Remote sensing image small target detection method and device based on frequency domain enhancement

Abstract

The invention provides a remote sensing image small target detection method and device based on frequency domain enhancement, and the method comprises the steps: extracting an input feature map based on an original remote sensing image through a first-level feature map layer of a feature pyramid architecture; determining the strongest activation point in the input feature map by adopting global maximum pooling; determining average response information of the input feature map by adopting global average pooling; gathering the strongest activation point and average response information through an attention mechanism, and generating a spatial sensitive feature weight; dynamically enhancing the high-frequency component through a learnable filter, obtaining a target space vector according to the enhanced high-frequency component, and generating a frequency domain channel attention weight; performing weighted fusion on the spatial features in the input feature map and the target spatial vector to obtain target features; connecting the target features and the average response information through residual connection to obtain an output feature map; and by adopting the synergistic effect of dynamic frequency domain filtering and an attention mechanism, the accuracy and robustness of target classification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image interpretation, and in particular to a remote sensing image small target detection method and device based on frequency domain enhancement. BACKGROUND

[0002] Target detection is one of the basic tasks in the field of remote sensing image interpretation, and plays an important role in practical applications such as intelligent monitoring, maritime management, precision agriculture, and urban planning. In recent years, with the rapid development of remote sensing technology, general target detection research has made significant progress. Small target detection is a subtask of general target detection. Because the small targets in unmanned aerial vehicle aerial images or satellite remote sensing images usually occupy only tens or even a few pixels, and the data set is often accompanied by complex background interference, target dense distribution, target type diversity, and significant scale difference, small target detection has gradually become a key and challenging research task in remote sensing image processing. Small target detection tasks have high requirements for feature expression ability.

[0003] The small target detection task is generally detected by a convolutional neural network combined with frequency domain technology. One method is to build a top-down feature pyramid based on the bottom-up default path of the convolutional neural network, extract features through the weight-sharing convolutional layer, and use down-sampling operations such as pooling or step convolution to build a hierarchical semantic structure. However, down-sampling reduces the spatial resolution of the feature map and weakens the expression of spatial information. The features of small targets are often hidden in high-order semantic information in the deep layers of the backbone network. The characteristics of low pixel ratio and weak feature expression of small targets lead to double challenges in frequency domain analysis: traditional attention mechanisms are difficult to effectively capture high-frequency detail features; and fixed frequency band filters cannot adapt to the frequency spectrum distribution characteristics of different scenes. SUMMARY

[0004] The present application provides a remote sensing image small target detection method and device based on frequency domain enhancement to solve the defects that it is difficult to effectively capture high-frequency detail features and cannot adapt to the frequency spectrum distribution characteristics of different scenes in the prior art, and to realize effective suppression of complex background interference, strengthening of high-frequency edge and texture information, and improvement of the accuracy and robustness of target classification. The present application provides a remote sensing image small target detection method based on frequency domain enhancement, comprising the following steps: Based on the original remote sensing image, the input feature map is extracted using the first level feature map layer of the feature pyramid architecture; wherein the feature pyramid architecture is obtained by removing the second level feature map layer of the feature model architecture and retaining the first level feature map layer of the feature model architecture; the level of the second level feature map layer is greater than the level of the first level feature map layer; Global max pooling is used to determine the strongest activation points in the input feature map; global average pooling is used to determine the average response information of the input feature map; and an attention mechanism is used to aggregate the strongest activation points and the average response information to generate spatially sensitive feature weights. The input feature map is transformed by discrete cosine transform to obtain high-frequency components; the high-frequency components are dynamically enhanced by a learnable filter; the target space vector is obtained based on the enhanced high-frequency components, and frequency domain channel attention weights are generated. The spatial features in the input feature map and the target spatial vector are weighted and fused according to the spatial sensitive feature weights and the frequency domain channel attention weights to obtain the target features; the target features and the average response information are connected through residual connections to obtain the output feature map.

[0005] The method for detecting small targets in remote sensing images based on frequency domain enhancement provided by this invention generates spatially sensitive feature weights by converging the strongest activation points and average response information through an attention mechanism, as achieved by the following formula: in, Weights for spatially sensitive features; Use the Sigmoid activation function; A bottleneck structure consisting of two fully connected layers is used to enhance the attention mechanism; This is the average response information; It is the strongest activation point.

