An aerial point target detection method and system based on high-frequency perception and dynamic scaling loss

Abstract

The present application belongs to the technical field of computer vision and remote sensing image processing, and provides a kind of aerial point target detection method and system based on high-frequency perception and dynamic scaling loss, comprising: receiving aerial image, extracting multi-scale spatial features containing high-frequency details using high-resolution Mamba backbone network;The multi-scale spatial features are fused by PANet neck network to generate a multi-scale fusion feature map;The multi-scale fusion feature map is enhanced using a high-frequency refocusing module HFRM;The enhanced multi-scale fusion feature map is used as input to predict the position and category of the target through decoupling detection head;In the model training stage, dynamic scaling loss function is used for model optimization, and the loss function dynamically adjusts the weights of classification loss and regression loss according to the target real area area.The present application effectively solves the detection problem of point targets in aerial images caused by semantic collapse, high-frequency information loss and unstable training.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing image processing technology, specifically a method and system for detecting aerial point targets based on high-frequency sensing and dynamic scaling loss. Background Technology

[0002] With the widespread adoption of drone technology, pedestrian detection in aerial images has become crucial for public safety, emergency search and rescue, and other fields. However, from a high-altitude perspective, pedestrian targets degenerate into dots, short lines, or small rectangles due to extreme feature compression, retaining only slight intensity or texture differences from the background. This leads to a sharp decline in the performance of traditional convolutional neural network detectors that rely on mid-level features. This phenomenon, known as "semantic collapse," has become a core challenge in the field of object detection.

[0003] Although existing methods have attempted to improve small target detection through multi-scale feature fusion, attention mechanisms, or super-resolution techniques, two fundamental shortcomings remain when dealing with aerial point targets: First, the inherent low-pass filtering effect of deep networks weakens key high-frequency details layer by layer, while the recognition of point targets relies precisely on these subtle edges and grayscale changes; second, general training strategies and loss functions do not fully consider extreme scale differences and annotation noise, resulting in the loss gradient dominating the optimization process for large targets, while for small targets, even slight annotation deviations can cause drastic gradient fluctuations, making training extremely unstable.

[0004] Therefore, how to effectively represent and amplify high-frequency features, and how to achieve stable and efficient supervised training for small targets, have become two major bottlenecks restricting the performance of contemporary aerial pedestrian detection systems. In existing technologies, CNN-based detectors, due to the stacking of their convolutional and downsampling layers, essentially act as low-pass filters, attenuating high-frequency signals in the image. However, the subtle edges and grayscale variations of point targets are precisely the high-frequency features necessary for their recognition. Furthermore, traditional regression loss functions (such as CIoU loss) treat targets of all scales equally, resulting in larger targets generating gradients of orders of magnitude higher due to their wider absolute error range. This dominates the optimization process, drowning out the supervised signals for small targets. At the same time, the bounding box annotations of point targets have inherent ambiguity; a single pixel offset can cause drastic changes in the Intersection over Union (IoU) value, generating extremely unstable gradients. This label noise severely hinders the effective convergence of the model.

[0005] Therefore, the present invention provides a method and system for aerial point target detection based on high-frequency sensing and dynamic scaling loss. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: an aerial point target detection method based on high-frequency sensing and dynamic scaling loss, comprising the following steps:

[0008] The system receives aerial images and uses a high-resolution Mamba backbone network to extract multi-scale spatial features containing high-frequency details.

[0009] The multi-scale spatial features are fused using the PANet neck network to generate a multi-scale fused feature map;

[0010] The multi-scale fused feature map is enhanced using a high-frequency refocusing module (HFRM). The enhancement process includes:

[0011] The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT.

[0012] The enhanced multi-scale fused feature map is used as input to predict the location and category of the target by decoupling the detection head.

[0013] As a further technical solution of the present invention: the high-resolution Mamba backbone network includes an initial patch embedding layer and three layered stages;

[0014] The process of extracting multi-scale spatial features containing high-frequency details is as follows:

[0015] The input image I is first downsampled through a convolutional layer to obtain features. Figure X 0; The three hierarchical stages output multi-scale feature maps with downsampling rates of 8x, 16x, and 32x, respectively. The output of the i-th stage is:

[0016]

[0017] Each MambaStage i It consists of a downsampling layer and Ni sequential HG-MFTBlocks, used to model global context dependencies while preserving high-frequency details, where, :No. Feature maps output from each hierarchical stage At that time, the corresponding multi-scale output feature map, The multi-scale feature maps output by the backbone network correspond to the input image at different scales. downsampling scale No. The overall mapping function for the hierarchical stage is used to summarize all combinations of operators within that stage. No. Feature maps output from each hierarchical stage :No. The specific network structure of the layered stage, Ni: the number of HG-MFTBlock stacks in the i-th layered stage (positive integer hyperparameter), used to characterize the depth of the stage.

