Adaptive heterogeneous decoupled detection head based on multi-dimensional scene representation

Abstract

The present application belongs to the technical field of computer vision and image processing, and particularly relates to a self-adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation. A panoramic statistical gating network outputs classification gating coefficients and regression gating coefficients in real time. A classification enhancement branch obtains a basic classification feature map through convolution, is synchronously input into a DHF-CM module for processing, and completes dynamic weighted fusion based on the classification gating coefficients from the PSGN. The fused attention map and the original line perform residual feature injection through a matrix element-by-element multiplication operator under a broadcast mechanism, and finally output a classification prediction probability distribution through an independent convolution layer and a prediction convolution layer. A regression enhancement branch obtains a basic regression feature map through convolution, and after the same operation as the classification enhancement branch, outputs accurate bounding box regression coordinates. The present application breaks through the technical barriers of traditional static and fixed attention mechanisms, and realizes dynamic weighted heterogeneous modulation in each branch.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and particularly relates to an adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation. Background Technology

[0002] To alleviate the representational conflict between the classification and localization subtasks, most commonly used single-stage detection head schemes employ a decoupled head topology. That is, after basic extraction, the input feature map is split into independent classification and regression branches through different convolutional paths. To further improve the representational ability of decoupled heads for small targets, the most similar existing technical solutions mainly present the following two structures: one is the standard unenhanced decoupled detection head, which performs dimensionality reduction and channel mapping only through ordinary convolutional layers after splitting. Because the features of small targets (such as distant vehicles and densely packed ships) in remote sensing scenarios are easily attenuated during deep network propagation, detection heads lacking specific feature enhancement cannot effectively extract the geometric details of small target boundaries. This results in a lack of high-level semantic capture capability for the classification path and extremely poor robustness of the regression path for localizing small boundaries. In the presence of complex background features (such as farmland, buildings, and road noise), it easily leads to large-area missed detections of small targets and severe drift of the localization boxes. Another approach is a decoupled detection head using a static, symmetrically stacked attention mechanism. Classification and regression tasks inherently have different feature preferences (classification prefers channel semantics, while regression prefers spatial coordinates). Staticly stacking the same attention mechanism or statically binding branch structures, when faced with dynamic scene fluctuations (e.g., extremely crowded targets causing overlapping spatial boundaries, or image degradation leading to blurred channel semantics), the fixed topology cannot dynamically find the optimal balance between "channel dimension selection" and "spatial location-sensitive representation," exacerbating the representational conflict between classification and regression tasks in scenarios with small and dense targets. Summary of the Invention

[0003] In view of this, the present invention aims to provide an adaptive heterogeneous decoupled detection head based on multi-dimensional scene representation to solve the technical problems of existing detection heads in the face of degradation in multi-source remote sensing dynamic imaging (such as blurring, haze) and drastic fluctuations in target density, such as insufficient elasticity in representing dense and weak targets, limited generalization performance, and increased conflict between classification and regression. The present invention constructs a panoramic scene perception and task-specific heterogeneous dynamic modulation system at the final output layer of the detection network. This system first constructs a panoramic scene descriptor using multi-dimensional global statistics (including background mean, strong activation density, and local gradient standard deviation reflecting edge texture loss); then, through a lightweight gating network, it predicts in real time the dynamic demand weights for high-frequency channel response enhancement and local multi-directional position enhancement in two decoupled branches (classification and regression); finally, it breaks through the technical barrier of static and fixed traditional attention mechanisms, realizing dynamic weighted heterogeneous modulation "tailored to the scene" within each branch.

[0004] To achieve the above objectives, the technical solution created by this invention is implemented as follows: An adaptive heterogeneous decoupling detection head based on multidimensional scene representation includes a classification enhancement branch, a panoramic statistical gating network, and a regression enhancement branch configured in parallel, wherein: The panoramic statistical gating network extracts local features from the input feature map based on a single convolutional layer to obtain a first feature map. The first feature map is then subjected to different global statistical operations through parallel global average pooling, global max pooling, and global standard deviation pooling operators to extract three different global statistical features. These three different global statistical features are then cascaded and fed into the classification gating predictor and regression gating predictor for processing, and the corresponding outputs are classification gating coefficients and regression gating coefficients. The classification enhancement branch performs convolution operations on the input feature map to obtain the basic classification feature map. After the basic classification feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using classification gating coefficients, and the classification prediction probability distribution is output through independent decoupled convolutional layers. The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After the basic regression feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using regression gating coefficients, and the bounding box regression coordinates are output through an independent decoupled convolutional layer.

[0005] Furthermore, the specific operational process for classifying and enhancing branches is as follows: The classification enhancement branch performs convolution operations on the input feature map to obtain a basic classification feature map. After processing by the DHF-CM module, the basic classification feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the classification gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain the first fused attention map. The first fused attention map and the basic classification feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the classification prediction probability distribution is output through an independent decoupled convolutional layer.

[0006] Furthermore, the specific operational process of the regression enhancement branch is as follows: The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After processing by the DHF-CM module, the basic regression feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the regression gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain a second fused attention map. The second fused attention map and the basic regression feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the bounding box regression coordinates are output through an independent decoupled convolutional layer.

[0007] Furthermore, the specific operation process of the DHF-CM module is as follows: The input feature map is preprocessed with 3×3 depthwise convolution and 1×1 pointwise convolution to obtain shared intermediate features; The shared intermediate features are processed by 3×3 convolution to obtain the second feature map. The second feature map is then processed by the LMDP module to obtain the orientation and position enhancement map. The shared intermediate features are processed by 3×3 convolution to obtain the third feature map. The third feature map is then processed by the HFCR module, and the processing result is multiplied with the third feature map to obtain the high-frequency channel enhancement map.

[0008] Furthermore, the processing procedure of the LMDP module is as follows: The input feature map is processed by horizontal local pooling, vertical local pooling, first diagonal local pooling, and second diagonal local pooling to obtain features with four different spatial orientations. These features are then projected and concatenated to obtain a fused feature map. The fused feature map is then processed by 1×1 convolution and Split to obtain first branch features and second branch features. The first branch features are then processed by 1×1 convolution and Sigmoid activation to obtain the first output feature. The second branch features are then processed by 1×1 convolution and Sigmoid activation to obtain the second output feature. The first output feature, the second output feature, and the input feature map are multiplied to obtain the orientation position enhancement map.

