Query key value guided asymmetric feature augmentation system and method

Abstract

This invention belongs to the field of computer vision and image processing technology, and particularly relates to a query key-guided asymmetric feature enhancement system and method. The system includes: Q-ASFE modules embedded in the backbone network, specifically in the link where features are transmitted from the C3K2 module to the CBS module. The P2 and P3 layer feature maps output by the backbone network are processed by the neck network to output three feature fusion maps at different scales. These three feature fusion maps are then processed by a detection head to obtain the final detection result. This invention achieves efficient enhancement of image features through the organic integration of spatial detail awareness and global semantic guidance.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and particularly relates to a query key-value-guided asymmetric feature enhancement system and method. Background Technology

[0002] Small target detection is widely used in remote sensing imaging, drone aerial photography, intelligent traffic monitoring, security patrol, and industrial visual analysis. Compared with salient targets in ordinary natural scenes, small targets typically have low pixel proportions, weak texture information, and indistinct edge features, making them easily submerged by noise or background textures in high-resolution complex backgrounds. Especially in remote sensing image scenes, targets such as vehicles, ships, and pedestrians often occupy only a small pixel area and have problems such as large scale variations, complex directional distribution, and dense target arrangement, making it difficult for traditional feature extraction methods to effectively preserve their key structural information. Therefore, how to enhance the local structural response capability of small targets under complex background conditions and improve the effective focus on target regions during feature representation has become an important research direction in the field of small target detection. Existing small target detection and feature enhancement techniques mainly include the following two categories: The first category is enhancement methods based on local convolution feature extraction. Most existing feature enhancement modules use standard convolutional structures to extract local spatial information from input features. Typical methods enhance feature representation capabilities by stacking 3×3 standard convolutions or multi-scale convolutional structures and combine them with feature pyramids to improve the detection performance of targets at different scales. However, standard convolution typically uses isotropic kernels, which have limited sensitivity to the horizontal, vertical, and slender structural targets, making it difficult to fully model the fine-grained directional structural features of small targets. In complex backgrounds, target edge information is easily weakened during feature propagation, thus affecting subsequent target recognition and localization. The second category is feature enhancement methods based on attention mechanisms. To further improve feature representation capabilities, researchers have introduced various attention mechanism modules, such as SE (Squeeze-and-Excitation), CBAM (Convolutional Block Attention Module), Coordinate Attention (CA), and EMA. These methods typically reweight input features by generating weight coefficients in the channel or spatial dimensions to enhance the target region response and suppress background noise. Furthermore, some methods introduce self-attention mechanisms based on query-key-value vector (QKV) structures to establish long-range dependencies between features and improve the ability to model contextual information. However, most existing lightweight QKV attention methods focus on global semantic relationship modeling. Their attention weight generation process is independent of local spatial structure extraction. They lack a joint enhancement mechanism for direction-sensitive detail features and cannot simultaneously take into account the local structural representation ability of small targets and the semantic suppression ability in complex backgrounds. Summary of the Invention

[0003] In view of this, the present invention aims to provide a query key-guided asymmetric feature enhancement system and method to address the problems of insufficient directional feature extraction, lack of global semantic guidance, and poor synergy between lightweight QKV attention and spatial feature extraction in existing small target detection technologies, thereby improving the feature representation and detection performance of small targets in complex backgrounds. The present invention provides a QKV-guided asymmetric spatial feature enhancement module (Q-ASFE), which achieves efficient enhancement of image features through the organic integration of spatial detail awareness and global semantic guidance.

[0004] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A query key-value-guided asymmetric feature enhancement system, which is an improvement on YOLOv11, includes a backbone network, a neck network, and a detection head, wherein: In the backbone network, Q-ASFE modules are embedded in the link from C3K2 module to CBS module to transmit features. The P2 layer feature map and P3 layer feature map output by the backbone network are processed by the neck network to output three feature fusion maps of different scales. The three feature fusion maps of different scales are processed by the detection head to obtain the final detection result. The Q-ASFE module consists of an asymmetric spatial feature extraction submodule, an efficient QKV semantic modulation submodule, and a joint fusion output submodule. The asymmetric spatial feature extraction submodule captures spatial features of three different categories and concatenates these features to obtain a feature mosaic map. The efficient QKV semantic modulation submodule, based on the Query-Key-Value mechanism, generates a spatial attention-gated weight map guided by global semantic vectors, and generates a QKV enhanced feature map based on the spatial attention-gated weight map. The joint fusion output submodule concatenates the feature mosaic map and the QKV enhanced feature map to obtain the final enhanced feature map.

[0005] Furthermore, the processing steps of the asymmetric spatial feature extraction submodule include: The input feature map is compressed through a 1×1 convolutional layer to obtain a low-channel feature map. The low-channel-number feature map is input into the standard isotropic branch, the lateral asymmetric branch, and the longitudinal asymmetric branch for processing, respectively, to obtain local spatial features, edge and texture features, and detail features. Local spatial features, edge and texture features, and detail features are stitched together to obtain a three-way feature stitched image.

[0006] Furthermore, the processing procedure for the standard isotropic branch is as follows: ; in, For local spatial features, This is a combination of 3×3 convolution, batch normalization, and activation functions. This is a combination of 1×1 convolution, batch normalization, and activation functions. Let C be the input feature map, H be the channel dimension of the input feature map, W be the height of the input feature map, and R be a real number. The process for handling lateral asymmetric branches is as follows: ; in, For edge and texture features, This is a combination of 1×3 convolution, batch normalization, and activation function operations, where d is the dilation rate of the dilated convolution. This is a combination of 3×1 convolution, batch normalization, and activation functions; The process for handling vertical asymmetric branches is as follows: ; in, For detailed features.