[0006] The method for detecting small targets in remote sensing images based on frequency domain enhancement provided by this invention employs discrete cosine transform to transform the input feature map, obtaining high-frequency components; dynamically enhances the high-frequency components using a learnable filter; and obtains the target space vector based on the enhanced high-frequency components and generates frequency domain channel attention weights, including: The input feature map is divided into multiple groups of input features; the discrete cosine transform is used to project each group of input features into the frequency domain to obtain high-frequency components. The weights of different frequency components in the high-frequency component are adaptively learned through backpropagation using a learnable filter; the enhanced high-frequency component is obtained by multiplying the learnable filter with each element in the high-frequency component. The enhanced high-frequency components are restored to the spatial domain by inverse discrete cosine transform to obtain the target spatial vector; frequency domain channel attention weights are generated based on the target spatial vector through an attention mechanism.

[0007] The method for detecting small targets in remote sensing images based on frequency domain enhancement provided by the present invention further includes: By using dual-path lateral connection, the high-level semantic feature map and the low-level detail feature map in the output feature map are decoupled to obtain dynamic components and static components. Multiple dynamic components are concatenated and then input into a deformable convolutional layer to obtain aligned features; the aligned features and static components are then fused to obtain a fused feature map.

[0008] The remote sensing image small target detection method based on frequency domain enhancement provided by the present invention decouples the high-level semantic feature map and the low-level detail feature map in the output feature map through dual-path lateral connection to obtain dynamic and static components, including: Convolution and group normalization are performed on the high-level semantic feature map and low-level detail feature map in the output feature map to obtain the processed high-level semantic feature map and the processed low-level detail feature map. Specifically, the high-level semantic feature map and the low-level detail feature map in the output feature map are convolved and group normalized, which is achieved through the following formula: in, This is a normalization operation for groups of 32. It is a 1×1 convolution; This is a high-level semantic feature map; This is a low-level detail feature map; This is the processed high-level semantic feature map; This is the processed low-level detail feature map; By using a dual-path lateral connection along the channel dimension, the processed high-level semantic feature map and the processed low-level detail feature map are decoupled by channel to obtain dynamic and static components.

[0009] According to the frequency domain enhancement-based remote sensing image small target detection method provided by the present invention, multiple dynamic components are concatenated and input into a deformable convolutional layer to obtain aligned features, which is achieved by the following formula: in, Features after alignment; For deformable convolutional layer learning functions; , For dynamic components; These are deformable convolution parameters; The aligned features and static components are fused to obtain a fused feature map, which is achieved by the following formula: in, To fuse feature maps; , This is a static component; It is a 1×1 convolution.

[0010] This invention provides a small target detection device for remote sensing images based on frequency domain enhancement, comprising: The input feature map extraction module is used to extract the input feature map based on the original remote sensing image and using the first-level feature layer of the feature pyramid architecture. The feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture and retaining the first-level feature layer of the feature model architecture. The level of the second-level feature layer is greater than the level of the first-level feature layer. The spatially sensitive feature weight generation module is used to determine the strongest activation point in the input feature map using global max pooling; determine the average response information of the input feature map using global average pooling; and generate spatially sensitive feature weights by pooling the strongest activation point and the average response information through an attention mechanism. The frequency domain channel attention weight generation module is used to transform the input feature map using discrete cosine transform to obtain high-frequency components; the high-frequency components are dynamically enhanced by a learnable filter; the target space vector is obtained based on the enhanced high-frequency components; and frequency domain channel attention weights are generated. The output feature map determination module is used to perform weighted fusion of spatial features and target spatial vectors in the input feature map according to spatially sensitive feature weights and frequency domain channel attention weights to obtain target features; the target features and average response information are connected through residual connections to obtain the output feature map.

[0011] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the frequency domain enhancement-based remote sensing image small target detection method as described above.

[0012] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the frequency domain enhancement-based remote sensing image small target detection method as described above.

[0013] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the frequency domain enhancement-based remote sensing image small target detection method as described above.