[0018] As a further technical solution of the present invention: the PANet neck network fuses multi-scale features through a dual-path mechanism, specifically including:

[0019] The top-down path starts from the deepest feature map C5 and gradually propagates high-level semantic information to shallower layers through upsampling and lateral connections, generating a semantically enhanced feature pyramid. The bottom-up path starts from the shallowest feature map P3, and injects low-level spatial details into deeper layers through downsampling and connections, ultimately outputting a multi-scale fused feature map. The calculation process is as follows:

[0020] ,in N represents intermediate features. :No. Layer Convolution, used for channel alignment to achieve feature fusion. It is a scale-level index in a multi-scale feature pyramid, used to distinguish feature layers of different resolutions. This indicates that the top-to-bottom path of PANet is at the [number]th [level]. Feature maps output at each scale express Indexed collection, Downsampling operators are used to extract intermediate features at higher resolution. Downsample to a spatial size consistent with the current scale features. It is the intermediate fused feature map of the previous layer scale in the PANet bottom-up path. : A convolutional module used for feature fusion and refinement, representing the fusion transformation. : A convolutional module used for feature fusion and refinement, representing the output refinement process, index. : indicates a multi-scale hierarchical index, typically This indicates the generation of layers through a gradual fusion process from shallow to deep within the bottom-to-top path. .

[0021] As a further technical solution of the present invention: the specific process by which the high-frequency refocusing module HFRM enhances the multi-scale fused feature map is as follows:

[0022] The multi-scale fused feature map is transformed to the frequency domain through fast Fourier transform to obtain a complex spectral tensor, which is then centered.

[0023] Using a binary low-pass mask, the centered spectrum is separated into low-frequency and high-frequency components through Hadamard product;

[0024] Design a channel attention function to learn the channel-specific amplification factor based on the amplitude spectrum of the high-frequency components. Let W represent the existence of a C-dimensional real vector, where each dimension corresponds to an intensity coefficient that amplifies the high-frequency components of the i-th feature channel, and W represents the channel-specific amplification coefficient. Let C represent the set of real numbers, and let C represent the number of channels in the feature map. The amplification factor is used to dynamically enhance the high-frequency components to obtain the enhanced high-frequency components.

[0025] The enhanced high-frequency components are recombined with the original low-frequency components to obtain the enhanced spectrum;

[0026] By performing inverse centering and inverse fast Fourier transform, the enhanced spectrum is converted back to the spatial domain, resulting in a multi-scale fusion feature map with enhanced details.

[0027] As a further technical solution of the present invention: the channel attention function includes global average pooling of the amplitude spectrum of high-frequency components to extract channel descriptors, and then processing the channel descriptors through a multilayer perceptron network composed of one-dimensional convolutional layers to generate channel-specific amplification coefficients. Let W represent the existence of a C-dimensional real vector, where each dimension corresponds to an intensity coefficient that amplifies the high-frequency components of the i-th feature channel, and W represents the channel-specific amplification coefficient. Let C represent the set of real numbers, and let C represent the number of channels in the feature map.

[0028] As a further technical solution of the present invention, it also includes the following steps:

[0029] A dynamically scaled loss function is used to optimize the overall HFA-Mamba model framework during end-to-end training. The loss function dynamically adjusts the weights of classification loss and regression loss based on the area of ​​the target's true region.

[0030] As a further technical solution of the present invention: the dynamic scaling loss function includes a regression loss weight function. A classification loss weight function .

[0031] As a further technical solution of the present invention: the regression loss weight function It is a negative exponential decay function with respect to the area A of the target real region, and its calculation expression is: Where A is the area of ​​the target region, α R ε is a positive hyperparameter that controls the decay rate, ε is a small positive constant to prevent division by zero, i is the number assigned to all positive samples in a training iteration, and Ai is the area calculated for each i-th target instance.

[0032] As a further technical solution of the present invention: the classification loss weight function It is an exponential enhancement function with respect to the area A of the target true region, and its calculation expression is: Where A is the area of ​​the target region, β C and γ C is a positive hyperparameter, i is the number assigned to all positive samples in one training iteration, and Ai is the area calculated for each i-th target instance.