[0009] Furthermore, the HFCR module's processing procedure is as follows: The input feature map is processed sequentially by high-frequency average pooling, MLP and Sigmoid to obtain high-frequency channel weights; The high-frequency channel weights are multiplied by the input feature map to obtain the high-frequency channel enhancement map.

[0010] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The adaptive heterogeneous decoupling detection head based on multidimensional scene representation created in this invention has panoramic multidimensional scene degradation and target density quantification perception capabilities, overcoming the single dimension limitation of traditional pooling. Existing technologies rely only on global average pooling (GAP) or global maximum pooling (GMP), which are first-order or boundary statistics that can only reflect the global brightness and strongest local response of the image. When faced with high-frequency texture erosion (smooth degradation) of the image caused by atmospheric turbulence, fog, or motion blur, they cannot achieve quantitative feature capture. This invention introduces global standard deviation pooling (GSDP) as a novel environmental degradation metric operator in the scene perception path, which can accurately quantify the discrete fluctuations of local gradients in spatial images. In low-contrast degradation scenes such as fog or blur, local edge blurring will cause the global standard deviation to drop sharply. Combined with the parameterless GAP and GMP operators, the panoramic scene description vector constructed in this invention It can capture "environmental degradation degree" and "ground object density" in a comprehensive and multi-dimensional way, providing accurate scene-level pre-statistical priors for subsequent adaptive modulation.

[0011] (2) The adaptive heterogeneous decoupled detection head based on multi-dimensional scene representation created in this invention realizes dynamic gating adaptive control based on the scene, breaking the "one-size-fits-all" constraint of traditional hard-coded fixed topology. Existing technical solutions adopt a static hard-coded topology, and the weight fusion ratio of the classification and regression branch enhancement operators remains unchanged regardless of whether the remote sensing image is clear, blurry, or ultra-dense. The performance drops sharply when switching between complex scenes. This invention designs a lightweight panoramic statistical gating network (PSGN), which uses an independent heterogeneous multilayer perceptron (MLP) to implicitly learn the mapping relationship between the current panoramic scene and the feature preferences of the two sub-tasks, and uses the Softmax operator to output the classification gating coefficients in real time. With regression gating coefficient This mechanism endows the detector head with excellent "representational flexibility," enabling the network to dynamically find the optimal weighted game balance between Local Multidirectional Position Enhancement (LMDP) and High Frequency Channel Response Enhancement (HFCR) based on the real-time imaging quality of the input image. This significantly improves the model generalization ability and anti-interference robustness of the YOLO11 prediction layer under extreme weather or non-uniform imaging fluctuations.

[0012] (3) The adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation described in this invention effectively alleviates the heterogeneous representation conflict between classification and regression tasks, and significantly suppresses the missed detection and localization box drift of dense small targets. Existing technologies symmetrically and indiscriminately apply the same attention mechanism to both ends of the decoupling head, failing to effectively solve the heterogeneous demand conflict problem of "classification prefers channel semantics, regression prefers spatial coordinates" in dynamic environments, resulting in the representation confusion of dense small targets under degradation imaging fluctuations. While maintaining the asymmetric decoupling topology, this invention forces the classification enhancement branch and the regression enhancement branch to perform asymmetric game-theoretic allocation constraints through gating coefficients (satisfying... and When image quality is good, increase the directional position enhancement map. The fusion weights are adjusted to ensure the accuracy of the regression path anchoring to the boundary of small targets; when environmental degradation causes blurring, the high-frequency channel enhancement map is automatically increased. The fusion weights are used to enhance the contextual semantic compensation capability of the classification path. This dual-path dynamic heterogeneous reconstruction mechanism maximizes the dynamic alignment of multiple tasks in the high-activation region of space, and significantly reduces the false detection and false negative rates of small targets in complex remote sensing backgrounds.

[0013] (4) The adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation described in this invention has extremely high energy efficiency and edge-friendly performance, achieving seamless and lossless lightweight plug-and-play deployment. Existing technologies, in pursuit of dynamic performance, typically introduce ultra-large receptive field convolutions or dense spatial matrix multiplications (such as the Transformer Head based on full-image self-attention mechanisms) into the detection head. This results in a massive number of parameters and computational overhead, making it difficult to run on embedded edge devices with extremely limited computing resources, such as UAVs or satellites. The panoramic statistical gating network (PSGN) of this invention runs entirely on low-dimensional global channel-level vectors (dimensional only...). At the [level] level. Simultaneously, a channel reduction rate hyperparameter was specifically introduced into the dual-path MLP predictor. This creates a bottleneck-level low-dimensional mapping. Without introducing any spatially dimensional dense matrix multiplications, this topology endows the detector head with powerful dynamic modulation capabilities with negligible (almost negligible) FLOPs of computation and parameter overhead, making it highly suitable for the real-time, lightweight online processing requirements of remote sensing airborne equipment. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1A schematic diagram of the structure of the adaptive heterogeneous decoupling detection head based on multidimensional scene representation as described in the embodiments of the present invention; Figure 2 A schematic diagram of the structure of the DHF-CM module described in the embodiment of the present invention; Figure 3 This is a schematic diagram of the overall embedding position of the adaptive heterogeneous decoupling detection head in the YOLO11 network according to an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0017] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0018] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] like Figure 1 As shown, this invention proposes an adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation, including a classification enhancement branch, a panoramic statistical gating network, and a regression enhancement branch set in parallel, wherein: The panoramic statistical gating network extracts local features from the input feature map based on a single convolutional layer to obtain a first feature map. The first feature map is then subjected to different global statistical operations through parallel global average pooling, global max pooling, and global standard deviation pooling operators to extract three different global statistical features. These three different global statistical features are then cascaded and fed into the classification gating predictor and regression gating predictor for processing, and the corresponding outputs are classification gating coefficients and regression gating coefficients. The classification enhancement branch performs convolution operations on the input feature map to obtain the basic classification feature map. After the basic classification feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using classification gating coefficients, and the classification prediction probability distribution is output through independent decoupled convolutional layers. The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After the basic regression feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using regression gating coefficients, and the bounding box regression coordinates are output through an independent decoupled convolutional layer.