[0007] Furthermore, the calculation formula used to stitch together local spatial features, edge and texture features, and detail features to obtain the three-way feature stitched image is as follows: ; in, This is a spliced ​​image of three features. For local spatial features, For edge and texture features, For detailed features, C is the channel dimension of the input feature map, H is the height of the input feature map, W is the width of the input feature map, and R is a real number.

[0008] Furthermore, the processing steps of the efficient QKV semantic modulation submodule include: Global semantic compression is performed on the input feature map to generate the query vector: ; in, For the query vector, For activation function, For 1×1 convolution, ( ) represents global average pooling, and X is the input feature map. R is the number of channels, and R is a real number; Local spatial awareness is applied to the input feature map to generate a key feature map: ; Where C is the channel dimension of the input feature map, H is the height of the input feature map, and W is the width of the input feature map. For the key feature map, For the number of channels, ( ) represents a 5×5 depthwise separable convolution. This is a combination of 1×1 convolution, batch normalization, and activation functions; Convolutional mapping is performed on the input features to generate the Value feature map: ; Where V is the Value feature map; Attention gating is calculated based on the query vector, key feature map, and value feature map to obtain a spatial attention gating weight map. ; in, For spatial attention gating weights, For element-wise multiplication; Introducing a learnable scaling factor The Value feature map is modulated to obtain the QKV enhanced feature map. : ; ; in, For the expanded spatial attention gating weight map, () represents an extension operation.

[0009] Furthermore, the joint fusion output submodule concatenates the feature stitched map and the QKV enhanced feature map to obtain the final enhanced feature map: ; in, For the final enhanced feature map, For activation function, t represents feature concatenation. For 1×1 convolution, This is a spliced ​​image of three features. Enhance the QKV feature map. This is the residual shortcut mapping function.

[0010] A query key-driven asymmetric feature enhancement method specifically includes the following steps: S1: Acquire the original remote sensing images, preprocess the original remote sensing images, and obtain the training set; S2: Train the query key-value-guided asymmetric feature enhancement system using the training set to obtain a trained query key-value-guided asymmetric feature enhancement model; S3: Input the remote sensing image to be processed into the trained query key-value-guided asymmetric feature enhancement model to obtain the prediction result.

[0011] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The query key-guided asymmetric feature enhancement system and method described in this invention significantly improves the spatial detail perception capability of directional targets through three asymmetric convolutional branches. Existing technologies (such as CBAM and EMA) rely on isotropic standard convolutional kernels for feature extraction, which are insufficient for perceiving the fine-grained spatial structure of small targets with directional shapes (such as slender vehicles and horizontally arranged ships) in remote sensing images, resulting in the accelerated loss of key features during downsampling. This invention designs three parallel asymmetric convolutional branches through the ASFE submodule, which simultaneously extract local spatial detail features in the isotropic, horizontal, and vertical directions through three paths: standard 3×3 convolution, horizontal 1×3→3×1→expanded 3×3, and vertical 3×1→1×3→expanded 3×3. Furthermore, an expanded convolution with an dilation rate of d=2 is introduced at the end of each branch to expand the receptive field. Compared to using ASFE alone (mAP50=48.1%) with any existing attention module, the Q-ASFE of this invention achieves the best result of mAP50=48.9% on the VisDrone dataset, with an APS of 20.5% for small targets, surpassing CBAM (18.6%), CA (19.1%), EMA (19.2%), DCNv3 (19.5%) and ASFE alone (19.9%), verifying the effectiveness of asymmetric multi-directional perception design in preserving detailed features of directional small targets.

[0012] (2) The query key-value guided asymmetric feature enhancement system and method described in this invention have a global semantic query guidance mechanism, which effectively suppresses complex background interference and reduces the false detection rate. Existing attention methods (CBAM, CA, EMA) rely entirely on the local statistical information (such as mean or maximum value) of the feature map itself for attention weights, lacking global semantic guidance. Under complex remote sensing backgrounds (ground spots, tree canopies, and other texture interference), it is difficult to accurately focus attention resources on the real target area, resulting in a high false detection rate. The Efficient QKV submodule of this invention compresses the entire feature map into a global semantic vector Q (shape Cr×1×1) through global average pooling (GAP), and guides the generation of gating weights of the key feature map with global context prior, so that the attention weights simultaneously encode two types of information: "global semantic location" and "local spatial structure". On the USOD dataset (99.9% small target ratio, low-light scene), Q-ASFE achieved an AP50 of 92.4%, surpassing all comparison methods. The AP50:95 reached 37.3%, an improvement of 2.0 percentage points compared to the baseline, directly reflecting the contribution of global semantic guidance to the accurate localization of targets in complex backgrounds.

[0013] (3) The query key-value guided asymmetric feature enhancement system and method described in this invention breaks through the traditional module stacking and constructs a deep collaborative mechanism of "direction awareness + semantic guidance". In existing fusion schemes (such as simply connecting CNN and Attention in parallel), attention calculation and spatial feature extraction are often independent of each other and lack interaction. In essence, it is a physical patching of features, which can easily lead to confusion between background noise and target structure features in complex remote sensing backgrounds. The advantage of this invention is that it is not a simple combination of asymmetric convolution and attention mechanism, but rather a deep collaborative coupling architecture. The global semantic prior generated by Query in the EfficientQKV branch not only completes the spatial gating of Value within the sub-module, but also, in the joint fusion stage (cross-channel interaction layer), directly participates in the recalibration process of the asymmetric directional features extracted by ASFE as a high-dimensional semantic guide. This design enables the network to dynamically modulate the response intensity of asymmetric directional features during inference by utilizing global semantic features (e.g., using global semantics to identify "road background" and thus suppress the "road edge" response extracted by ASFE; using global semantics to lock onto "port area" and thus amplify the "slender ship" response extracted by ASFE). Through this joint enhancement mechanism of "direction awareness + semantic guidance", this module generates a synergistic gain that far exceeds the "1+1>2" effect of each substructure operating alone.