[0033] An aerial point target detection system based on high-frequency sensing and dynamic scaling loss includes the following modules:

[0034] The feature extraction module receives aerial images and uses a high-resolution Mamba backbone network to extract multi-scale spatial features containing high-frequency details.

[0035] The feature fusion module fuses the multi-scale spatial features through the PANet neck network to generate a multi-scale fused feature map;

[0036] The feature enhancement module uses a high-frequency refocusing module (HFRM) to enhance the multi-scale fused feature map. The enhancement process includes:

[0037] The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT.

[0038] The target detection output module takes the enhanced multi-scale fused feature map as input and predicts the target's location and category by decoupling the detection head;

[0039] The global detection optimization module uses a dynamically scaled loss function to optimize the overall HFA-Mamba model framework during end-to-end training.

[0040] The beneficial effects of this invention are as follows:

[0041] 1. This invention constructs a high-resolution Mamba backbone network, which preserves high-frequency information crucial for point target detection by minimizing downsampling, and at the same time utilizes the linear complexity of the Mamba model to efficiently model the global context, thereby alleviating the inherent detail degradation problem in deep CNNs;

[0042] 2. This invention designs a high-frequency refocusing module (HFRM), which operates in the frequency domain and uses Fast Fourier Transform (FFT) to accurately amplify key high-frequency feature signals, thereby breaking the performance bottleneck of spatial domain convolution in processing fine details.

[0043] 3. This invention designs a Dynamic Scaling Loss (DSL) function, which dynamically balances the training focus by associating regression and classification weights with the target size, thereby robustly handling the challenges of extreme scale and label noise in small target detection. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart of an aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to an embodiment of the present invention;

[0046] Figure 2 This is a flowchart of an aerial point target detection system based on high-frequency sensing and dynamic scaling loss according to an embodiment of the present invention.

[0047] Figure 3 This is a schematic diagram of the overall architecture of the HFA-Mamba framework of this invention;

[0048] Figure 4 This is a detailed architectural diagram of the hierarchical gated Mamba feature transformer module (HG-MFTBlock) used in the Mamba backbone network of this invention;

[0049] Figure 5 This is a detailed architectural diagram of the high-frequency refocusing module (HFRM) proposed in this invention;

[0050] Figure 6 This is a schematic diagram of the heat map of the guided object of the high frequency refocusing module (HFRM) of the present invention;

[0051] Figure 7 This is a visual comparison chart of the detection results of HFA-Mamba on the AFO dataset in this invention;

[0052] Figure 8 This is a visualization comparison of the detection results of HFA-Mamba on the TinyPerson dataset in this invention;

[0053] Figure 9 This is a visualization comparison of the detection results of HFA-Mamba on the Heridal dataset of this invention;

[0054] Figure 10This is a diagram showing the field test results of the HFA-Mamba invention in a real-world scenario. Detailed Implementation

[0055] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0056] Example 1: Please refer to Figure 1 As shown in the figure, the aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to an embodiment of the present invention includes the following steps:

[0057] Step S10: Receive aerial images and extract multi-scale spatial features containing high-frequency details using a high-resolution Mamba backbone network.

[0058] In step S10, the main architecture of the high-resolution Mamba backbone network includes an initial patch embedding layer and a series of layered phases;

[0059] In step S10, the process of extracting multi-scale spatial features containing high-frequency details is as follows:

[0060] Using aerial images as input, input image I is first processed by a convolutional layer. The convolutional layer performs initial spatial downsampling and transforms the image into a sequence of feature patches suitable for hierarchical processing. in This is the initial feature map; the backbone network is constructed into three main feature extraction stages, producing feature maps with progressively decreasing spatial resolution, specifically downsampling rates of 8x, 16x, and 32x relative to the original input image; the multi-scale output is represented as... , where the i-th hierarchical stage ( The calculation process for ) is as follows: Each MambaStage i It consists of a downsampling layer and Ni sequential HG-MFTBlocks, used to model global context dependencies while preserving high-frequency details, where, :No. Feature maps output from each hierarchical stage At that time, the corresponding multi-scale output feature map, The multi-scale feature maps output by the backbone network correspond to the input image at different scales. downsampling scale No. The overall mapping function for the hierarchical stage is used to summarize all combinations of operators within that stage. No. Feature maps output from each hierarchical stage :No. The specific network structure of the hierarchical stage, Ni: the number of HG-MFTBlock stacks in the i-th hierarchical stage, used to characterize the depth of the stage;