[0014] The present invention provides a method and apparatus for small target detection in remote sensing images based on frequency domain enhancement. The method extracts an input feature map from the original remote sensing image using a first-level feature layer of a feature pyramid architecture. The feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture while retaining the first-level feature layer. The second-level feature layer has a higher level than the first-level feature layer. Global max pooling is used to determine the strongest activation points in the input feature map. Global average pooling is used to determine the average response information of the input feature map. An attention mechanism is used to aggregate the strongest activation points and the average response information to generate spatially sensitive feature weights. Discrete cosine transform is then applied to the input feature map to obtain... High-frequency components are dynamically enhanced using learnable filters. The enhanced high-frequency components are used to obtain the target spatial vector and generate frequency domain channel attention weights. The spatial features in the input feature map and the target spatial vector are weighted and fused based on the spatially sensitive feature weights and the frequency domain channel attention weights to obtain the target features. The target features and average response information are connected via residual connections to obtain the output feature map. A lightweight feature pyramid architecture is designed to simplify redundant network layers while retaining high-resolution feature layers, significantly enhancing the ability to express detailed features of small targets. The synergistic effect of dynamic frequency domain filtering and attention mechanisms effectively suppresses interference from complex backgrounds, strengthens high-frequency edge and texture information, and improves the accuracy and robustness of target classification. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the method for detecting small targets in remote sensing images based on frequency domain enhancement provided by the present invention.

[0017] Figure 2 This is a schematic diagram of the structure of the remote sensing image small target detection device based on frequency domain enhancement provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] The following is combined with Figures 1-3 This invention is described.

[0021] Figure 1 This is one of the flowcharts illustrating the frequency domain enhancement-based remote sensing image small target detection method provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Based on the original remote sensing image, extract the input feature map using the first-level feature layer of the feature pyramid architecture; wherein, the feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture and retaining the first-level feature layer of the feature model architecture; the level of the second-level feature layer is greater than the level of the first-level feature layer.

[0022] In step 101 above, an improved model architecture based on the YOLOv8 network was designed in this embodiment of the invention. In the YOLOv8 model, the deeper the feature layer level, for example, P4 and P5, the lower its spatial resolution, which is an inevitable result of multiple downsampling operations. This low resolution causes tiny targets in the original image to occupy only a few pixels or even a single pixel in the deep feature map, resulting in the severe loss or blurring of their key shape, edges, texture, and other detailed features, making them difficult to distinguish from the background, especially when small targets are densely distributed, making them more prone to sticking together. By removing the lowest resolution P4 and P5 layers and focusing on using shallower P2 and P3 layers with fewer downsampling operations for detection, higher resolution feature maps can be effectively preserved. In P2 and P3 layers, tiny targets occupy a relatively larger number of pixels, and their fine structure and local features, for example, edges, corners, and internal textures, are more clearly preserved and represented, providing the detector with richer and more accurate visual information to identify and locate these tiny targets, thereby directly alleviating the problem of difficult detail representation caused by insufficient spatial resolution of deep feature maps. Therefore, in this embodiment of the invention, only layers P2 and P3 are used for target detection.

[0023] The second-level feature layer is a higher-level layer in the model architecture; for example, levels P4 and P5 in a YOLOv8 model. The first-level feature layer is a lower-level layer in the model architecture; for example, levels P2 and P3 in a YOLOv8 model.

[0024] Step 102: Use global max pooling to determine the strongest activation point in the input feature map; use global average pooling to determine the average response information of the input feature map; and use an attention mechanism to aggregate the strongest activation point and the average response information to generate spatially sensitive feature weights.

[0025] In step 102 above, this embodiment of the invention proposes a feature enhancement module based on discrete cosine transform, which optimizes the feature representation of small targets through a collaborative mechanism of frequency domain dynamic modulation branch and spatial feature calibration branch; it adopts a dual-path hybrid architecture, which dynamically enhances the high-frequency components related to the target through spatially sensitive feature weights and frequency domain feature weights.

[0026] The spatial calibration branch utilizes both Global Average Pooling (GAP) and Global Max Pooling (GMP) convergence information to construct a channel attention mechanism, focusing on channels sensitive to high-frequency textures. This channel attention weight is used to calibrate the importance of feature channels.

[0027] Given the input feature maps (P2, P3) extracted by the backbone network. First, channel recombination is performed using 1×1 convolution: in For batch size, For the number of channels, and For the height and width of the input features, subsequent... The original dimensions are preserved to ensure compatibility with subsequent processing. A dual-channel attention unit is then designed. The module generates spatially sensitive feature weights.

[0028] Optionally, in step 102, the strongest activation points and average response information are aggregated through an attention mechanism to generate spatially sensitive feature weights, which is achieved through the following formula: in, Weights for spatially sensitive features; Use the Sigmoid activation function; A bottleneck structure consisting of two fully connected layers is used to enhance the attention mechanism; This is the average response information; It is the strongest activation point.