[0014] (4) The query key-value-guided asymmetric feature enhancement system and method described in this invention have excellent lightweight deployment potential due to the extremely limited increase in the number of parameters and computational cost. Existing high-performance alternatives (such as DCNv3 and FFCA-YOLO) are usually accompanied by a significant increase in the number of parameters and computational cost. DCNv3 has 1.12M parameters and 26.42 GFLOPs, while FFCA-YOLO has as many as 7.14M parameters and 131.18 GFLOPs. However, the Q-ASFE module of this invention only increases the number of parameters from the baseline of 0.88M to 1.02M, a net increase of 0.14M (an increase of about 15.9%), and the GFLOPs increase from 23.32 to 30.58, with a controllable increase. With minimal cost, Q-ASFE achieved the best mAP50 (48.9%) and APS (20.5%) among all comparison methods, fully demonstrating the excellent balance between accuracy and efficiency achieved by the present invention, and is suitable for resource-constrained deployment scenarios such as UAV onboard processors and edge computing devices.

[0015] (5) The query key-value guided asymmetric feature enhancement system and method described in this invention achieve optimal system performance at extremely low cost based on the precise deployment strategy of shallow high-resolution nodes. Existing methods usually deploy feature enhancement modules evenly in each layer of the network, lacking a basis for layer selection for small target detection tasks. This invention clearly verifies the optimal deployment position of Q-ASFE through systematic ablation experiments. Taking the YOLO11n network as an example: Q-ASFE deployed in layer P2 (APS=19.3%) is better than layer P3 (APS=18.5%), and jointly deployed in P2+P3 (APS=20.5%) is better than any single-node configuration. Moreover, under the lightweight P2–P4 architecture, with a configuration of only 1.02M parameters, it comprehensively surpasses all deployment schemes under the P2–P5 architecture (P2 / P4 configuration APS=20.2%, P2 / P3 / P4 configuration APS=19.9%). This demonstrates that the present invention not only proposes an effective feature enhancement module, but also guides the optimal use of the module through an experiment-driven deployment strategy, achieving better detection performance without retaining deep P5 nodes, and has significant engineering practice guidance value.

[0016] (6) The present invention creates an asymmetric feature enhancement system and method guided by query key values. Most existing feature enhancement modules use standard convolutional structures for local spatial feature extraction. Standard convolution usually uses isotropic convolution kernels, which have limited sensitivity to the direction of horizontal, vertical and slender structural targets, making it difficult to fully model the fine-grained directional structural features of small targets. Under complex background conditions, the target edge and local texture information are easily weakened during feature propagation, thus affecting the accuracy of subsequent target recognition and localization. To address the above problems, the present invention designs an asymmetric spatial feature extraction structure, which uses standard convolutional branches, horizontal asymmetric convolutional branches and vertical asymmetric convolutional branches to form a multi-branch direction perception mechanism, so as to simultaneously extract isotropic, horizontal and vertical local structural features, enhance the module's ability to express the details of direction-sensitive targets, and reduce the loss of key structural information during feature extraction.

[0017] (7) The query key-value guided asymmetric feature enhancement system and method described in this invention largely rely on the local statistical information of the input feature map itself to generate weight coefficients, lacking the ability to guide the spatial response region through global semantic priors. In complex background scenes, feature confusion easily occurs between background textures and real targets, making it difficult for the model to effectively focus on the target region, resulting in false detections or background false responses. To address the above problems, this invention introduces a lightweight query generation mechanism based on global average pooling, using global semantic vectors to guide and modulate the spatial feature response, enabling the feature enhancement process to simultaneously combine global semantic information and local spatial structure information, improving the response capability of the target region and suppressing background noise interference.

[0018] (8) The present invention creates a query key-value-guided asymmetric feature enhancement system and method. Existing methods, when attempting to combine attention mechanisms with spatial convolution, typically employ simple serial stacking or parallel independent computation, where attention weight generation and spatial feature extraction are independent of each other. This mechanical combination leads to a disconnect between "semantic guidance" and "structure awareness"—global semantics cannot accurately guide the extraction of local structures, and local details cannot affect the allocation of global weights. Ultimately, this not only results in a large amount of redundant computation but also easily leads to semantic and structural feature misalignment in complex backgrounds. To address the above problems, the present invention proposes a QKV-guided asymmetric spatial feature enhancement module (Q-ASFE). The present invention directly uses the global semantic prior generated by the query to modulate the asymmetric directional feature response, forming a joint enhancement mechanism of "directional awareness + semantic guidance." Through the synergistic effect of global semantic guidance and directional structure modeling, the local structure enhancement and background noise suppression are unified, achieving deep synergy between local structure enhancement and background noise suppression in the feature dimension. Attached Figure Description

[0019] 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 1 This is a schematic diagram illustrating the deployment of the Q-ASFE module in a YOLO11n network according to an embodiment of the present invention. Figure 2 The internal structural topology diagram of the Q-ASFE module described in the embodiments of the present invention; Figure 3 This is a flowchart illustrating the query key-value-guided asymmetric feature enhancement method described in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0022] 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.

[0023] 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.

[0024] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] like Figure 1 As shown, this invention provides a query key-value-guided asymmetric feature enhancement system. This asymmetric feature enhancement system is an improvement upon YOLOv11 and includes a backbone network, a neck network, and a detection head, wherein: In the backbone network, Q-ASFE modules are embedded in the link from C3K2 module to CBS module to transmit features. The P2 layer feature map and P3 layer feature map output by the backbone network are processed by the neck network to output three feature fusion maps of different scales. The three feature fusion maps of different scales are processed by the detection head to obtain the final detection result. The Q-ASFE module consists of an asymmetric spatial feature extraction submodule, an efficient QKV semantic modulation submodule, and a joint fusion output submodule. The asymmetric spatial feature extraction submodule captures spatial features of three different categories and concatenates these features to obtain a feature mosaic map. The efficient QKV semantic modulation submodule, based on the Query-Key-Value mechanism, generates a spatial attention-gated weight map guided by global semantic vectors, and generates a QKV enhanced feature map based on the spatial attention-gated weight map. The joint fusion output submodule concatenates the feature mosaic map and the QKV enhanced feature map to obtain the final enhanced feature map.