[0061] It should be noted that, see also Figure 4 As shown, HG-MFTBlock plays a crucial role in performing deep feature extraction, capturing complex local patterns, and modeling broad global contextual understanding through state-space mechanisms at their respective scales. Ultimately, this hierarchical backbone network robustly outputs three key feature maps: in Where D i Indicates the downsampling factor as The channel dimension of the feature map, this set of semantically rich multi-scale feature maps is then accurately forwarded to the subsequent feature fusion neck for further processing;

[0062] Understandably, the significance of step S10 lies in its construction of a high-resolution Mamba backbone network. Its core function is to extract multi-scale spatial features rich in high-frequency details and global context from the input aerial image. By strictly limiting downsampling operations, this network effectively alleviates the problem of high-frequency information loss caused by multiple pooling and convolutions in traditional deep convolutional networks, preserving the subtle edge and texture features upon which point targets depend. Simultaneously, the introduction of a state-space model-based Mamba module efficiently models long-distance dependencies with linear complexity, giving the network powerful global perception capabilities and providing a feature foundation for subsequent processing that includes both fine local details and broad semantic understanding. The final output feature map... This forms a multi-scale feature pyramid, providing suitable semantic and spatial information for the detection of targets of different sizes.

[0063] Step S20: The multi-scale spatial features are fused through the PANet neck network to generate a multi-scale fused feature map;

[0064] In step S20, the PANet neck network strategically enhances the feature pyramid through a dual-path mechanism, which includes a top-down path for semantic propagation and a bottom-up path for spatial detail flow.

[0065] In step S20, the process of generating the multi-scale fused feature map is as follows:

[0066] The top-down path starts from the deepest, semantically richest feature map. C Starting with level 5, through a series of upsampling operations and lateral connections, high-level semantic information is gradually propagated to shallower layers. For level i (where... The feature map is used to generate semantically enhanced features. The calculation process is as follows:

[0067] Subsequently, a refined feature of level i is generated by applying another C2f module. This iterative process generates a feature pyramid rich in deep semantic context. .

[0068] The bottom-up approach aims to inject precise, fine-grained localization information, starting from the shallowest semantic enhancement features. Initially, low-level spatial details are propagated to deeper layers. For level j (where... The feature map is used to generate spatially refined features. The calculation process is as follows: Final output features Obtained after another C2f refinement: This iterative process ultimately results in a carefully fused set of multi-scale feature maps. ;

[0069] Among them, N represents intermediate features. :No. Layer Convolution, used for channel alignment to achieve feature fusion. It is a scale-level index in a multi-scale feature pyramid, used to distinguish feature layers of different resolutions. This indicates that the top-to-bottom path of PANet is at the [number]th [level]. Feature maps output at each scale express Indexed collection, Downsampling operators are used to extract intermediate features at higher resolution. Downsample to a spatial size consistent with the current scale features. It is the intermediate fused feature map of the previous layer scale in the PANet bottom-to-top path. : A convolutional module used for feature fusion and refinement, representing the fusion transformation. : A convolutional module used for feature fusion and refinement, representing the output refinement process, index. : indicates a multi-scale hierarchical index, typically This indicates the generation of layers through a gradual fusion process from shallow to deep within the bottom-to-top path. .

[0070] Understandably, the significance of step S20 lies in its efficient fusion of multi-scale features through the PANet neck network. Its core lies in constructing a top-down semantic propagation path and a bottom-up detail enhancement path, thereby deeply fusing deep semantic information with shallow spatial details. The top-down path transmits rich semantic information from high-level features to the shallow layer, improving the ability to classify small targets; the bottom-up path injects precise localization information from shallow features into the deep layer, enhancing the regression accuracy of target bounding boxes. This bidirectional feature pyramid structure significantly enhances the model's ability to represent point-like targets with drastic scale changes and weak details, outputting a set of multi-scale fused feature maps with balanced semantics and details. This provides high-quality, complete input features for subsequent frequency domain enhancement modules.

[0071] Step S30: Enhance the multi-scale fused feature map using the high-frequency refocusing module HFRM. The enhancement process includes:

[0072] The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT.

