performs dense prediction, utilizing low-level high-resolution features to enhance the localization accuracy of small targets and utilizing high-level strong semantic features to improve the discriminative ability of target categories, generating candidate detection results at the corresponding scale. Subsequently, the candidate detection results generated by the three scale branches are merged, and redundant boxes are eliminated and optimally aggregated using a non-maximum suppression algorithm, finally outputting the small target detection result.

[0065] To verify the effectiveness of this embodiment, experiments were conducted on a SIMD dataset containing 5000 high-resolution satellite aerial images covering 15 subcategories of targets. Data preprocessing strategies such as discrete rotation enhancement and sliding window slicing were employed. The total loss function was a weighted sum of classification loss, regression loss, and L1 loss, and the dataset was trained for 150 epochs. Results show that the proposed method achieves an average precision of 76.4% on the SIMD test set, a 2.8% improvement over the baseline model. Recall rates for small cars and long-haul trailers are improved by 5.1% and 4.5%, respectively. Furthermore, on the NVIDIA RTX 4090 platform, the full-image inference latency for 1024×768 resolution input is controlled within 35ms, validating the effectiveness of this embodiment in detecting small targets and demonstrating good real-time performance.

[0066] This embodiment processes images by inputting them into a two-layer W-LDS module and a Mamba perception module. This process preserves high-frequency details and physical pixel information of small targets while reducing resolution, resulting in shallow physical fidelity features. The W-LDS module utilizes discrete wavelet transform decomposition to achieve lossless downsampling. These shallow physical fidelity features are then processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features. These mid-level semantic enhancement features are further processed by the Mamba perception module to expand the receptive field, suppress background noise, and refine deep semantic context, resulting in high-level semantic abstraction features. The Mamba perception module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain using fast Fourier transform and a frequency-domain state-space model module. The spatial flow branch refines the extraction of local context information of small targets through a hybrid extended selection kernel unit. The textural and geometric features are analyzed. The high-level semantic abstract features are processed by a dynamic interaction module to generate enhanced high-level features. This dynamic interaction module includes a spectrum-aware branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. The shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features are input into EAF-Net for multi-scale feature alignment and fusion, generating shallow-aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module performs secondary semantic enhancement on the features in the frequency domain. The feature alignment module guides the features to perform spatial position correction. The context fusion module achieves progressive complementary fusion of shallow and deep features. The shallow-aligned fused features, mid-level aligned fused features, and enhanced aligned fused features are input into a decoupled convolutional detection head to obtain small target detection results. By employing lossless downsampling of the W-LDS module, dual-stream collaboration of the Mamba sensing module, holographic frequency-space interaction of the dynamic interaction module, and explicit alignment and fusion of EAF-Net, the high-frequency details of small targets are preserved, global context information is perceived, and cross-scale features are spatially aligned, thereby improving the detection accuracy and precision of small targets.

[0067] Example 2

[0068] Figure 12 This is a schematic diagram of the micro-target detection device based on feature alignment and interaction provided in Embodiment 2 of the present invention, as shown below. Figure 12 As shown, the device includes:

[0069] The shallow physical fidelity feature acquisition module 210 is used to input the image into the two-layer W-LDS module and the Mamba sensing module for processing. It is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, so as to obtain shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling.

[0070] The mid-level semantic enhancement feature acquisition module 220 is used to process the shallow physical fidelity features through the Mamba perception module, and to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism to obtain mid-level semantic enhancement features.

[0071] The high-level semantic abstraction feature acquisition module 230 is used to process the mid-level semantic enhancement features through the Mamba perception module to further expand the receptive field, suppress background noise, and refine the deep semantic context to obtain high-level semantic abstraction features.

[0072] The enhanced high-level feature acquisition module 240 is used to process the high-level semantic abstract features through the dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit.

[0073] The multi-scale alignment and fusion module 250 is used to input the shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features into EAF-Net for multi-scale feature alignment and fusion, generating shallow alignment fusion features, mid-level alignment fusion features, and enhanced alignment fusion features. The EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features.

[0074] The target detection module 260 is used to input the shallow alignment fusion feature, the middle alignment fusion feature and the enhanced alignment fusion feature into the decoupled convolutional detection head to obtain the small target detection result.