[0029] Global max pooling (GMP) directly reflects the strongest activation points in the feature map. These points often correspond to the locations with the richest high-frequency information, such as edges, corners, and texture details, or where the target is most prominent. Global average pooling (GAP), on the other hand, provides the average response of the entire feature map, representing the overall distribution of background and foreground, providing contextual information, and helping to highlight the prominence of high-frequency texture regions relative to the average background.

[0030] This represents a bottleneck structure consisting of two fully connected layers: Compression ratio It is used to balance computational efficiency and nonlinear modeling capabilities.

[0031] In step 102 above, the enhanced attention mechanism can generate more accurate and robust channel attention weights, and perform feature selection and recalibration more effectively.

[0032] Step 103: The input feature map is transformed using discrete cosine transform to obtain high-frequency components; the high-frequency components are dynamically enhanced using a learnable filter, and the target space vector is obtained based on the enhanced high-frequency components, and frequency domain channel attention weights are generated.

[0033] In step 103 above, the feature enhancement module first decomposes the input features to the frequency domain through discrete cosine transform, adaptively enhances the high-frequency components related to the target using a learnable filter, and reconstructs the spatial features with enhanced details through inverse transform.

[0034] In the frequency domain, a signal or image can be decomposed into components of different frequencies. Low-frequency components typically correspond to the overall trend or smooth areas in an image, while high-frequency components correspond to edges, details, and noise. In remote sensing images, the contours, edges, and details of small targets are usually mainly reflected in high-frequency information, and random noise in the image also often belongs to high-frequency components because they are irregularly distributed in space or the signal and change rapidly.

[0035] Optionally, step 103 includes steps 1031 to 1033: Step 1031: Divide the input feature map into multiple groups of input features; use discrete cosine transform to project each group of input features into the frequency domain to obtain high-frequency components.

[0036] Step 1032: Adaptively learn the weights of different frequency components in the high-frequency component through backpropagation using a learnable filter; multiply the learnable filter by each element in the high-frequency component to obtain the enhanced high-frequency component.

[0037] Step 1033: The enhanced high-frequency components are restored to the spatial domain by inverse discrete cosine transform to obtain the target spatial vector; frequency domain channel attention weights are generated based on the target spatial vector through an attention mechanism.

[0038] In steps 1031 to 1033 above, to achieve frequency domain feature modulation, the original input features are first divided into... Group, default To reduce computational complexity: Next, a two-dimensional DCT transform is performed on each set of features, and they are projected into the frequency domain: in, Using frequency domain coordinates, a learnable filter is then introduced. Enhance the details of the target texture, among which For element-wise multiplication: filter Initialized as a full-1 tensor, the importance of different frequency components is adaptively learned through backpropagation. When the input size changes, bilinear interpolation is used to dynamically adjust the filter size to maintain compatibility. The modulated features are reconstructed into the spatial domain using IDCT. Finally, through the dual-channel attention unit Generate frequency domain channel attention weights The calculation process of CA2 and They are the same, but the parameters are learned independently.

[0039] Step 104: Based on the spatial sensitive feature weights and frequency domain channel attention weights, the spatial features in the input feature map and the target spatial vector are weighted and fused to obtain the target features; the target features and the average response information are connected through residual connections to obtain the output feature map.

[0040] In step 104 above, the dual-branch output is generated through attention-guided weighted fusion, while to preserve the low-frequency background information of the original features, the final output is generated through residual connections: in, This is for outputting feature maps.

[0041] By connecting the outputs of the frequency domain dynamic modulation branch and the spatial feature calibration branch to the original feature residual after attention-weighted fusion, high-frequency details are enhanced in a targeted manner while suppressing background noise. This effectively solves the feature inaccuracy problem caused by the decoupling of fixed-band filtering and spatial attention mechanism in traditional methods.

[0042] Optionally, the method for detecting small targets in remote sensing images based on frequency domain enhancement provided by the present invention further includes steps A and B: Step A: Decouple the high-level semantic feature map and the low-level detail feature map in the output feature map by using dual-path lateral connection to obtain dynamic components and static components; Step B: Concatenate multiple dynamic components and input them into a deformable convolutional layer to obtain aligned features; fuse the aligned features and static components to obtain a fused feature map.