[0073] In step S30, the process of transforming the multi-scale fused feature map to the frequency domain using FFT is as follows:

[0074] Receive multi-scale fused feature maps from the path aggregation neck output of the PANet neck network. By each channel Independent application of the two-dimensional discrete Fourier transform (DFT) operator This achieves the transformation to the frequency domain, which produces a complex-valued spectral tensor. , where each element The calculation is as follows:

[0075] in (x,y) are spatial coordinates, (u,v) are frequency coordinates, F(u,v,c): the complex value of the c-th channel at coordinate (u,v) in the frequency domain, P(x,y,c): the pixel value of the c-th channel at coordinate (x,y) in the spatial domain, H,W: the height and width of the feature map. The kernel function of the Fourier transform;

[0076] Understandably, signal processing theory posits that low-frequency (LF) components of an image correspond to macroscopic structures (such as the background), while high-frequency (HF) components encode fine-grained details and abrupt changes (such as edges and textures). Point-like pedestrian targets, due to their small spatial extent and subtle intensity variations, are fundamentally encoded as local HF signal anomalies. The inherent low-pass filtering properties of deep CNNs systematically suppress these critical HF signals; this phenomenon is termed semantic collapse. Therefore, the motivation for the High-Frequency Refocusing Module (HFRM) is to operate in the frequency domain, an orthogonal representation that allows for explicit separation and direct manipulation of these HF / LF components. The architecture of the HFRM can be found in [reference needed]. Figure 5 As shown;

[0077] In step S30, the process of separating low-frequency and high-frequency components using a binary mask is as follows:

[0078] Construct a binary circular low-pass mask parameterized by radius r Using this mask, the spectrum is centered. Through Hadama accumulation ( It is strictly decomposed into its low-frequency and high-frequency components: , Where 1 is a tensor consisting entirely of 1s. F low The overall structural information was preserved, while F high This encapsulates the target details;

[0079] In step S30, the process of dynamically amplifying high-frequency components through the channel attention mechanism is as follows:

[0080] Design a lightweight channel attention function Used to learn the channel-specific amplification factor Let W represent the existence of a C-dimensional real vector, where each dimension corresponds to an intensity coefficient that amplifies the high-frequency components of the i-th feature channel, and W represents the channel-specific amplification coefficient. Let C represent the set of real numbers, and C represent the number of channels in the feature map. The function is global average pooling (...). ) applied to amplitude spectrum and a compact multilayer perceptron (

[0081] A complex of ) in and ;

[0082] The final enhancement operation employs residual additive gain to ensure robust preservation of intrinsic high-frequency information and adaptive amplification: in It is a version of w broadcast in space, which dynamically enhances the high-frequency components. With the original low-frequency component F retained low To reorganize;

[0083] In step S30, the process of returning the enhanced recombined spectrum to the spatial domain via IFFT is as follows:

[0084] Enhanced detail feature map By applying the inverse two-dimensional fast Fourier transform And extract the real part to recover: in It reversed the centralized operation;

[0085] The obtained feature map It is then passed to the downstream decoupling detection head;

[0086] Understandably, the significance of step S30 lies in its enhancement of the feature map in the frequency domain through the high-frequency refocusing module. Its core function is to explicitly separate and dynamically amplify the key high-frequency signals representing point-like targets, overcoming the inherent limitations of spatial domain convolution in extracting subtle details. This module first transforms the feature map to the frequency domain, using a mask to separate low-frequency background information from high-frequency target details. Then, through a lightweight channel attention mechanism, it adaptively learns the importance of high-frequency components in each channel and dynamically amplifies them. Finally, it reassembles the enhanced high-frequency components with the retained low-frequency components and inversely transforms them back to the spatial domain. This process directly strengthens the easily suppressed high-frequency components in the network, effectively combating the "semantic collapse" phenomenon, enabling subsequent detection heads to more accurately identify and locate small targets based on the enhanced feature map.

[0087] Step S40: The enhanced multi-scale fused feature map is used as input to predict the location and category of the target by decoupling the detection head.

[0088] Example 2: Please refer to Figure 1 As shown in the embodiment of the present invention, an aerial point target detection method based on high-frequency sensing and dynamic scaling loss further includes the following steps:

[0089] Step S50: The dynamic scaling loss function (DSL) is used to detect and optimize the overall HFA-Mamba model framework during end-to-end training.

[0090] In step S50, the overall HFA-Mamba model framework is specifically as follows:

[0091] HFA-Mamba model framework (see also) Figure 3(As shown) It begins with a high-resolution Mamba backbone (Spatial Domain Backbone), which extracts a set of hierarchical multi-scale features under strict downsampling constraints to preserve fine spatial details. These multi-scale features are then fed into a PANet (Frequency Domain Neck) network, which facilitates the effective and comprehensive fusion of shallow spatial information and deep semantic information. To combat the suppression of weak high-frequency signals indicated by point targets, a High Frequency Refocusing Module (HFRM) is strategically embedded between the PANet neck and the decoupled detection head. The HFRM dynamically amplifies key high-frequency components encoding minute target details in the frequency domain by transforming the fused multi-scale feature map to the frequency domain, thereby achieving targeted feature sharpening before the final detection stage. After enhancement, the enhanced features are fed into a computationally efficient decoupled detection head responsible for bounding box classification and regression.