[0026] It should be noted that embedding the Q-ASFE module into the shallow high-resolution feature layers (such as P2 and P3 nodes) of the backbone layer of the target detection network (such as YOLO11n) is used to enhance the structural response of small targets and suppress interference from complex backgrounds, thereby significantly improving detection accuracy.

[0027] like Figure 1As shown, the deployment position of the Q-ASFE module in the YOLO11n detection network of this invention is as follows: The invention is divided into three parts: the backbone network, the neck network, and the head network. The backbone network adopts a three-scale structure of P2–P4, removing the P5 deep downsampling node in the traditional YOLO architecture. The Q-ASFE module (highlighted in bold in the figure) is inserted after the P2 node (output size 256×256×256) and the P3 node (output size 512×128×128) of the backbone network, and before the concat operation of the neck network, to enhance the shallow high-resolution feature maps. The neck network achieves multi-scale feature fusion through upsampling and concat concatenation, and finally, the three scale detection heads (Head P2 / P3 / P4) output detection results at three resolutions of 256×256, 128×128, and 64×64, respectively.

[0028] Next, we will introduce each module included in the Q-ASFE module in detail.

[0029] The asymmetric spatial feature extraction submodule (ASFE) is designed with three parallel asymmetric convolutional branches, including a standard isotropic branch (by...). Convolution and Composed of convolutional cascades), and lateral asymmetric branches (sequentially composed of...) Horizontal Vertical and an expansion rate of 2 (composed of convolutional cascades) and longitudinal asymmetric branches (sequentially composed of...) Vertical Horizontal and an expansion rate of 2 (Composed of convolutional cascades). Each branch extracts local spatial detail features from different directions, and after channel concatenation (Concat), a multi-directional perceptual spatial feature map is obtained.

[0030] The Efficient QKV semantic modulation submodule generates lightweight query vectors (Query) representing global semantic priors through global average pooling (GAP); Depthwise separable convolutions generate key feature maps that preserve spatial location information; and then... Convolutional mapping generates a value feature map (Value). The query and key are multiplied element-wise via broadcast and gating with a sigmoid gate to generate a spatial attention-gated weight map. This weight map is then repeatedly expanded and applied to the value, which is then multiplied by a learnable scaling factor. This yields enhanced features guided by global semantics.

[0031] like Figure 2 As shown, the two parallel substructures of the efficient QKV semantic modulation submodule and their fusion method are as follows: the left side of the efficient QKV semantic modulation submodule is the efficient QKV semantic modulation branch, and the input feature map is... After global average pooling (AvgPool) and Convolution generates query vectors ,through Depthwise Separable Convolution (DWConv) and Convolution generates key feature maps and through Convolution generates Value feature maps Subsequently, and The spatial attention gating weight map is obtained by multiplying the elements one by one through broadcasting and then activating it with a Sigmoid function. Then through learnable scaling factors right Spatial modulation is performed to obtain QKV enhanced features. The right side of this module is the Asymmetric Spatial Feature Extraction (ASFE) branch, which includes standard... Branches, horizontal (including) , and expansion Convolutional branches and vertical (including) branches , and expansion The three parallel asymmetric structures (convolutional branches) are concatenated to obtain multi-directional spatial features. Finally, the features from the left and right sides are concatenated using Concat, and then... Convolutions are fused and combined with residual shortcut branches. Element-wise addition is performed, and finally, the result is processed by the SiLU activation function to output the final enhanced feature map. .

[0032] Joint Fusion Output Submodule: This submodule serves as the collaborative interaction center for "direction awareness" and "semantic guidance." It concatenates the spatial features of multi-directional perception with the enhanced features of global semantic guidance along the channels, and then outputs the results. Convolutional layers are used for cross-dimensional feature recalibration. In this process, the global semantic guidance features output by the QKV branch serve as contextual priors. In the joint fusion layer, these features are concatenated and interacted across channels with the asymmetric directional features output by the ASFE branch. This allows the fused features to possess both directional detail awareness and global semantic suppression capabilities.

[0033] In some embodiments, the processing of the P2 layer feature map or the P3 layer feature map by the asymmetric spatial feature extraction submodule includes: The input feature map is compressed through a 1×1 convolutional layer to obtain a low-channel feature map. The low-channel-number feature map is input into the standard isotropic branch, the lateral asymmetric branch, and the longitudinal asymmetric branch for processing, respectively, to obtain local spatial features, edge and texture features, and detail features. Local spatial features, edge and texture features, and detail features are stitched together to obtain a three-way feature stitched image.

[0034] In some embodiments, the processing procedure for standard isotropic branches is as follows: ; in, For local spatial features, This is a combination of 3×3 convolution, batch normalization, and activation functions. This is a combination of 1×1 convolution, batch normalization, and activation functions. Let C be the input feature map, H be the channel dimension of the input feature map, W be the height of the input feature map, and R be a real number. The process for handling lateral asymmetric branches is as follows: ; in, For edge and texture features, This is a combination of 1×3 convolution, batch normalization, and activation function operations, where d is the dilation rate of the dilated convolution. The process for handling vertical asymmetric branches is as follows: ; in, For detailed features.