[0021] It should be noted that this invention establishes a dynamic multidimensional representation mechanism for environmental degradation and scene density: by breaking away from the inherent pattern of traditional attention mechanisms that only extract average or maximum values, it collaboratively introduces Global Standard Deviation Pooling (GSDP) as a novel environmental degradation metric operator. Through changes in standard deviation, it perceives the blur, fog, and background noise complexity of the current image in real time and quantitatively, constructing a highly elastic panoramic scene descriptor. This invention achieves dynamic heterogeneous gating modulation based on the scene: it designs a lightweight panoramic statistical gating network (PSGN), which, based on extracted multidimensional scene statistics, predicts in real time the dynamic gating coefficients of the two decoupled branches of classification and regression for feature enhancement requirements of "channel dimension" and "spatial coordinate dimension," breaking the technical constraint of hard-coded fixed operators inside the detection head. This invention maximizes the suppression of missed detections, false detections, and localization drift while maintaining a lightweight design. It forces the classification enhancement branch and the regression enhancement branch to dynamically reconstruct the fusion ratio of their internal attention mechanisms when facing fluctuations in different imaging environments. In clear scenes, it strengthens spatial coordinate guidance, and in degraded scenes, it enhances semantic channel compensation. Without introducing dense spatial matrix calculations, it significantly improves the detection accuracy and robustness of single-stage detection frameworks such as YOLO11 in multi-source remote sensing small target tasks.

[0022] Input feature map Parallel three-way splitting is immediately activated upon entering the detection head: (1) Mid-path (Statistical Panoramic Gated Network PSGN): The lossless input passes through a convolutional layer and then flows in parallel to the Global Average Pooling (GAP), Global Max Pooling (GMP), and Global Standard Deviation Pooling (GSDP) operators to extract scene-level statistics. After being concatenated through channels, these statistics are fed into the classification gated predictor and the regression gated predictor, respectively. Each predictor consists of two cascaded linear fully connected layers, and ultimately outputs the classification gating coefficients in real time. With regression gating coefficient .

[0023] (2) Upper path (classification enhancement branch): The basic classification feature map is obtained through convolution and is synchronously input into the DHF-CM module to generate the orientation and position enhancement map in parallel. With high-frequency channel enhancement map The two sets of augmented maps are based on classification gating coefficients from PSGN. Dynamic weighted fusion is performed within the module. The fused attention map and the original circuit are injected with residual features by the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the classification prediction probability distribution is output through independent convolutional layers and prediction convolutional layers.

[0024] (3) Lower path (regression enhancement branch): The basic regression feature map is obtained through convolution, and is also input into the DHF-CM module for extraction. and Based on the regression gating coefficients from PSGN Asymmetric weighted fusion is performed and injected into the regression path, and finally, accurate bounding box regression coordinates are output through independent decoupled convolutional layers.

[0025] In some embodiments, the specific operation process of the classification enhancement branch is as follows: The classification enhancement branch performs convolution operations on the input feature map to obtain a basic classification feature map. After processing by the DHF-CM module, the basic classification feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the classification gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain the first fused attention map. The first fused attention map and the basic classification feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the classification prediction probability distribution is output through an independent decoupled convolutional layer.

[0026] In some embodiments, the specific operation process of the regression enhancement branch is as follows: The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After processing by the DHF-CM module, the basic regression feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the regression gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain a second fused attention map. The second fused attention map and the basic regression feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the bounding box regression coordinates are output through an independent decoupled convolutional layer.

[0027] In some embodiments, the specific operation process of the DHF-CM module is as follows: The input feature map is preprocessed with 3×3 depthwise convolution and 1×1 pointwise convolution to obtain shared intermediate features; The shared intermediate features are processed by 3×3 convolution to obtain the second feature map. The second feature map is then processed by the LMDP module to obtain the orientation and position enhancement map. The shared intermediate features are processed by 3×3 convolution to obtain the third feature map. The third feature map is then processed by the HFCR module, and the processing result is multiplied with the third feature map to obtain the high-frequency channel enhancement map.

[0028] like Figure 2 As shown, the Directional-High Frequency Co-modulation (DHF-CM) module consists of two parallel candidate enhancement paths: the Local Multi-Directional Position Aggregation (LMDP) module and the High Frequency Channel Response Modulation (HFCR) module. Any input feature F first undergoes lightweight preprocessing via 3×3 depthwise convolution (DWConv) and 1×1 pointwise convolution (PWConv) to obtain shared intermediate features Φ. The upper LMDP module then performs horizontal local pooling... Vertical local pooling Local pooling in the first diagonal direction Local pooling in the second diagonal direction The directional position response of small targets is extracted. The local pooling features in the first and second diagonal directions are mapped through horizontal and vertical projections, respectively, to horizontal projection features that retain height position information and vertical projection features that retain width position information. Then, after concat, 1×1 convolution, split, orientation mapping convolution, and sigmoid activation, an orientation position enhancement map is generated. The HFCR module below extracts edges, texture residuals, and local high-frequency details through high-frequency average pooling (HFAP). It then uses MLP and sigmoid to obtain high-frequency channel weights, which are multiplied with the branch features to generate a high-frequency channel enhancement map. The above. and As two candidate execution units of the PSGN dynamic gating mechanism, they serve the task-related adaptive fusion of the classification branch and the regression branch, respectively.

[0029] In some embodiments, the processing procedure of the LMDP module is as follows: The input feature map is processed by horizontal local pooling, vertical local pooling, first diagonal local pooling, and second diagonal local pooling to obtain features with four different spatial orientations. The features with four different spatial orientations are concatenated to obtain a fused feature map. The fused feature map is processed by 1×1 convolution and Split to obtain the first branch features and the second branch features. The first branch features are processed by 1×1 convolution and Sigmoid activation to obtain the first output feature. The second branch features are processed by 1×1 convolution and Sigmoid activation to obtain the second output feature. The first output feature, the second output feature, and the input feature map are multiplied to obtain the orientation position enhancement map.