[0092] It should be noted that the entire HFA-Mamba training process is supervised by dynamic scaling loss (DSL), which aims to adaptively adjust the learning weights according to the target scale, ensuring stable and efficient optimization for targets in extreme size ranges (especially tiny instances).

[0093] In step S50, the dynamic scaling loss function is specifically:

[0094] The core of the dynamic scaling loss function lies in its two weighting functions that depend on the target size. and ,in, and It concerns the actual target area. A i Continuous functions;

[0095] To mitigate unstable gradients from small targets, the contribution of the regression loss is dynamically suppressed through negative exponential decay: Where α R It is a positive hyperparameter that controls the decay rate. Conversely, to prioritize detecting challenging small instances, the classification loss is adaptively focused through an enhancement factor:

[0096] Where β C and γ C It is a positive hyperparameter;

[0097] set up It is the set of positive samples matched by the allocator. Total loss Defined as all allocations

[0098] The above represents the average contribution from the dynamic modulation of classification, bounding box regression, and distribution focusing loss:

[0099] in, With fixed weights, the loss functions for each component are defined as follows:

[0100] Classification loss ( The classification part uses the standard binary cross-entropy (BCE) loss:

[0101] in Ensure numerical stability of bounding box loss ( The regression part utilizes the complete intersection-union (CIoU) loss, which penalizes differences in overlap, center point distance, and aspect ratio. ,in It is the Euclidean distance between the center points. It is the diagonal length of the smallest bounding box. Inconsistency in quantification aspect ratio It is a dynamic trade-off coefficient;

[0102] Distributed focusing loss ( To mitigate boundary ambiguity, a distributed focusing loss is employed, which applies to the boundaries of the four bounding boxes. The regression is reformulated as a discrete probability distribution task:

[0103] For each boundary j, It is a continuous true value, while and It consists of two discrete points that surround it, with a predictive probability. and ;

[0104] Understandably, the dual adaptive weighting strategy ensures that the dynamically scaled loss function (DSL) carefully balances the learning signal across different target scales, providing the best training scheme for point target detection.

[0105] Understandably, the significance of step S50 lies in its use of a dynamically scaled loss function to perform end-to-end optimization of the overall HFA-Mamba model. Its core function is to adaptively adjust the weights of the classification and regression losses based on the true area of ​​the target, thus addressing the training instability issues caused by extreme scale differences and annotation noise in point target detection. This loss function suppresses the contribution of regression loss to small targets through a negative exponential function, mitigating gradient oscillations caused by slight deviations in bounding box annotations. Simultaneously, it increases the focus on the classification loss of small targets through an exponential enhancement function, improving the model's ability to discriminate difficult samples. This scale-aware loss modulation mechanism ensures that the model obtains a balanced learning signal across targets of different sizes during training, thereby significantly improving the robustness and overall accuracy of point target detection.

[0106] Example 3: Please refer to Figure 2 As shown in the embodiment of the present invention, an aerial point target detection system based on high-frequency sensing and dynamic scaling loss includes the following modules:

[0107] The feature extraction module receives aerial images and uses a high-resolution Mamba backbone network to extract multi-scale spatial features containing high-frequency details.

[0108] The feature fusion module fuses the multi-scale spatial features through the PANet neck network to generate a multi-scale fused feature map;

[0109] The feature enhancement module uses a high-frequency refocusing module (HFRM) to enhance the multi-scale fused feature map. The enhancement process includes:

[0110] The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT.

[0111] The target detection output module takes the enhanced multi-scale fused feature map as input and predicts the target's location and category by decoupling the detection head;

[0112] The global detection optimization module uses a dynamically scaled loss function (DSL) to optimize the overall HFA-Mamba model framework during end-to-end training.

[0113] Regarding this invention, the following example is provided to intuitively demonstrate the detection capability of the HFA-Mamba of this invention: (See also...) Figure 6 , Figure 7 and Figure 8It visualizes the detection results compared with other representative methods in several highly challenging scenarios on three benchmark datasets, highlighting key difficulties, including: wide-area surveillance of sparsely distributed, point-like targets occupying only a few pixels. Figure 6 In marine environments, the contrast between the target and the complex reflective water surface background is extremely low. Figure 7 ); and in highly dense crowds, there are situations where there is severe occlusion or complex background textures (such as waves, rocks) that mimic the target. Figure 8 );

[0114] In sparse Heridal scenes ( Figure 6 (first line) and the AFO ocean scene with extremely low contrast ( Figure 7 In all rows, other models produced a large number of missed detections, failing to detect small, blurry targets. In contrast, the HFA-Mamba of this invention successfully identified most of these challenging point targets.