[0035] In some embodiments, the calculation formula used to stitch together local spatial features, edge and texture features, and detail features to obtain a three-way feature stitched image is as follows: ; in, This is a spliced ​​image of three features. For local spatial features, For edge and texture features, For detailed features, C is the channel dimension of the input feature map, H is the height of the input feature map, W is the width of the input feature map, and R is a real number.

[0036] In some embodiments, the processing procedure of the efficient QKV semantic modulation submodule includes: Global semantic compression is performed on the input feature map to generate the query vector: ; in, For the query vector, For activation function, For 1×1 convolution, ( ) represents global average pooling, X is the input feature map, C is the channel dimension of the input feature map, H is the height of the input feature map, W is the width of the input feature map, and R is a real number; Local spatial awareness is applied to the input feature map to generate a key feature map: ; in, For the key feature map, For the number of channels, ( ) represents a 5×5 depth-separable convolution; Convolutional mapping is performed on the input features to generate the Value feature map: ; Where V is the Value feature map; Attention gating is calculated based on the query vector, key feature map, and value feature map to obtain a spatial attention gating weight map. ; in, For spatial attention gating weights, For element-wise multiplication; Introducing a learnable scaling factor The Value feature map is modulated to obtain the QKV enhanced feature map. : ; ; in, For the expanded spatial attention gating weight map, () represents an extension operation.

[0037] In some embodiments, the joint fusion output submodule concatenates the feature stitched map and the QKV enhanced feature map to obtain the final enhanced feature map: ; in, For the final enhanced feature map, For activation function, t represents feature concatenation. For 1×1 convolution, This is a spliced ​​image of three features. Enhance the QKV feature map. This is a residual shortcut mapping function used to map input features. Map to the same channel dimension as the main branch output; when the dimensions are the same, it is an identity mapping; when the dimensions are different, it is a different mapping. Convolution is used for channel alignment.

[0038] like Figure 3 As shown, the present invention also provides a query key-value-guided asymmetric feature enhancement method, which specifically includes the following steps: S1: Acquire the original remote sensing images, preprocess the original remote sensing images, and obtain the training set; S2: Train the query key-value-guided asymmetric feature enhancement system using the training set to obtain a trained query key-value-guided asymmetric feature enhancement model; S3: Input the remote sensing image to be processed into the trained query key-value-guided asymmetric feature enhancement model to obtain the prediction result.

[0039] It should be noted that the specific process of the query key-value-guided asymmetric feature enhancement system performing relevant detection is as follows: Step S1: Preprocess the input image Acquire raw remote sensing images (multi-source images such as drone aerial photography and satellite remote sensing), uniformly scale them to a fixed resolution (e.g., 1024×1024 pixels), and perform normalization processing to obtain the input tensor. ,in, For batch size, .

[0040] Step S2: Backbone Network Feature Extraction The preprocessed image is input into the backbone network (using YOLO11n as the base network in this invention) for layer-by-layer feature extraction. After stacking multiple CBS (convolution-batch normalization-SiLU activation) modules and C3K2 modules, feature maps of different scales are obtained sequentially. 1. Feature map of P2 layer: (1 / 4 downsampling to retain the richest fine-grained spatial information); 2. Feature map of P3 layer: (1 / 8 downsampling); 3. Feature map of layer P4: (1 / 16 downsampling); Design Description: This invention removes the P5 downsampling node (1 / 32 scale) from the traditional YOLO architecture, retaining the three detection scales: P2, P3, and P4. The P2 layer retains the shallowest features with the highest resolution, is the layer with the richest detail information for tiny targets, and is also the core layer of the Q-ASFE module.

[0041] Step S3: Q-ASFE module feature enhancement The P2 layer feature map output by the backbone network With P3 layer feature map Each feature map is fed into its corresponding Q-ASFE module for asymmetric spatial feature extraction and QKV semantic-guided modulation, and then outputs an enhanced feature map, which is then fed into the neck network for multi-scale feature fusion.

[0042] Step S4: Multi-scale neck feature fusion The enhanced P3 feature map and the upsampled result of the P4 feature map are concatenated and then fused using C3K2 to obtain the P3 fused feature map; the P2 feature map is then concatenated with the upsampled result of the P3 fused feature map and then fused using C3K2, finally outputting fused feature maps of three scales, corresponding to three detection resolutions of 256×256 (P2), 128×128 (P3), and 64×64 (P4), respectively.

[0043] Step S5: The detection head outputs the prediction result. The fused feature maps at the three scales are fed into the detection head, and the target bounding box coordinates (regression branch) and class confidence (classification branch) at each scale are output. The final detection result is obtained by post-processing with NMS (non-maximum suppression).

[0044] Next, the internal structure of the Q-ASFE module will be described in detail.

[0045] The Q-ASFE module consists of two parallel sub-structures: the Asymmetric Spatial Feature Extraction (ASFE) sub-module and the Efficient QKV semantic modulation sub-module. The outputs of the two are jointly fused at the end of the module and the output is enhanced through residual connections.

[0046] Overall input / output definition: Input: Feature map ; Output: Final enhanced feature map (Same shape as input, facilitating residual connection); 4.1 Submodule 1: Asymmetric Spatial Feature Extraction (ASFE) The ASFE module uses three parallel asymmetric convolutional branches to perceive the spatial detail features of small targets from different directions. The specific steps are as follows: Step S411: Channel Compression Input feature map First, go through a Convolution performs channel compression, reducing the number of channels from... Compress to The low-channel-number feature map is used as the common input for the three branches, reducing the amount of subsequent computation.

[0047] Step S412: Standard isotropic branch (l-branch): ; pass After convolution and channel adjustment, use Standard convolution extracts balanced local spatial features in all directions, which serve as the basic perceptual branch.

[0048] Step S413: Lateral asymmetric branch (x-branch) ; Passing through in sequence convolution, Lateral convolution (perceives only in the horizontal direction) Vertical convolution (locally assisted), and dilation rate of Dilated convolution (expands the receptive field). This branch focuses on extracting edge and texture features extending horizontally from the target.