[0043] In steps A and B above, to address the spatial misalignment problem during feature fusion between adjacent layers of the Feature Pyramid Network (FPN) in small target detection tasks in remote sensing images, this invention proposes a Dynamic Feature Alignment Module (DFAM). For example... Figure 3 As shown, this module significantly improves the spatial alignment accuracy of cross-level features through a collaborative design of feature decoupling and recombination with adaptive deformation learning. The dynamic feature alignment module, based on deformable convolution, aims to address the feature alignment problem in remote sensing images caused by direct upsampling operations in feature pyramid networks. Upsampling operations in traditional feature pyramid networks can lead to misalignment between feature maps of different scales, thus affecting the detection of small targets. By combining the alignment mechanism of deformable convolution, this module can accurately align features at different levels, effectively avoiding feature misalignment and improving detection accuracy in complex backgrounds and occlusion conditions.

[0044] DFAM uses high-level semantic feature maps from adjacent layers of FPN. and low-level detail feature map As input, the module employs a channel decoupling strategy to balance feature deformation adaptability and geometric stability.

[0045] In step B above, considering the anisotropic nature of remote sensing targets, the module uses an improved Deformable Convolution v2 as the alignment core.

[0046] Optionally, step A includes: Step A1: Perform convolution and group normalization on the high-level semantic feature map and low-level detail feature map in the output feature map to obtain the processed high-level semantic feature map and the processed low-level detail feature map. Specifically, the high-level semantic feature map and the low-level detail feature map in the output feature map are convolved and group normalized, which is achieved through the following formula: in, This is a normalization operation for groups of 32. It is a 1×1 convolution; This is a high-level semantic feature map; This is a low-level detail feature map; This is the processed high-level semantic feature map; This is the processed low-level detail feature map; Step A2: By using dual-path lateral connections along the channel dimension, channel decoupling is performed on the processed high-level semantic feature map and the processed low-level detail feature map to obtain dynamic and static components.

[0047] In steps A1 to A2 above, by setting the number of groups for the group normalization operation to 32, experiments show that this configuration can effectively stabilize the feature distribution during mini-batch training, as evidenced by ablation experiments. Along the channel dimension... Divided into dynamic components With static components The split ratio is 1:1. The dynamic component is used to learn spatial deformation parameters, while the static component retains the original geometric structure information. The two are decoupled through parallel paths.

[0048] Optionally, step B includes steps B1 to B2: Step B1: After concatenating multiple dynamic components, input them into a deformable convolutional layer to obtain aligned features using the following formula: in, Features after alignment; For deformable convolutional layer learning functions; , For dynamic components; These are deformable convolution parameters; Step B2: Fuse the aligned features and static components to obtain a fused feature map, achieved using the following formula: in, To fuse feature maps; , This is a static component; It is a 1×1 convolution.

[0049] In step B1 above, It is configured with a 3×3 convolution kernel and 8 sets of deformable offsets. The offset range is limited. By applying L2 regularization constraints to the offset to avoid excessive deformation leading to feature distortion, this design enables the module to adaptively adjust the receptive field, exhibiting more accurate feature responses in small target areas such as ships and vehicles.

[0050] In step B2 above, to preserve the detailed information of cross-level features, the module fuses the dynamically aligned features with the static components. Here, the 1×1 convolution increases the number of channels from... Compress to This reduces computational complexity while enabling cross-level information integration to generate the final fused feature map.

[0051] This invention provides a method for small target detection in remote sensing images based on frequency domain enhancement. The method extracts an input feature map from the original remote sensing image using a first-level feature layer of a feature pyramid architecture. The feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture while retaining the first-level feature layer. The second-level feature layer has a higher level than the first-level feature layer. Global max pooling is used to determine the strongest activation points in the input feature map. Global average pooling is used to determine the average response information of the input feature map. An attention mechanism is used to aggregate the strongest activation points and the average response information to generate spatially sensitive feature weights. Discrete cosine transform is applied to the input feature map to obtain high-resolution features. The system dynamically enhances high-frequency components using learnable filters, obtains the target spatial vector based on the enhanced high-frequency components, and generates frequency domain channel attention weights. It then weights and fuses the spatial features in the input feature map and the target spatial vector using spatially sensitive feature weights and frequency domain channel attention weights to obtain target features. Finally, it connects the target features and average response information through residual connections to obtain the output feature map. A lightweight feature pyramid architecture is designed to simplify redundant network layers while retaining high-resolution feature layers, significantly enhancing the ability to express detailed features of small targets. The synergistic effect of dynamic frequency domain filtering and attention mechanisms effectively suppresses interference from complex backgrounds, strengthens high-frequency edge and texture information, and improves the accuracy and robustness of target classification.