[0030] In some embodiments, the processing procedure of the HFCR module is as follows: The input feature map is processed sequentially by high-frequency average pooling, MLP and Sigmoid to obtain high-frequency channel weights; The high-frequency channel weights are multiplied by the input feature map to obtain the high-frequency channel enhancement map.

[0031] The basic testing process of this invention is as follows: Step S1: Preparation for multi-path splitting of input features and task decoupling The small target detection feature map output by the feature fusion network (such as the YOLO11 feature fusion network with high-resolution feature pyramids) The input is fed into the adaptive heterogeneous decoupling detection head of this invention, and then split into two parallel processing paths: Path 1 (Scene-Aware Path): The original input feature map The data is directly input into the panoramic statistical gating network (PSGN) to perform quantitative statistical feature extraction of scene degradation and target density; Path 2 (Task Decoupling Path): Input feature map Through two independent decoupled paths, they are mapped to basic classification feature maps used to identify target attributes. and the basic regression feature map used for precise target range localization .

[0032] Step S2: Collaborative Extraction of Multi-Dimensional Panoramic Scene Representation In the scene perception pathway, this invention collaboratively executes three global pooling operations with different physical and geometric scales. Global average pooling (GAP) captures the background brightness and mean information of the macroscopic image; global max pooling (GMP) extracts the peak response of local strong activation points to reflect the density of the target distribution; and global standard deviation pooling (GSDP) measures the spatial discreteness of the feature map response to quantify the degree of high-frequency texture erosion and environmental degradation caused by atmospheric turbulence or fog.

[0033] The calculated global average pooling vector Global max pooling vector and global standard deviation pooling vector By cascading and stitching along the channel dimension, a panoramic scene description vector that can comprehensively reflect the current remote sensing imaging status is dynamically reconstructed. .

[0034] Step S3: Construction of Adaptive Heterogeneous Gated Prediction Mechanism Panoramic scene description vector Within the panoramic statistical gating network, the data is further split and fed into two independent lightweight multilayer perceptrons (MLPs). These two MLPs establish implicit mappings with the representation preferences of the classification and regression tasks, respectively, and finally perform normalization mapping through the Softmax activation function to predict the heterogeneous modulation coefficients applicable to the classification and regression branches in real time and dynamically.

[0035] The heterogeneous modulation coefficients include the required weights of the classification enhancement branch for the high-frequency channel response and the local directional position response, as well as the required weights of the regression enhancement branch for the local directional position response and the high-frequency channel response. The two sets of coefficients are subject to a dynamic game constraint in the topology, which sums to 1.

[0036] Step S4: Task-driven dynamic adaptive heterogeneous feature modulation The dynamic gating coefficients predicted in step S3 are injected in real time into the classification and regression base feature maps obtained from step S1 for dynamic weighted heterogeneous feature modulation. Within the classification enhancement branch: the high-frequency channel enhancement map generated by the high-frequency channel response modulation module (HFCR) is the main component, and the directional position enhancement map generated by the local multi-directional position aggregation module (LMDP) is the auxiliary component. Adaptive weighted fusion is performed based on the classification gating coefficient to generate a classification enhancement feature map that is highly robust to background degradation noise. Within the regression branch: the orientation and position enhancement map generated by the Local Multi-Directional Position Aggregation Module (LMDP) is the primary component, while the high-frequency channel enhancement map generated by the High-Frequency Channel Response Modulation Module (HFCR) is the secondary component. Dynamic weighted fusion is performed based on the regression gating coefficients to generate a regression enhancement feature map that is sensitive to the boundary and orientation and position information of small targets.

[0037] Step S5: Multi-task decoupling prediction output The classification-enhanced feature map generated after adaptive weighted heterogeneous feature modulation With regression-enhanced feature maps These are inputs to the final task output mapping layer, respectively. The two paths are respectively processed through independent... Convolutional layers and The task prediction mapping layer performs channel dimensionality reduction and category space projection on target features, thereby outputting the classification prediction probability distribution of multi-source remote sensing micro-targets in parallel. With precise bounding box regression coordinates, i.e. bounding box regression coordinates It enables intelligent identification of dense, weak targets in complex and fluctuating scenarios.

[0038] Example 1 This embodiment uses the current mainstream single-stage target detection network YOLO11 as the baseline network framework, and adjusts its default P3–P5 detection pyramid to a high-resolution P2–P4 detection pyramid for small targets. The adaptive heterogeneous modulation detection head (DAH-Head) proposed in this invention is seamlessly replaced and embedded as its final prediction component.

[0039] like Figure 3 As shown, the Adaptive Heterogeneous Modulation Detection Head (DAH-Head) proposed in this invention serves as the final prediction component, seamlessly replacing and embedding it at the output of a single-stage object detection network (taking YOLO11 as an example). The input image (dimension 1) is... First, the feature maps enter the backbone network, where multi-scale feature extraction is performed through consecutive CBS operators, C3K2 modules, and SPPF+C2PSA cascaded modules. The extracted multi-level feature maps are then fed into the feature fusion network (Neck), where multi-scale fusion is performed through a bidirectional feature fusion path (including Concat, Upsample, C3K2, and CBS operators) from top to bottom and bottom to top. Finally, the Neck layer outputs the feature maps. ( (1 / 4 scale) ( (1 / 8 scale) and ( Three parallel feature tensors (1 / 16 scale) are generated, and the corresponding DAH-Head is entered respectively. ), DAH-Head ) and DAH-Head( Adaptive heterogeneous prediction output is performed in ).

[0040] This invention uses the original object detection loss function of YOLO11 during the training phase without introducing additional auxiliary losses. The total loss consists of classification loss, bounding box geometric regression loss, and bounding box discrete distribution regression loss, expressed as follows: ; in, The binary cross-entropy classification loss is used to constrain the category prediction results; The complete intersection-union loss is used to constrain the overlap, center distance, and aspect ratio between the predicted bounding box and the ground truth bounding box. The distributed focus loss is used to optimize the discrete distribution prediction of the distances in the four directions of the bounding box; , and These are the weighting coefficients for the corresponding loss terms.