[0049] Step S414: Vertical asymmetric branch (y-branch) ; Symmetrical to the x-branch structure, and By reversing the order of convolutions, the focus is on extracting detailed features in the vertical direction of the target.

[0050] Step S415: Three-way feature splicing ; Design highlights: The three branches capture three types of spatial features: isotropic, horizontal, and vertical. Without significantly increasing the parameters, the module’s ability to perceive small, directional targets (such as slender vehicles, horizontally lined boats, etc.) is significantly improved.

[0051] 4.2 Submodule 2: Efficient QKV Semantic Modulation The Efficient QKV module is based on the Query-Key-Value mechanism. It guides the generation of a spatial attention-gated weight map through a global semantic vector and performs semantic modulation on the ASFE output features.

[0052] Step S421: Query generation (global semantic compression): ; For the input feature map Perform global average pooling (GAP) to compress the spatial dimension to After Convolution compresses the number of channels to Finally, it is activated by the SiLU function. Obtain the global semantic query vector.

[0053] Step S422: Key Generation (Local Spatial Awareness): ; For the input feature map First Depthwise separable convolution (DWConv) is used for local spatial awareness, followed by... Convolution compresses the number of channels to This yields a Key feature map that retains spatial location information.

[0054] Step S423: Value generation: ; Input feature map through Convolution directly maps and preserves the full number of channels. This serves as the Value feature map to be modulated, i.e., the Value feature map.

[0055] Step S424: Calculation of spatial attention gating weight map: ; The query vector is broadcast and multiplied element-wise with the key feature map, and then activated by Sigmoid to obtain a spatial attention gating weight map.

[0056] Step S5: Weight expansion and modulation of the value feature map: ; ; Introducing a learnable scaling factor (Initialized to 0, adaptively adjusted during training) Modulate the Value feature map.

[0057] 4.3 Fusion Output: Joint Feature Fusion and Residual Connectivity The output of the ASFE module With the output of the Efficient QKV module Perform Concat stitching along the channel dimension, then... Convolution restores the number of channels to Finally, it is compared with the input feature map. Residual shortcut (through Convolutional alignment channels are performed element-wise, and the final enhanced feature map is output: ; ; 4.4 Description of the deployment location of the Q-ASFE module In a preferred embodiment, the present invention embeds the Q-ASFE module at the junction of the backbone network and the neck network, specifically deployed at two locations: the P2 scale (1 / 4 resolution, high fine-grained features) and the P3 scale (1 / 8 resolution), corresponding to spatial resolutions of 256×256 and 128×128 for the feature maps, respectively. The two Q-ASFE modules operate independently without sharing parameters, enhancing the feature maps at their respective scales. The P4 scale (64×64) feature map has rich semantic information but suffers from significant loss of detail; therefore, no Q-ASFE module is deployed there to control the overall parameter and computational load.

[0058] This invention uses the original object detection loss function of YOLO11 during the training phase, without introducing any additional auxiliary loss. Total Loss It consists of classification loss, bounding box geometric regression loss, and bounding box discrete distribution regression loss, and is expressed as follows: ; in, This is the binary cross-entropy classification loss, 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.

[0059] To demonstrate the effectiveness of the present invention, a core ablation experiment was conducted to verify the effectiveness of the Q-ASFE module components.

[0060] To verify the individual contributions of the two substructures in the Q-ASFE module, namely the Asymmetric Spatial Feature Extraction (ASFE) module and the Efficient QKV Modulation (EQKV) module, as well as the gain of their synergistic fusion, ablation experiments were conducted on two publicly available remote sensing target detection datasets, VisDrone and USOD, with YOLO11n (P2–P4 architecture, without any feature enhancement modules) as the baseline. The results are shown in Table 1.

[0061] Table 1. Performance comparison results of Q-ASFE module and its substructure with existing feature enhancement modules.

[0062] The following four beneficial effects can be derived from Table 1: Effect 1: Compared to the baseline, accuracy is improved across the board, and the ability to detect small targets is significantly enhanced.

[0063] Effect 2: The synergistic fusion of ASFE and QKV produces gains that surpass those of using them individually.

[0064] Effect 3: Compared with similar attention modules, it has the best overall performance.

[0065] Effect 4: The increase in the number of parameters and computation is extremely limited, highlighting its lightweight characteristics.

[0066] The deployment location of Q-ASFE was verified using a whole-system ablation experiment.

[0067] Table 2 Comparison of target detection performance of Q-ASFE module at different feature layer deployment locations.

[0068] The following three conclusions can be drawn from Table 2: Conclusion 1: Under the P2–P5 architecture, the gain of the Q-ASFE module extending to deeper levels does not accumulate continuously.

[0069] Conclusion 2: The P2–P4 architecture, combined with the P2+P3 joint deployment, surpasses the configuration of all Q-ASFE modules under the P2–P5 architecture with extremely low parameter cost.

[0070] Conclusion 3: The Q-ASFE module contributes more to the P2 layer than to the P3 layer, and the shallow high-resolution features are the core source of the module gain.

[0071] In summary, the Q-ASFE module proposed in this invention achieves a comprehensive improvement in accuracy, small target detection capability, and anti-interference ability of the baseline model in multi-source remote sensing small target detection tasks with minimal parameter and computational cost. It effectively solves the three core defects mentioned in the background technology and has significant technical value and practical significance.