[0052] The following describes the frequency-domain-enhanced remote sensing image small target detection device provided by the present invention. The frequency-domain-enhanced remote sensing image small target detection device described below can be referred to in correspondence with the frequency-domain-enhanced remote sensing image small target detection method described above.

[0053] Figure 2 This is one of the flowcharts of the remote sensing image small target detection device based on frequency domain enhancement provided by the present invention, such as... Figure 2 As shown, the device includes the following: The input feature map extraction module 201 is used to extract the input feature map based on the original remote sensing image and using the first-level feature layer of the feature pyramid architecture. The feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture and retaining the first-level feature layer of the feature model architecture. The level of the second-level feature layer is greater than the level of the first-level feature layer.

[0054] The spatially sensitive feature weight generation module 202 is used to determine the strongest activation point in the input feature map by using global max pooling; to determine the average response information of the input feature map by using global average pooling; and to generate spatially sensitive feature weights by pooling the strongest activation point and the average response information through an attention mechanism.

[0055] The frequency domain channel attention weight generation module 203 is used to transform the input feature map using discrete cosine transform to obtain high-frequency components; to dynamically enhance the high-frequency components through a learnable filter; and to obtain the target space vector and generate frequency domain channel attention weights based on the enhanced high-frequency components.

[0056] The output feature map determination module 204 is used to perform weighted fusion of spatial features and target spatial vectors in the input feature map according to spatial sensitive feature weights and frequency domain channel attention weights to obtain target features; and to connect the target features and average response information through residual connections to obtain the output feature map.

[0057] This invention provides a remote sensing image small target detection device based on frequency domain enhancement. It extracts an input feature map from the original remote sensing image using a first-level feature layer of a feature pyramid architecture. The feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture while retaining the first-level feature layer. The second-level feature layer has a higher level than the first-level feature layer. Global max pooling is used to determine the strongest activation points in the input feature map. Global average pooling is used to determine the average response information of the input feature map. An attention mechanism is used to aggregate the strongest activation points and the average response information to generate spatially sensitive feature weights. Discrete cosine transform is used to transform the input feature map to obtain high-resolution features. The system dynamically enhances high-frequency components using learnable filters, obtains the target spatial vector based on the enhanced high-frequency components, and generates frequency domain channel attention weights. It then weights and fuses the spatial features in the input feature map and the target spatial vector using spatially sensitive feature weights and frequency domain channel attention weights to obtain target features. Finally, it connects the target features and average response information through residual connections to obtain the output feature map. A lightweight feature pyramid architecture is designed to simplify redundant network layers while retaining high-resolution feature layers, significantly enhancing the ability to express detailed features of small targets. The synergistic effect of dynamic frequency domain filtering and attention mechanisms effectively suppresses interference from complex backgrounds, strengthens high-frequency edge and texture information, and improves the accuracy and robustness of target classification.

[0058] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute a frequency-domain enhanced remote sensing image small target detection method.

[0059] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the frequency domain enhancement-based remote sensing image small target detection method provided by the above methods.

[0061] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the frequency-domain-enhanced remote sensing image small target detection method provided by the above methods.

[0062] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0063] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting small targets in remote sensing images based on frequency domain enhancement, characterized in that, include: Based on the original remote sensing image, the input feature map is extracted using the first-level feature layer of the feature pyramid architecture; wherein, the feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture and retaining the first-level feature layer of the feature model architecture; the level of the second-level feature layer is greater than the level of the first-level feature layer. The strongest activation point in the input feature map is determined by global max pooling; the average response information of the input feature map is determined by global average pooling; and the strongest activation point and the average response information are aggregated by an attention mechanism to generate spatially sensitive feature weights. The input feature map is transformed by discrete cosine transform to obtain high-frequency components; the high-frequency components are dynamically enhanced by a learnable filter; the target space vector is obtained based on the enhanced high-frequency components, and frequency domain channel attention weights are generated. The spatial features in the input feature map and the target spatial vector are weighted and fused according to the spatial sensitive feature weights and the frequency domain channel attention weights to obtain the target features; the target features and the average response information are connected through residual connections to obtain the output feature map.