[0041] (1) Task flow of input features and decoupling preprocessing In this embodiment, the feature map of the input adaptive heterogeneous modulation detection head is used. The feature fusion network derived from YOLO11 (P2, P3 or P4 scale output of the Neck layer).

[0042] Let the dimension of the current input tensor be . ,in, Batch size. Number of feature channels (e.g.) ), and These represent the height and width of the spatial feature map, respectively.

[0043] like Figure 2 As shown, feature map Parallel three-way splitting is immediately performed after entering the adaptive heterogeneous modulation detection head: 1) Classification benchmark feature splitting: using a standard Two-dimensional convolutional layers (containing convolution, batch normalization (BatchNorm), and SiLU activation functions, collectively known as CBS operators) preserve or map the number of channels to... The basic classification feature map is obtained. .

[0044] 2) Regression baseline feature splitting: In parallel, through another independent The CBS operator maps the number of channels to Generate basic regression feature maps .

[0045] 3) Scene-aware statistical routing: This involves routing the original, lossless data... The data is synchronously input into the panoramic statistical gating network (PSGN).

[0046] (2) Explicit matrix mapping of panoramic statistical gating network (PSGN) The goal of Panoramic Statistical Gated Network (PSGN) is to quantify the degree of image degradation and target density of the current input remote sensing image using parameter-free physical statistics.

[0047] (2.1) Details of three-way collaborative panoramic pooling For the input F, the network moves along the spatial dimension Three different mathematical aggregation operators are executed in parallel to preserve the most complete panoramic representation of the scene at the channel level: 1) Global Average Pooling (GAP): Extracts the macroscopic background mean of the feature map. For any channel The calculation formula is as follows: ; in, This represents the batch index, used to indicate the index of the current input batch. Image 1 or 2 One sample; Indicates the feature channel index; This represents the spatial index of the feature map in the height direction; This represents the spatial index of the feature map in the width direction.

[0048] 2) Global Max Pooling (GMP): Captures the strongest local geometric activations in the feature map to reflect the clustering density of weak targets. : ; 3) Global Standard Deviation Pooling (GSDP): Used to quantify the loss of local gradient edges in an image (fog, motion blur, or large areas of shadow can cause a significant reduction in the standard deviation of spatial dispersion).

[0049] To prevent gradient explosion or division by zero errors when taking the square root during backpropagation due to variance approaching zero, this invention explicitly introduces a hyperparameter adjustment operator (numerical stability factor) inside the square root operator. Its standard deviation vector The precise formula is expressed as follows: ; get Then, they are fused together using the channel concatenation (Concat) operator. The fused panoramic scene description vector... The dimensions have increased to 3 times the original size: ; (2.2) Dual Fully Connected Heterogeneous Gated Predictor Topology This invention does not use a single large model for prediction, but instead establishes a dual-path parallel heterogeneous multilayer perceptron (MLP) specifically for classification and regression.

[0050] To achieve extremely lightweight edge deployment, each MLP employs a bottleneck structure, introducing a bottleneck dimension ratio parameter. (Selected in this embodiment) ).

[0051] Predictors based on classification paths For example, its hierarchical cascading relationship and dimensional changes are as follows: 1) Input layer: dimension is (like ); 2) Intermediate dimensionality reduction fully connected layer (Linear Layer 1): This layer converts the input features from... Mapped to (when (The dimension becomes 64 when the time interval is 4), followed by the application of a SiLU nonlinear activation function. This step aims to perform nonlinear high-order fusion of cross-pooling statistics in a low-dimensional space; 3) Output Linear Layer 2: Maps the channel dimension from the low-dimensional bottleneck space to 2D gated logarithmic probability values ​​(Logits). These 2D Logits represent the immediate demand of the current task branch for High Frequency Channel Response Enhancement (HFCR) and Local Multidirectional Position Enhancement (LMDP).

[0052] 4) Normalization Layer: The output 2D channels are normalized using the Softmax activation function to obtain the final classification gating coefficients. Let the classification gating coefficients be... Its satisfaction The calculation process can be expressed as follows: ; ; in, Enhancement diagram of high-frequency channel response in classification enhancement branches Dynamic fusion weights, Enhanced directional location maps in branch paths for classification Dynamic fusion weights; GAP ( ) represents global average pooling, GMP ( ) represents global max pooling, GSDP ( ) indicates global standard deviation pooling.

[0053] Similarly, the control coefficients of the regression branch are dynamically predicted using a completely independent regression gating predictor. ,in, This indicates the enhancement effect of the regression enhancement branch on the high-frequency channel. The fusion weight, This indicates the direction and position enhancement map of the regression enhancement branch. The fusion weights, and the two satisfy the following: .

[0054] (2.3) Task-driven direction - high-frequency collaborative heterogeneous feature modulation In the decoupled head path, the basic classification features and basic regression features, during forward propagation, no longer employ the simple parallel structure of traditional channel attention and coordinate attention. Instead, they are simultaneously input into the Direction-High Frequency Co-modulation Module (DHF-CM). The DHF-CM consists of two candidate enhancement execution units: the Local Multi-Direction Position Aggregation Module (LMDP) and the High Frequency Channel Response Modulation Module (HFCR), which output the direction and position enhancement map M, respectively. dir With high-frequency channel enhancement map M hf .

[0055] First, for the input feature F, DHF-CM performs lightweight preprocessing using 3×3 depthwise convolution (DWConv) and 1×1 pointwise convolution (PWConv) to obtain shared intermediate features Φ. This preprocessing preserves the local texture and boundary response of small targets while avoiding the introduction of large-scale spatial matrix calculations in the detection head.

[0056] Secondly, in the LMDP module, the shared features are fed into horizontal local pooling after convolutional integrators. Vertical local pooling Local pooling in the first diagonal direction Local pooling in the second diagonal direction The statistical results from the four directions are concatenated using Concat and compressed using 1×1 convolution, and then decomposed into response components in different directions using a Split operation. and ,in, This represents the characteristic response components that are correlated in the horizontal direction. This represents the feature response components related to the vertical direction. A direction-sensitive weighted response map is obtained by performing a 1×1 convolution followed by sigmoid activation. and Finally, a local multi-directional position enhancement map is generated through element-wise multiplication. .