[0075] The micro-target detection device based on feature alignment and interaction provided in this embodiment processes images into a two-layer W-LDS module and a Mamba perception module. This process preserves high-frequency details and physical pixel information of micro-targets while reducing resolution, resulting in shallow physical fidelity features. The W-LDS module utilizes discrete wavelet transform decomposition to achieve lossless downsampling. The shallow physical fidelity features are then processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of micro-targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features. These mid-level semantic enhancement features are further processed by the Mamba perception module to expand the receptive field, suppress background noise, and refine deep semantic context, resulting in high-level semantic abstraction features. The Mamba perception module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain through fast Fourier transform and a frequency domain state-space model module. The spatial flow branch uses a hybrid extended selection kernel unit. The local texture and geometric features of the small target are extracted with fine detail. The high-level semantic abstract features are then processed by a dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum-aware branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. The shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features are then input into EAF-Net for multi-scale feature alignment and fusion, generating shallow-aligned fusion features, mid-level aligned fusion features, and enhanced aligned fusion features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement of the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to achieve progressive complementary fusion of deep and shallow features. The shallow-aligned fusion features, mid-level aligned fusion features, and enhanced aligned fusion features are then input into a decoupled convolutional detection head to obtain the small target detection results. By employing lossless downsampling of the W-LDS module, dual-stream collaboration of the Mamba sensing module, holographic frequency-space interaction of the dynamic interaction module, and explicit alignment and fusion of EAF-Net, the high-frequency details of small targets are preserved, global context information is perceived, and cross-scale features are spatially aligned, thereby improving the detection accuracy and precision of small targets.

[0076] Example 3

[0077] Figure 13 This is a schematic diagram of the structure of a server provided in Embodiment 3 of the present invention. Figure 13 A block diagram is shown of an exemplary server 12 suitable for implementing embodiments of the present invention. Figure 13 The server 12 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0078] like Figure 13 As shown, server 12 is presented in the form of a general-purpose computing server. The components of server 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0079] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0080] Server 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by server 12, including volatile and non-volatile media, removable and non-removable media.

[0081] System memory 28 may include computer system readable media in the form of volatile memory, such as RAM 30 and / or cache 32. Server 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media ( Figure 13 Not shown; usually referred to as a "hard drive"). Although Figure 13 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0082] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0083] Server 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing server, display 24, etc.), and with one or more servers that enable users to interact with server 12, and / or with any server (e.g., network card, modem, etc.) that enables server 12 to communicate with one or more other computing servers. This communication can be performed via I / O interface 22. Furthermore, server 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of server 12 via bus 18. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with server 12, including but not limited to: microcode, server drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0084] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the small target detection method based on feature alignment and interaction provided in the embodiments of the present invention.

[0085] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for detecting small targets based on feature alignment and interaction, characterized in that, include: The image is input into a two-layer W-LDS module and a Mamba sensing module for processing, which is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, thus obtaining shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling. The shallow physical fidelity features are processed by the Mamba perception module to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism, resulting in mid-level semantic enhancement features. The mid-level semantic enhancement features are processed by the Mamba perception module to further expand the receptive field, suppress background noise, and extract deep semantic context to obtain high-level semantic abstract features. The Mamba sensing module includes a frequency flow branch, a spatial flow branch, and a channel fusion unit. The frequency flow branch captures global context information in the frequency domain through a fast Fourier transform and a frequency domain state space model module. The spatial flow branch refines the extraction of local texture and geometric features of small targets through a hybrid extended selection kernel unit. The high-level semantic abstract features are processed by a dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. The shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features are respectively input into EAF-Net for multi-scale feature alignment and fusion, generating shallow aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features. The shallow alignment fusion feature, the middle alignment fusion feature, and the enhanced alignment fusion feature are input into the decoupled convolutional detection head to obtain the small target detection result.

2. The method according to claim 1, characterized in that, The frequency flow branch includes: The system comprises a Fast Fourier Transform (FFT) unit, a frequency domain state-space model module, and an Inverse Fast Fourier Transform (IFFT) unit. The frequency domain state-space model module includes: The system comprises a two-dimensional fast Fourier transform unit, a two-dimensional selective scanning mechanism unit, and a two-dimensional fast Fourier inverse transform. The two-dimensional fast Fourier transform unit is used to transform the input features from the spatial domain to the complex frequency domain. The two-dimensional selective scanning mechanism unit is used to perform state-space sequence modeling in the frequency domain to capture long-distance dependencies, thereby obtaining global context information. The two-dimensional fast Fourier inverse transform is used to restore the processed frequency domain features to the spatial domain.