[0115] In environments with complex background textures, such as the lush vegetation of Heridal ( Figure 6 (second line) or TinyPerson's textured water surface and rocky coastline ( Figure 8 (In rows 2, 4, and 5), other models generated a large number of false alarms, incorrectly identifying the background pattern as a person. The HFA-Mamba of this invention effectively suppresses this noise and exhibits superior accuracy.

[0116] In TinyPerson's densely populated scenes ( Figure 8 (in lines 1 and 3), the HFA-Mamba of this invention maintains accurate positioning of occluded individual targets, avoiding the merging or misalignment of bounding boxes that occurs in other methods.

[0117] Figure 9 A qualitative visualization of the HFRM mechanism is provided, demonstrating how the module generates highly localized heatmaps based on the input feature map, thereby providing accurate spatial guidance for subsequent detection. Figure 10 The results of field tests of the invention in real-world scenarios are demonstrated, including two different environments: jungle and wilderness. At a drone flight altitude of approximately 30-80 meters and a variable viewing angle of 45°-90°, HFA-Mamba achieved high-precision and robust detection of pedestrian samples with camouflage, different viewing angles, different altitudes, and occlusion.

[0118] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting aerial point targets based on high-frequency sensing and dynamic scaling loss, characterized in that: Includes the following steps: The system receives aerial images and uses a high-resolution Mamba backbone network to extract multi-scale spatial features containing high-frequency details. The multi-scale spatial features are fused using the PANet neck network to generate a multi-scale fused feature map; The multi-scale fused feature map is enhanced using a high-frequency refocusing module (HFRM). The enhancement process includes: The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT. The enhanced multi-scale fused feature map is used as input to predict the location and category of the target by decoupling the detection head; A dynamic scaling loss function is used to optimize the overall HFA-Mamba model framework during end-to-end training. The loss function dynamically adjusts the weights of classification loss and regression loss based on the area of ​​the target's true region. The dynamic scaling loss function includes a regression loss weight function. A classification loss weight function ; The regression loss weight function It is a negative exponential decay function with respect to the area A of the target real region, and its calculation expression is: Where A is the area of ​​the target region. ε is a positive hyperparameter that controls the decay rate, ε is a small positive constant to prevent division by zero, i is the number of all positive samples in one training iteration, and Ai is the area calculated for each i-th target instance. The classification loss weight function It is an exponential enhancement function with respect to the area A of the target true region, and its calculation expression is: Where A is the area of ​​the target region, β c and γ c is a positive hyperparameter, i is the number assigned to all positive samples in one training iteration, and Ai is the area calculated for each i-th target instance.

2. The aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to claim 1, characterized in that: The high-resolution Mamba backbone network comprises an initial patch embedding layer and three layered phases; The process of extracting multi-scale spatial features containing high-frequency details is as follows: The input image I is first downsampled through a convolutional layer to obtain the feature map X0; the three hierarchical stages output multi-scale feature maps with downsampling rates of 8x, 16x, and 32x, respectively. The output of the i-th stage is: Each MambaStage i It consists of a downsampling layer and Ni sequential HG-MFTBlocks, used to model global context dependencies while preserving high-frequency details, where, :No. Feature maps output from each hierarchical stage At that time, the corresponding multi-scale output feature map, The multi-scale feature maps output by the backbone network correspond to the input image at different scales. downsampling scale No. The overall mapping function for the hierarchical stage is used to summarize all combinations of operators within that stage. No. Feature maps output from each hierarchical stage :No. The specific network structure of the hierarchical stage, Ni: the number of HG-MFTBlock stacks in the i-th hierarchical stage, used to characterize the depth of the stage.