2. The method for detecting small targets in remote sensing images based on frequency domain enhancement according to claim 1, characterized in that, The process of converging the strongest activation points and the average response information through an attention mechanism to generate spatially sensitive feature weights is achieved through the following formula: in, Weights for spatially sensitive features; Use the Sigmoid activation function; A bottleneck structure consisting of two fully connected layers is used to enhance the attention mechanism; This is the average response information; It is the strongest activation point.

3. The method for detecting small targets in remote sensing images based on frequency domain enhancement according to claim 1, characterized in that, The input feature map is transformed using discrete cosine transform to obtain high-frequency components; The high-frequency components are dynamically enhanced using a learnable filter. A target space vector is obtained based on the enhanced high-frequency components, and frequency domain channel attention weights are generated, including: The input feature map is divided into multiple groups of input features; each group of input features is projected into the frequency domain using discrete cosine transform to obtain high-frequency components. The weights of different frequency components in the high-frequency component are adaptively learned through backpropagation using a learnable filter; the enhanced high-frequency component is obtained by multiplying the learnable filter with each element of the high-frequency component. The enhanced high-frequency components are restored to the spatial domain by inverse discrete cosine transform to obtain the target spatial vector; frequency domain channel attention weights are generated based on the target spatial vector through an attention mechanism.

4. The method for detecting small targets in remote sensing images based on frequency domain enhancement according to claim 1, characterized in that, Also includes: By using dual-path lateral connection, the high-level semantic feature map and the low-level detail feature map in the output feature map are decoupled to obtain dynamic components and static components. The multiple dynamic components are concatenated and then input into a deformable convolutional layer to obtain aligned features. The aligned features and the static components are fused to obtain a fused feature map.

5. The method for detecting small targets in remote sensing images based on frequency domain enhancement according to claim 4, characterized in that, The process of decoupling the high-level semantic feature map and the low-level detail feature map in the output feature map through dual-path lateral connection to obtain dynamic and static components includes: Convolution and group normalization are performed on the high-level semantic feature map and low-level detail feature map in the output feature map to obtain the processed high-level semantic feature map and the processed low-level detail feature map. The convolution and group normalization processing of the high-level semantic feature map and low-level detail feature map in the output feature map is achieved through the following formula: in, This is a normalization operation for groups of 32. It is a 1×1 convolution; This is a high-level semantic feature map; This is a low-level detail feature map; This is the processed high-level semantic feature map; This is the processed low-level detail feature map; By using a dual-path lateral connection along the channel dimension, the processed high-level semantic feature map and the processed low-level detail feature map are decoupled by channel to obtain dynamic and static components.

6. The method for detecting small targets in remote sensing images based on frequency domain enhancement according to claim 4, characterized in that, The process of concatenating multiple dynamic components and inputting them into a deformable convolutional layer to obtain aligned features is achieved through the following formula: in, Features after alignment; For deformable convolutional layer learning functions; , For dynamic components; These are deformable convolution parameters; The fusion of the aligned features and the static components to obtain the fused feature map is achieved through the following formula: in, To fuse feature maps; , This is a static component; It is a 1×1 convolution.

7. A small target detection device for remote sensing images based on frequency domain enhancement, characterized in that, include: The input feature map extraction module is used to extract the input feature map based on the original remote sensing image and using the first-level feature layer of the feature pyramid architecture; wherein, the feature pyramid architecture is obtained by removing the second-level feature layer of the feature model architecture and retaining the first-level feature layer of the feature model architecture; the level of the second-level feature layer is greater than the level of the first-level feature layer. The spatially sensitive feature weight generation module is used to determine the strongest activation point in the input feature map by global max pooling; to determine the average response information of the input feature map by global average pooling; and to generate spatially sensitive feature weights by converging the strongest activation point and the average response information through an attention mechanism. The frequency domain channel attention weight generation module is used to transform the input feature map using discrete cosine transform to obtain high-frequency components; dynamically enhance the high-frequency components through a learnable filter; obtain the target space vector based on the enhanced high-frequency components; and generate frequency domain channel attention weights. The output feature map determination module is used to perform weighted fusion of the spatial features in the input feature map and the target spatial vector according to the spatial sensitive feature weight and the frequency domain channel attention weight to obtain the target features; and to connect the target features and the average response information through residual connection to obtain the output feature map.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for detecting small targets in remote sensing images based on frequency domain enhancement as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for detecting small targets in remote sensing images based on frequency domain enhancement as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for detecting small targets in remote sensing images based on frequency domain enhancement as described in any one of claims 1 to 6.

Images