[0072] Each core component of this invention has feasible alternative implementations, as detailed below: 1. Alternatives to ASFE submodules: Alternative Solution 1: Replacement for the number of branches in asymmetric convolution This invention employs a three-way parallel branching approach (isotropic, lateral, and longitudinal). An alternative can be extended to a four-way branching approach, adding an additional diagonal asymmetric branch (e.g., a combination of 1×1 and oblique convolution) to further enhance the perception of obliquely arranged targets (e.g., slanted roads, inclined runways, etc.). Alternatively, it can be simplified to a two-way branching approach (retaining only lateral and longitudinal branches), reducing computational load in edge deployment scenarios with extremely limited computing resources while still retaining the core advantages of asymmetric perception.

[0073] Alternative Solution 2: Replacement of Dilated Convolutional Expansion Rate In this invention, the ends of each asymmetric branch are expanded using a 3×3 dilated convolution with a dilation rate of d=2 to increase the receptive field. This can be replaced with a dilated convolution with d=3 or d=4 to adapt to remote sensing scenes with larger target scales; alternatively, the dilated convolution can be replaced with an ASPP (Spatial Pyramid Pooling with Holes) structure, which integrates features of multiple dilation rates to further enhance the multi-scale receptive field coverage.

[0074] Alternative Solution 3: Alternatives to Channel Compression Ratio This invention compresses the number of input channels to C / 4 (compression ratio of 4). It can be replaced by C / 2 (compression ratio of 2) or C / 8 (compression ratio of 8). The former increases the feature capacity when the number of parameters is sufficient, while the latter further reduces the computational overhead when lightweight deployment requirements are higher. Neither of them affects the overall structural logic of the module.

[0075] Alternative Solution 4: Replacement of Branch Activation Functions The SiLU activation function is used in each CBS module of this invention. It can be replaced by activation functions such as ReLU, GELU, or Mish to adapt to different training frameworks or hardware acceleration requirements without affecting the core mechanism of asymmetric spatial perception.

[0076] 2. Alternatives to the Efficient QKV submodule: Alternative Solution 1: Replacement of Query Vector Generation Method This invention generates query vectors using Global Average Pooling (GAP), which can be replaced by Global Max Pooling (GMP). Alternatively, the outputs of GAP and GMP can be concatenated to generate the query vector, simultaneously encoding the mean semantics and peak response information of the feature map, thus enhancing the ability to perceive salient target regions. It can also be replaced by the SE module generation method in channel attention (GAP + two-layer MLP) to achieve stronger channel selectivity.

[0077] Alternative Solution 2: Key-Generated Convolution Alternative This invention uses a 5×5 depthwise separable convolution (DWConv) to generate key feature maps. It can be replaced by a 3×3 depthwise separable convolution (lower computational cost) or a 7×7 depthwise separable convolution (larger receptive field) to adapt to feature layers of different resolutions; it can also be replaced by a regular 3×3 convolution to ensure compatibility in deployment scenarios where hardware support for depthwise separable convolution is incomplete.

[0078] Alternative Solution 3: Replacement of Gated Activation Functions This invention employs the Sigmoid function in the calculation of attention gating weights. It can be replaced by the Softmax function (which normalizes the global attention weights) or the Tanh function (which extends the weight range to [-1,1], allowing for inhibitory modulation). Different activation functions correspond to different assumptions about the attention weight distribution, and can be selected according to the specific task characteristics.

[0079] Alternative Option 4: An alternative to initialization with learnable scaling factor γ This invention initializes γ to 0 to ensure that the modules are equivalent to identity mappings in the early stages of training. Alternatively, it can be initialized to a small positive value (such as 0.01 or 0.1), which can accelerate the activation of module feature modulation capabilities and shorten the number of training rounds required for convergence in certain pre-training transfer scenarios.

[0080] 3. Alternatives to the merged output method: Alternative Solution 1: Replacement of ASFE and QKV Feature Fusion Method This invention performs channel-dimensional concatenation between the outputs of the ASFE branch and the QKV branch, followed by 1×1 convolutional fusion. Alternatively, element-wise addition can be used for fusion to further reduce the number of parameters; alternatively, concatenation followed by depthwise separable convolutional fusion can be used to enhance cross-channel feature interaction capabilities while maintaining lightweight design.

[0081] Alternative Solution 2: Replacement of Residual Connection Method This invention employs standard residual connections (identical mapping when input and output dimensions are the same, otherwise 1×1 convolutional alignment). It can be replaced by weighted residual connections (introducing learnable scalars to weight the main path and residual paths separately), or by dense connections (introducing intermediate layer features into the residual paths) to further enhance gradient flow and feature reuse capabilities in specific scenarios.

[0082] 4. Alternative deployment locations for modules: Alternative Option 1: Alternative Deployment Scale This invention deploys the Q-ASFE module on two shallow nodes, P2 and P3. An alternative approach is to deploy it only on a single node, P2 (minimizing the number of parameters, suitable for extremely lightweight scenarios), or simultaneously on three nodes, P2, P3, and P4 (covering more scales, suitable for scenarios with greater target scale variations). Experiments show that the latter offers limited performance improvement under the architecture of this invention, but may produce additional gains in deeper backbone networks (such as larger-scale models like ResNet-50 and YOLO11m / x).

[0083] Alternative Solution 2: Replacement of Backbone Network Architecture This invention is based on YOLO11n (P2–P4 three-scale) as its backbone. The Q-ASFE module is a plug-and-play module that can be migrated to other detection frameworks, including but not limited to single-stage detectors such as YOLOv8, YOLOv9, YOLOv10, and RT-DETR, as well as the feature extraction backbone layers of two-stage detectors such as Faster R-CNN. Q-ASFE modules can be inserted into the corresponding shallow high-resolution nodes to obtain similar small target detection gains.

[0084] Other uses of the present invention include: Application 1: Other remote sensing target detection tasks This invention is designed for remote sensing micro-target detection, and is also applicable to other remote sensing target detection tasks, including ship detection in satellite images, airport runway target detection, and vehicle surveys in high-resolution optical remote sensing images. The Q-ASFE module of this invention can be directly used as a feature enhancement component in all these scenarios.