3. The method according to claim 2, characterized in that, The spatial flow branch includes: A hybrid extended selection core unit, comprising a four-branch parallel unit and an SSG mechanism unit; The four-branch parallel unit includes: 3×3 convolution, 11×1 convolution, 1×11 convolution and feature expansion branch, the feature expansion branch including parallel dilated convolution branch, 3×3 convolution branch and identity mapping branch, are used to enrich gradient flow through multi-scale feature extraction to expand the receptive field of small targets; The SSG mechanism unit is used for: The output features of the four-branch parallel units are concatenated to obtain the first concatenated feature; The first concatenated feature is subjected to average pooling and max pooling operations respectively to obtain average pooling feature and max pooling feature; The average pooling feature and the max pooling feature are concatenated and aggregated along the channel dimension to obtain the concatenated spatial feature. The spliced ​​spatial features are passed through a 7×7 convolutional layer and an activation function to obtain a spatial attention map; The spatial attention map is multiplied element-wise with the first splicing feature to obtain the output feature of the spatial flow branch; The channel fusion unit is used to add the output features of the frequency flow branch and the output features of the spatial flow branch element by element to obtain the added features; the added features are then processed sequentially through global average pooling, multilayer perceptron and activation function to obtain the channel weight vector; the channel weight vector is then multiplied with the added features channel by channel to obtain the channel fusion features.

4. The method according to claim 3, characterized in that, The spectrum sensing branch is used for: The high-level semantic abstract features are processed sequentially through a two-dimensional fast Fourier transform unit, a global spectrum self-attention unit, and a two-dimensional fast Fourier transform unit to obtain spectral features. The global spectrum self-attention unit performs frequency domain transformation and self-attention calculation on the output features of the two-dimensional fast Fourier transform unit to achieve efficient capture of global context information and enhancement of frequency domain features. The amplitude guides the sparse branch for: The high-level semantic abstract features are divided into S×S local grids, and an importance map reflecting each local grid is generated by a module length-aware routing operator. High-response regions are selected based on the importance map, and these high-response regions are clustered to obtain clustered region features. The aggregated region features are then refined using a sliding window hole attention operator to obtain enhanced region features. The enhanced regional features are filled back into the original positions of the high-response regions in the high-level semantic abstract features by the inverse mapping feature restoration operation, resulting in sparse spatial features. The local aggregation branch is used for: The high-level semantic abstract features are decoupled into a first feature subgroup and a second feature subgroup by channel grouping operation; The first feature subgroup is input and subjected to 3×3 convolution to obtain multi-scale spatial features; The second feature subgroup is input and subjected to 1×1 convolution to obtain point-by-point channel features; By using residual connections, the spatial weight map generated from the multi-scale spatial features is injected into the point-by-point channel features to obtain the first interactive enhancement feature. The channel weights extracted from the point-by-point channel features are then injected back into the multi-scale spatial features to obtain the second interactive enhancement feature. The first interaction enhancement feature and the second interaction enhancement feature are concatenated and fused to obtain a local aggregated feature.

5. The method according to claim 4, characterized in that, The module length-aware routing operator is used for: Perform linear mapping on the features within each local grid after partitioning to generate query vectors and key vectors; The L2 norm of the query vector is calculated by the modulus calculation module, and the L2 norm is input into the dynamic factor generator to generate a dynamic scaling factor. and dynamic offset factor ; The query vector and the key vector are respectively mapped to a high-dimensional space through a kernel function, and linear attention is calculated to obtain the linear attention calculation result; The linear attention calculation result is compared with the dynamic scaling factor. Perform multiplication modulation to generate the multiplication modulation result; The multiplication modulation result is combined with the dynamic offset factor. Additive modulation is performed to generate an importance map that reflects the importance of each local grid.

6. The method according to claim 5, characterized in that, The gated fusion unit is used for: The spectral features are processed by an activation function to generate gating weights; The first intermediate feature is obtained by performing an element-wise multiplication operation between the gating weight and the sparse space feature. The gating weights are multiplied element-wise with the local aggregated features to obtain the second intermediate features; The spectral features, the first intermediate feature, and the second intermediate feature are added, fused, and normalized to obtain the enhanced high-level features.