[0057] Furthermore, in the HFCR module, the shared features are fed into the High Frequency Average Pooling (HFAP) operator after convolutional integration. HFAP is used to extract small object edges, texture residuals, and local high-frequency detail responses, and is expressed as: ;in, This is the high-frequency channel description vector extracted by the HFAP module.

[0058] Subsequently, high-frequency channel weights are generated through lightweight MLP and Sigmoid activation. , The Sigmoid activation is performed, and element-wise modulation is applied with the branch features to obtain the high-frequency channel enhancement map M. hf .

[0059] Finally, the task-related gating coefficients output by PSGN are injected into the two candidate augmentation graphs of DHF-CM. For the classification branch, the following is adopted: Dynamic fusion is performed; for the regression branch, the following is adopted: Dynamic fusion is performed. Because... and Generated by two sets of independent predictors, the classification augmentation branch and the regression augmentation branch can learn different augmentation preferences based on the same input scene state.

[0060] In the feature injection stage, this invention employs residual attention modulation to avoid directly suppressing the response of weak targets. (Classification Enhancement Features) With regression enhancement features They are represented as follows: ; ; Here, ⊙ represents element-wise multiplication. and The modulation intensity coefficient can be learned or preset.

[0061] Through the above design, the DHF-CM module no longer relies on a fixed single attention structure, but provides the PSGN with two types of adjustable execution units: "direction and position" and "high-frequency channel". This enables the detection head to dynamically reconstruct the enhancement ratio of the classification and regression branches based on remote sensing imaging degradation, target density, and mission attributes.

[0062] (4) Multi-task decoupling calculation and final prediction Obtain DHF-CM dynamic modulation and The features are then fed into the final decoupled output mapping layer.

[0063] Classified output: After a Ordinary convolution is used to decouple local features, and then through a... The predictive convolution directly maps the number of channels to the total number of target categories in the current remote sensing dataset. Output classification tensor .

[0064] Regression Output: After the same decoupling and dimensionality reduction mapping, its feature channel number is seamlessly transformed into... (When using target bounding box distribution representations such as DFL loss), output regression tensor .

[0065] Through the complete tensor evolution process described above, this invention achieves heterogeneous decoupled computation in the final prediction stage of YOLO11, driven by scene degradation descriptors, executed by direction-high frequency candidate enhancement graphs, and adaptively adjusted for classification and regression tasks respectively.

[0066] To verify the feasibility, robustness, and scalability of the adaptive heterogeneous modulation detection head (DAH-Head) based on multidimensional scene representation proposed in this invention, this embodiment seamlessly embeds it as the final prediction component into the single-stage object detection network YOLO11.

[0067] The experiment was conducted using VisDrone, a globally recognized authoritative benchmark dataset for remote sensing and UAV ground-based small and dense target detection, through systematic training, simulation, and comparative testing. The explicit physical experimental environment configuration for this embodiment is as follows: Hardware platform: Intel(R) Xeon(R) Gold 6240R CPU@2.40GHz processor, NVIDIA Georce RTX 5090 (32GB VRAM) graphics accelerator card, onboard edge deployment emulator; Software system: The operating system is Ubuntu 22.04 LTS, the deep learning framework is PyTorch 2.1, and the parallel computing architecture is CUDA 12.1; Training hyperparameters: The input image resolution is uniformly set to... The pixel count and batch size were set to 8. An SGD optimizer with cosine annealing was used, the base learning rate was set to 0.01, and a total of 300 epochs were trained.

[0068] This experiment, while maintaining the same backbone topology, jointly performed two types of comparative tests: 1. Comparison with mainstream existing technologies (SOTA): Conventional unenhanced single-stage detectors (such as YOLOv8n, YOLOv10n, YOLO11n) are introduced as the existing technology benchmark; 2. Ablation comparison of improved detection head components: Detection heads (SE-Head, CA-Head) with symmetrical static attention superimposed on the shunt path were introduced as a control group.

[0069] Evaluation metrics include: average accuracy across all categories and single-order threshold. (%), Average accuracy specific to small targets (%), Model physical parameters (M) and floating-point operations per second (G). The overall evaluation results after fusion are shown in Table 1: Table 1. Comparison of detection performance and computational complexity for different basic detection networks and detection head configurations.

[0070] Based on the comprehensive physical test data presented in Table 1, the following rigorous conclusions regarding feasibility and innovative technology can be drawn: (1) The generational leap in precision establishes the absolute technical advantage of this invention over the mainstream existing technology (SOTA) in the industry. A longitudinal comparison was made between the lightweight version of this invention and existing technology benchmarks. Under the same benchmark forward transmission conditions, the DAH-Head of this invention exhibits powerful ground-based small and dense target identification and localization capabilities. It reached 49.6%, which is an improvement of 10.4% compared to the regular YOLOv8n, an improvement of 8.1% compared to the regular YOLOv10n, and an improvement of 5.8% compared to the latest benchmark YOLO11n.

[0071] More significantly, there are specific indicators that directly determine the ability to capture small objects in distant scenes during aerospace remote sensing. A significant improvement was achieved, reaching 29.8% (a jump of 5.6 percentage points compared to the original YOLO11n). This quantitatively demonstrates that the present invention, through feature reconstruction using an adaptive perception mechanism, can greatly suppress the erosion of weak targets by complex backgrounds, proving its high feasibility.

[0072] (2) Successfully broke through the characterization bottleneck of the traditional static hard-coded detection head, verifying the creativity of dynamic heterogeneous modulation. Option 4 (SE-Head) and Option 5 (CA-Head) represent the most advanced static enhancement detection heads in the industry. However, because their internal attention operators and weight fusion logic are hard-coded and fixed, they cannot dynamically sense environmental fluctuations, leading to performance degradation.