[0085] Application 2: General small target detection scenarios The asymmetric spatial perception and QKV semantic modulation collaborative mechanism proposed in this invention is not limited to remote sensing scenarios, but can also be extended to tasks such as the detection of small lesions in medical images (e.g., early nodules in CT / MRI), the identification of small defects in industrial quality inspection (e.g., scratches on chip surfaces, solder joint defects on printed circuit boards), and the detection of pedestrians and non-motorized vehicles at long distances in traffic monitoring.

[0086] Application 3: Preprocessing module for feature enhancement The output of the Q-ASFE module is an enhanced feature map with the same shape as the input. It can be embedded as an independent feature enhancement preprocessing module into the feature extraction backbone network of other visual tasks such as image classification, semantic segmentation, and instance segmentation. It can improve the feature expression quality of the backbone network in scenarios with rich directional texture features or sparse semantics of small objects.

[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. A query key-value-guided asymmetric feature enhancement system, characterized in that: The asymmetric feature enhancement system is an improvement on YOLOv11. The asymmetric feature enhancement system includes a backbone network, a neck network, and a detection head, wherein: In the backbone network, Q-ASFE modules are embedded in the link from C3K2 module to CBS module to transmit features. The P2 layer feature map and P3 layer feature map output by the backbone network are processed by the neck network to output three feature fusion maps of different scales. The three feature fusion maps of different scales are processed by the detection head to obtain the final detection result. The Q-ASFE module consists of an asymmetric spatial feature extraction submodule, an efficient QKV semantic modulation submodule, and a joint fusion output submodule. The asymmetric spatial feature extraction submodule captures spatial features of three different categories and concatenates these features to obtain a feature mosaic map. The efficient QKV semantic modulation submodule, based on the Query-Key-Value mechanism, generates a spatial attention-gated weight map guided by global semantic vectors, and generates a QKV enhanced feature map based on the spatial attention-gated weight map. The joint fusion output submodule concatenates the feature mosaic map and the QKV enhanced feature map to obtain the final enhanced feature map.

2. The query key-guided asymmetric feature enhancement system according to claim 1, characterized in that: The processing steps of the asymmetric spatial feature extraction submodule include: The input feature map is compressed through a 1×1 convolutional layer to obtain a low-channel feature map. The low-channel-number feature map is input into the standard isotropic branch, the lateral asymmetric branch, and the longitudinal asymmetric branch for processing, respectively, to obtain local spatial features, edge and texture features, and detail features. Local spatial features, edge and texture features, and detail features are stitched together to obtain a three-way feature stitched image.

3. The query key-guided asymmetric feature enhancement system according to claim 2, characterized in that: The processing procedure for standard isotropic branches is as follows: ； in, For local spatial features, This is a combination of 3×3 convolution, batch normalization, and activation functions. This is a combination of 1×1 convolution, batch normalization, and activation functions. Let C be the input feature map, H be the channel dimension of the input feature map, W be the height of the input feature map, and R be a real number. The process for handling lateral asymmetric branches is as follows: ； in, For edge and texture features, This is a combination of 1×3 convolution, batch normalization, and activation function operations, where d is the dilation rate of the dilated convolution. This is a combination of 3×1 convolution, batch normalization, and activation functions; The process for handling vertical asymmetric branches is as follows: ； in, For detailed features.

4. The query key-guided asymmetric feature enhancement system according to claim 1, characterized in that: The calculation formula used to stitch together local spatial features, edge and texture features, and detail features to obtain a three-way feature stitched image is as follows: ； in, This is a spliced ​​image of three features. For local spatial features, For edge and texture features, For detailed features, C is the channel dimension of the input feature map, H is the height of the input feature map, W is the width of the input feature map, and R is a real number.

5. The query key-value-guided asymmetric feature enhancement system according to claim 1, characterized in that: The processing steps of the high-efficiency QKV semantic modulation submodule include: Global semantic compression is performed on the input feature map to generate the query vector: ； in, For the query vector, For activation function, For 1×1 convolution, ( ) represents global average pooling, and X is the input feature map. R is the number of channels, and R is a real number; Local spatial awareness is applied to the input feature map to generate a key feature map: ； Where C is the channel dimension of the input feature map, H is the height of the input feature map, and W is the width of the input feature map. For the key feature map, For the number of channels, ( ) represents a 5×5 depthwise separable convolution. This is a combination of 1×1 convolution, batch normalization, and activation functions; Convolutional mapping is performed on the input features to generate the Value feature map: ； Where V is the Value feature map; Attention gating is calculated based on the query vector, key feature map, and value feature map to obtain a spatial attention gating weight map. ； in, For spatial attention gating weights, For element-wise multiplication; Introducing a learnable scaling factor The Value feature map is modulated to obtain the QKV enhanced feature map. : ； ； in, For the expanded spatial attention gating weight map, () represents an extension operation.

6. The query key-guided asymmetric feature enhancement system according to claim 1, characterized in that: The joint fusion output submodule concatenates the feature stitched map and the QKV enhanced feature map to obtain the final enhanced feature map: ； in, For the final enhanced feature map, For activation function, t represents feature concatenation. For 1×1 convolution, This is a spliced ​​image of three features. Enhance the QKV feature map. This is the residual shortcut mapping function.

7. A query key-value-guided asymmetric feature enhancement method, characterized in that: Specifically, the steps include the following: S1: Acquire the original remote sensing images, preprocess the original remote sensing images, and obtain the training set; S2: Train the query key-value guided asymmetric feature enhancement system as described in any one of claims 1-6 using the training set to obtain a trained query key-value guided asymmetric feature enhancement model; S3: Input the remote sensing image to be processed into the trained query key-value-guided asymmetric feature enhancement model to obtain the prediction result.

Images