7. The method according to claim 6, characterized in that, The global frequency domain modulation module is used for: After the enhanced high-level features are processed by a lightweight MLP branch and a learnable weight branch, the output features of the lightweight MLP branch and the output features of the learnable weight branch are fused to obtain dual-branch output features. The lightweight MLP branch includes global average pooling, fully connected layers, and activation functions to extract global semantic information of the enhanced high-level features and capture global dependency information of the image. The learnable weight branch includes feature aggregation for specific regions and learnable center weight encoding operations to aggregate local corner information of the enhanced high-level features and capture salient topological structures in the image. The bi-branch output features are sequentially processed through Fast Fourier Transform, frequency-domain selective scanning mechanism, and inverse Fast Fourier Transform to achieve secondary global semantic enhancement of the bi-branch output features, resulting in enhanced global features. The frequency-domain selective scanning mechanism is used to flatten the two-dimensional spectrum into a one-dimensional sequence, establish long-range dependencies between frequency components using the Mamba operator, and achieve secondary refinement of the semantics of the whole graph by selectively enhancing the target periodic frequency components and suppressing noise components. The feature alignment module includes: The wavelet-guided downsampling unit is used to decompose the first input feature into a low-frequency sub-band and high-frequency sub-bands in three directions (horizontal, vertical, and diagonal) through discrete wavelet transform, and then splices the four sub-bands in the channel dimension to generate a downsampling guided feature with resolution matching deep features. The spatial offset prediction unit is used to concatenate the downsampled guiding feature with the second input feature, and predict the spatial offset field and modulation mask through a convolutional layer; A deformable convolutional unit is used to perform pixel-level adaptive resampling of the second input features according to the spatial offset field, and to perform feature modulation using the modulation mask to output aligned features; The context fusion module includes a first-level fusion unit and a second-level fusion unit. The first-level fusion unit performs preliminary feature fusion and background noise filtering through splicing, convolutional layers, and activation function operations. The second-level fusion unit performs secondary purification and spatial location calibration of the primary fused features obtained by the first-level fusion unit through splicing, two convolutional layers, activation functions, and weighted fusion operations.

8. The method according to claim 7, characterized in that, The process of inputting the shallow alignment fusion features, mid-layer alignment fusion features, and enhanced alignment fusion features into the decoupled convolutional detection head to obtain the small target detection result includes: The shallow alignment fusion feature, the middle alignment fusion feature and the enhanced alignment fusion feature are respectively input into three decoupled convolutional branches. The three decoupled convolutional branches perform dense prediction respectively to generate shallow feature candidate detection results, middle feature candidate detection results and high-level feature candidate detection results. The shallow, mid-level, and high-level feature candidate detection results are merged, and redundant boxes are eliminated and optimally aggregated using a non-maximum suppression algorithm to obtain the small target detection results.

9. A small target detection device based on feature alignment and interaction, characterized in that, include: The shallow physical fidelity feature acquisition module is used to input the image into the two-layer W-LDS module and the Mamba sensing module for processing. It is used to retain the high-frequency details and physical pixel information of small targets while reducing the resolution, so as to obtain shallow physical fidelity features. The W-LDS module uses discrete wavelet transform decomposition to achieve lossless downsampling. The mid-level semantic enhancement feature acquisition module is used to process the shallow physical fidelity features through the Mamba perception module, and to capture long-distance dependencies and enhance the semantic representation of small targets through a dual-stream collaborative mechanism to obtain mid-level semantic enhancement features. The high-level semantic abstraction feature acquisition module is used to process the mid-level semantic enhancement features through the Mamba perception module to further expand the receptive field, suppress background noise, and extract deep semantic context to obtain high-level semantic abstraction features. The enhanced high-level feature acquisition module is used to process the high-level semantic abstract features through the dynamic interaction module to generate enhanced high-level features. The dynamic interaction module includes a spectrum sensing branch, an amplitude-guided sparse branch, a local aggregation branch, and a gated fusion unit. The multi-scale alignment and fusion module is used to input the shallow physical fidelity features, mid-level semantic enhancement features, and enhanced high-level features into EAF-Net for multi-scale feature alignment and fusion, generating shallow aligned fused features, mid-level aligned fused features, and enhanced aligned fused features. The EAF-Net includes a global frequency domain modulation module, a feature alignment module, and a context fusion module. The global frequency domain modulation module is used to perform secondary semantic enhancement on the features in the frequency domain. The feature alignment module is used to guide the features to perform spatial position correction. The context fusion module is used to realize the progressive complementary fusion of deep and shallow features. The target detection module is used to input the shallow alignment fusion feature, the middle alignment fusion feature and the enhanced alignment fusion feature into the decoupled convolutional detection head to obtain the small target detection result.

10. A server, characterized in that, The server includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the small target detection method based on feature alignment and interaction as described in any one of claims 1-8.

Images