[0073] The DAH-Head of this invention reconstructs the enhancement ratio of heterogeneous branches in real time and dynamically through a panoramic statistical gating network (PSGN), achieving a detection accuracy that surpasses that of static SE detection heads by 4.5% and static CA detection heads by 3.4%. This strongly supports, at both mathematical and physical levels, the "scene-specific" dynamic heterogeneous modulation mechanism, representing a fundamental technological leap compared to traditional static enhancement, and is highly innovative.

[0074] (3) Excellent energy consumption control and lightweight computing characteristics meet the real-time deployment requirements of remote sensing edge airborne terminals. Computational overhead is a core consideration related to feasibility. Compared with Scheme 3, after introducing a complete set of adaptive mechanisms including standard deviation collaborative pooling and dual-gated multilayer perceptron (MLP), the number of physical parameters of the model increased by only 0.38M and the computing power consumption GFLOPs increased by only 0.58G.

[0075] This remarkable experimental performance of "high precision, low overhead, and adaptability" is due to the fact that the gating mechanism of this invention operates entirely on a low-dimensional global channel feature vector (with a dimension of only...). By performing forward matrix mapping at the [level] level, spatially dense matrix multiplication at the full-graph level is completely avoided. Data strongly demonstrates that this invention significantly improves network resilience while maintaining excellent inference energy efficiency, fully meeting the real-time computation feasibility requirements for airborne, constrained edge embedded devices such as UAVs.

[0076] The following section introduces some feasible alternatives. 1. Equivalent replacement of scene-aware statistical descriptor operators In specific embodiments, the Panoramic Statistical Gated Network (PSGN) captures multidimensional statistics of feature maps through Global Average Pooling (GAP), Global Max Pooling (GMP), and Global Standard Deviation Pooling (GSDP). As an equivalent alternative design, operators used to quantify image degradation and density include, but are not limited to, Global Range Pooling, Global Skewness Pooling, or Global Kurtosis Pooling.

[0077] Furthermore, the region from which these statistics are extracted is not limited to the entire image; the feature map can also be divided into fixed local spatial regions (such as...). or The mean, maximum and standard deviation of the local feature blocks are calculated within the local region to construct a local panoramic scene description vector rich in spatial block information. This is also an evolutionary variant of the multidimensional statistical perception of the present invention.

[0078] 2. Changes to the topology of the Panoramic Statistical Gated Network (PSGN) In this embodiment, the dynamic gating coefficients are predicted and generated by a dual-path multilayer perceptron (MLP). As an alternative, lightweight one-dimensional convolutional layers can be used instead of fully connected layers to further utilize the local correlations between channels; or extremely lightweight self-attention layers (transformer encoder blocks) can be used to model the nonlinear high-order coupling relationships between the mean, maximum, and standard deviation vectors.

[0079] In addition, the final normalization layer is not limited to using the Softmax activation function to establish complementary rigid game constraints (summing to 1). Alternatively, the output layer can be changed to two independent Sigmoid activation functions, so that the channel and spatial enhancement coefficients of the classification and regression branches exhibit an independent adaptive and elastic state.

[0080] 3. Replacement of heterogeneous enhancement operators within branches The core of this invention lies in the dynamic heterogeneous feature modulation of the classification and regression paths, which is tailored to specific scenarios. In the preferred embodiment, a Directional-High Frequency Co-modulation (DHF-CM) module is used, which includes a Local Multi-Directional Position Aggregation (LMDP) module and a High Frequency Channel Response Modulation (HFCR) module. As an equivalent technical extension, the candidate enhancement operators within the branches can be replaced with any feature extraction mechanism that emphasizes channel-level semantics, high-frequency texture, or spatial geometric location sensitivity.

[0081] For example, the high-frequency channel response modulation module can be replaced with efficient channel recalibration, frequency domain residual pooling, multi-scale grouped channel enhancement, or a large kernel selection mechanism; the local multi-directional position aggregation module can be replaced with strip convolution, directional convolution, deformable convolution, local block coordinate aggregation, or an efficient multi-scale spatial enhancement mechanism. As long as its core logic follows the heterogeneous ratio of the two task branches of multi-dimensional scene prior adaptive modulation, it falls within the protection scope of this invention.

[0082] 4. Changes and replacements to the prediction mapping layer and loss function The DAH-Head component of this invention, during final calculation, is not only applicable to the anchor-free and distributed focus loss (DFL) decoupled calculation architectures described in the embodiments, but can also be directly applied to traditional anchor-based remote sensing detection and prediction layers. Furthermore, the calculation of its bounding box regression loss function can be seamlessly replaced from DFL to CIoU, DIoU, GIoU, or EIoU loss, without substantially altering the final dynamic feature modulation mechanism's topology.

[0083] Other uses of the present invention: The Adaptive Heterogeneous Modulation Detection Head (DAH-Head) designed in this invention is a highly versatile deep learning execution component. Besides its application in detecting small, dense targets in remote sensing images from high-altitude satellites and low-altitude drones, it can also be directly extended and ported to the following application scenarios: 1) Perception of Degraded Targets in Distant Scenes in Autonomous Driving and Intelligent Transportation Environments: In autonomous driving scenarios with onboard cameras, severe weather conditions such as fog, heavy rain, and low light at night can lead to the loss of high-frequency edges (similar to the smooth degradation of remote sensing images). Furthermore, distant pedestrians and vehicles, due to their low resolution, also exhibit the physical properties of tiny targets. This invention can be embedded in the output of an onboard perception network to dynamically perceive the contrast collapse caused by weather through standard deviation, and adjust the classification and localization focus weights of the visual model in real time.

[0084] 2) Microscopic Lesion and Cell Tracking in Intelligent Assisted Diagnosis of Medical Imaging: In microscopic cell imaging, X-ray, or computed tomography (CT) images, early malignant tumor lesions, microthrombi, or specific cells typically occupy only a very small number of pixels, and medical images suffer from non-uniform patchy noise and artifact degradation interference. This invention can be seamlessly applied to the predictive head of medical detectors, dynamically enhancing the characterization of microscopic lesions based on the quality of different scanning scenarios and suppressing noise-induced missed detections.