3. The aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to claim 1, characterized in that: The PANet neck network fuses multi-scale features through a dual-path mechanism, specifically including: The top-down path starts from the deepest feature map C5 and gradually propagates high-level semantic information to shallower layers through upsampling and lateral connections, generating a semantically enhanced feature pyramid. The bottom-up path starts from the shallowest feature map P3, and injects low-level spatial details into deeper layers through downsampling and connections, ultimately outputting a multi-scale fused feature map. The calculation process is as follows: in N represents intermediate features. :No. Layer Convolution, used for channel alignment to achieve feature fusion. It is a scale-level index in a multi-scale feature pyramid, used to distinguish feature layers of different resolutions. This indicates that the top-to-bottom path of PANet is at the [number]th [level]. Feature maps output at each scale express Indexed collection, Downsampling operators are used to extract intermediate features at higher resolution. Downsample to a spatial size consistent with the current scale features. It is the intermediate fused feature map of the previous layer scale in the PANet bottom-to-top path. : A convolutional module used for feature fusion and refinement, representing the fusion transformation. : A convolutional module used for feature fusion and refinement, representing the output refinement process, index. : indicates a multi-scale hierarchical index. This indicates the generation of layers through a gradual fusion process from shallow to deep within the bottom-to-top path. .

4. The aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to claim 1, characterized in that: The specific process by which the high-frequency refocusing module HFRM enhances the multi-scale fused feature map is as follows: The multi-scale fused feature map is transformed to the frequency domain through fast Fourier transform to obtain a complex spectral tensor, which is then centered. Using a binary low-pass mask, the centered spectrum is separated into low-frequency and high-frequency components through Hadamard product; Design a channel attention function to learn the channel-specific amplification factor based on the amplitude spectrum of the high-frequency components. Let W represent the existence of a C-dimensional real vector, where each dimension corresponds to an intensity coefficient that amplifies the high-frequency components of the i-th feature channel, and W represents the channel-specific amplification coefficient. Let C represent the set of real numbers, and let C represent the number of channels in the feature map. The amplification factor is used to dynamically enhance the high-frequency components to obtain the enhanced high-frequency components. The amplification factor is obtained by adaptive learning of the amplitude spectrum of the high-frequency components by the channel attention function, and is applied to each spatial position of the high-frequency components through broadcasting to achieve dynamic enhancement of high-frequency information of different channels. The enhanced high-frequency components are recombined with the original low-frequency components to obtain the enhanced spectrum; By performing inverse centering and inverse fast Fourier transform, the enhanced spectrum is converted back to the spatial domain, resulting in a multi-scale fusion feature map with enhanced details.

5. The aerial point target detection method based on high-frequency sensing and dynamic scaling loss according to claim 4, characterized in that: The channel attention function involves global average pooling of the amplitude spectrum of high-frequency components to extract channel descriptors, which are then processed by a multilayer perceptron network consisting of one-dimensional convolutional layers to generate channel-specific amplification coefficients. Let W represent the existence of a C-dimensional real vector, where each dimension corresponds to an intensity coefficient that amplifies the high-frequency components of the i-th feature channel, and W represents the channel-specific amplification coefficient. Let C represent the set of real numbers, and let C represent the number of channels in the feature map.

6. An aerial point target detection system based on high-frequency sensing and dynamic scaling loss, characterized in that: Includes the following modules: The feature extraction module receives aerial images and uses a high-resolution Mamba backbone network to extract multi-scale spatial features containing high-frequency details. The feature fusion module fuses the multi-scale spatial features through the PANet neck network to generate a multi-scale fused feature map; The feature enhancement module uses a high-frequency refocusing module (HFRM) to enhance the multi-scale fused feature map. The enhancement process includes: The multi-scale fused feature map is transformed to the frequency domain using FFT; low-frequency and high-frequency components are separated using a binary mask; high-frequency components are dynamically amplified using a channel attention mechanism; and the enhanced reconstructed spectrum is returned to the spatial domain using IFFT. The target detection output module takes the enhanced multi-scale fused feature map as input and predicts the target's location and category by decoupling the detection head; The global detection optimization module uses a dynamically scaled loss function to optimize the overall HFA-Mamba model framework during end-to-end training. The loss function dynamically adjusts the weights of classification loss and regression loss based on the area of ​​the target's true region. The dynamic scaling loss function includes a regression loss weight function. A classification loss weight function ; The regression loss weight function It is a negative exponential decay function with respect to the area A of the target's true region, and its calculation expression is: Where A is the area of ​​the target region. ε is a positive hyperparameter that controls the decay rate, ε is a small positive constant to prevent division by zero, i is the number of all positive samples in one training iteration, and Ai is the area calculated for each i-th target instance. The classification loss weight function It is an exponential enhancement function with respect to the area A of the target true region, and its calculation expression is: Where A is the area of ​​the target region, β c and γ c is a positive hyperparameter, i is the number assigned to all positive samples in one training iteration, and Ai is the area calculated for each i-th target instance.

Images