[0085] 3) Trace Surface Defect Detection in Intelligent Precision Manufacturing: In online industrial vision inspection, trace defects such as cracks, micro-sand holes, and scratches on the surface of metal strips are typically below the millimeter level. Furthermore, due to fluctuations in the light source angle of the production line or surface reflection, the imaging background noise is extremely high. This invention can seamlessly replace the output head of a single-stage defect detection network in industrial applications, utilizing an adaptive gating mechanism to ensure robustness in trace defect classification and boundary localization even under non-uniform fluctuations in the light source.

[0086] 4) Marine and Space Multi-Source Sonar / Synthetic Aperture Radar (SAR) Image Detection: Sonar imaging and SAR radar imaging inherently suffer from severe speckle scattering noise, resulting in extremely low image clarity and severe degradation of target physical contours. This invention can improve the robustness of capturing ship or small underwater debris targets by extracting multi-dimensional global scene sub-sub-targets and forcing the network to adaptively amplify the semantic response of low-degraded dimensions under strong speckle noise interference.

[0087] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0088] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation, characterized in that: This includes parallel configurations of classification enhancement branches, panoramic statistical gating networks, and regression enhancement branches, among which: The panoramic statistical gating network extracts local features from the input feature map based on a single convolutional layer to obtain a first feature map. The first feature map is then subjected to different global statistical operations through parallel global average pooling, global max pooling, and global standard deviation pooling operators to extract three different global statistical features. These three different global statistical features are then cascaded and fed into the classification gating predictor and regression gating predictor for processing, and the corresponding outputs are classification gating coefficients and regression gating coefficients. The goal of panoramic statistical gating networks is to quantify the degree of image degradation and target density of the current input remote sensing image using parameter-free physical statistics. The classification enhancement branch performs convolution operations on the input feature map to obtain the basic classification feature map. After the basic classification feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using classification gating coefficients, and the classification prediction probability distribution is output through independent decoupled convolutional layers. The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After the basic regression feature map is processed by the DHF-CM module, the processing result of the DHF-CM module is processed using regression gating coefficients, and the bounding box regression coordinates are output through an independent decoupled convolutional layer. The specific operation process of the DHF-CM module is as follows: The input feature map is preprocessed with 3×3 depthwise convolution and 1×1 pointwise convolution to obtain shared intermediate features; The shared intermediate features are processed by 3×3 convolution to obtain the second feature map. The second feature map is then processed by the LMDP module to obtain the orientation and position enhancement map. After the shared intermediate features are processed by 3×3 convolution, the third feature map is obtained. The third feature map is then processed by the HFCR module, and the processing result is multiplied with the third feature map to obtain the high-frequency channel enhancement map. The DHF-CM module does not rely on a fixed single attention structure, but provides two types of adjustable execution units for the panoramic statistical gating network: orientation position and high-frequency channel. This enables the detection head to dynamically reconstruct the enhancement ratio of the classification enhancement branch and the regression enhancement branch based on remote sensing imaging degradation, target density, and task attributes. The processing procedure of the LMDP module is as follows: The input feature map is processed by horizontal local pooling, vertical local pooling, first diagonal local pooling, and second diagonal local pooling to obtain features with four different spatial orientations. The features with four different spatial orientations are projected and stitched together to obtain a fused feature map. The fused feature map is processed by 1×1 convolution and Split to obtain the first branch feature and the second branch feature. The first branch feature is processed by 1×1 convolution and Sigmoid activation to obtain the first output feature. The second branch feature is processed by 1×1 convolution and Sigmoid activation to obtain the second output feature. The first output feature, the second output feature, and the input feature map are multiplied to obtain the orientation position enhancement map. The processing procedure of the HFCR module is as follows: The input feature map is processed sequentially by high-frequency average pooling, MLP and Sigmoid to obtain high-frequency channel weights; The high-frequency channel weights are multiplied by the input feature map to obtain the high-frequency channel enhancement map; Within the classification enhancement branch: the high-frequency channel enhancement map generated by the HFCR module is the main component, and the orientation and position enhancement map generated by the LMDP module is the auxiliary component. Adaptive weighted fusion is performed based on the classification gating coefficient to generate a classification enhancement feature map that is highly robust to background degradation noise. Within the regression enhancement branch: the orientation and position enhancement map generated by the LMDP module is the main component, and the high-frequency channel enhancement map generated by the HFCR module is the auxiliary component. Dynamic weighted fusion is performed based on the regression gating coefficient to generate a regression enhancement feature map that is sensitive to the boundary and orientation and position information of small targets. The classification-enhanced feature maps and regression-enhanced feature maps generated after adaptive weighted heterogeneous feature modulation are respectively input into the final task output mapping layer. The two paths pass through independent... Convolutional layers and The task prediction mapping layer performs channel dimensionality reduction and category space projection of target features, thereby outputting the classification prediction probability distribution and bounding box regression coordinates of multi-source remote sensing micro-targets in parallel, and completing the intelligent identification of dense and weak targets in complex fluctuating scenarios.

2. The adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation according to claim 1, characterized in that: The specific operation process of classifying and enhancing branches is as follows: The classification enhancement branch performs convolution operations on the input feature map to obtain a basic classification feature map. After processing by the DHF-CM module, the basic classification feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the classification gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain the first fused attention map. The first fused attention map and the basic classification feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the classification prediction probability distribution is output through an independent decoupled convolutional layer.

3. The adaptive heterogeneous decoupling detection head based on multi-dimensional scene representation according to claim 1, characterized in that: The specific operation process of the regression enhancement branch is as follows: The regression enhancement branch performs a convolution operation on the input feature map to obtain a basic regression feature map. After processing by the DHF-CM module, the basic regression feature map generates a direction and position enhancement map and a high-frequency channel enhancement map in parallel. Based on the regression gating coefficient, the direction and position enhancement map and the high-frequency channel enhancement map are dynamically weighted and fused to obtain a second fused attention map. The second fused attention map and the basic regression feature map are injected with residual features through the matrix element-wise multiplication operator under the broadcast mechanism. Finally, the bounding box regression coordinates are output through an independent decoupled convolutional layer.

Images