Bidirectional feature fusion and sparse dynamic enhancement method for small target detection of unmanned aerial vehicle and target detection system

Abstract

The invention discloses a bidirectional feature fusion and sparse dynamic enhancement method for unmanned aerial vehicle small target detection and a target detection system, and belongs to the technical field of target detection. The method comprises the steps of inputting preprocessed data into a backbone network Backbone module, extracting multi-level features of an input image, and obtaining a multi-scale feature map; transmitting the multi-scale feature map to a neck network for feature processing and fusion; and inputting the fused feature map into a detection head, and completing target classification, bounding box regression and confidence coefficient prediction. Partial standard convolution is replaced by wavelet transform through the C3WT module, down-sampling and filtering of features are realized in a frequency domain, and the calculation complexity is reduced; a dynamic gating mechanism introduced by the SDGE module can adaptively activate important neurons and sparise unimportant connections according to input features, so that'on-demand allocation 'computing resource scheduling is realized, and the model parameter quantity is greatly reduced on the premise that the BiSD-YOLO keeps strong feature extraction capability.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a bidirectional feature fusion and sparse dynamic enhancement method and target detection system for small target detection of UAVs. Background Technology

[0002] In recent years, with the accelerated advancement of industrial intelligence and the iterative upgrading of sensor technology, small object detection technology based on UAV platforms has gradually become a key technical support for fields such as smart city management and natural geological exploration.

[0003] Traditional object detection methods rely heavily on hand-crafted feature engineering and conventional classification methods, which fundamentally limits detection accuracy and practical applicability across various scenarios. The paradigm shift driven by advancements in deep learning, through data-driven learning, has endowed modern detection systems with sophisticated feature extraction and spatial localization capabilities. Contemporary detection architectures are generally divided into two distinct frameworks: two-stage detectors, such as Fast R-CNN and Faster R-CNN, and single-stage detectors, such as YOLO, SSD, CenterNet, and DETR.

[0004] It is worth noting that, compared to two-stage detectors, single-stage detectors exhibit superior inference speed while maintaining competitive accuracy, making them increasingly favored in industrial applications and edge computing scenarios requiring real-time processing. As a pioneering single-stage detection framework, the YOLO family of methods has undergone significant development since its inception, continuously iterating to achieve state-of-the-art performance and being widely adopted in real-time applications across various fields.

[0005] However, this field still faces two major problems: First, smaller than Pixel-based objects, limited by low-resolution imaging conditions, suffer from inherent defects such as lack of pixel information, blurred morphological features, and degraded edge textures, easily leading to loss of semantic information and aliasing of spatial features. Secondly, existing deep neural networks generally suffer from numerous parameters and high computational complexity, making it difficult to meet the requirements of edge computing devices for lightweight models and real-time inference. This significantly limits the engineering application of this technology in time-sensitive scenarios.

[0006] Previous studies have shown that small object detection performance is highly dependent on low-level features in shallow feature maps. However, as the network deepens, the rich spatial information and detailed features of small objects are significantly attenuated during hierarchical transmission, while the target representation region in deep features is gradually contaminated by background noise, leading to decreased classification confidence and shifted localization accuracy. Therefore, the key to improving small object detection accuracy lies in feature enhancement and information flow optimization, which can effectively improve the representation ability of small objects and enhance the fusion of shallow detailed features and deep semantic features.

[0007] So, how to solve the problems of insufficient pixels for small targets in images, redundant background information, and excessive model parameters, thereby enhancing the fusion of shallow detail features and deep semantic features in target detection, is the technical problem that this invention aims to solve. Summary of the Invention

[0008] The purpose of this invention is to provide a bidirectional feature fusion and sparse dynamic enhancement method and target detection system for small target detection in unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background art.

[0009] The objective of this invention is achieved as follows: a bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles (UAVs), characterized by the following steps:

[0010] Step S1: Acquire drone aerial photography data and perform preprocessing;

[0011] Step S2: Input the preprocessed data into the backbone network module to extract multi-level features of the input image and obtain multi-scale feature maps;

[0012] Step S3: Pass the multi-scale feature maps to the neck network for feature processing and fusion;

[0013] Step S4: Input the fused feature map into the detection head to complete target classification, bounding box regression and confidence prediction.

[0014] Preferably, the Backbone module includes a CBS module, a C3k2 module, an SPPF module, and a C3WT module. The CBS module consists of a 3×3 convolution with a stride of 2, batch normalization, and a SiLU activation function to achieve downsampling of the feature map.

[0015] The C3k2 module combines a multi-branch convolutional structure and a C3k sub-module with a shortcut connection mechanism. The input feature map is first adjusted to twice the number of hidden channels by a 3×3 standard convolution. Then, it is divided into two parts along the channel dimension. The second half is taken and processed by the C3k sub-module for feature extraction. Finally, the first half is concatenated and output by a 3×3 standard convolution.

[0016] The C3k submodule consists of a single bottleneck block and residual links. The bottleneck block consists of two convolutions with a kernel size of 3×3 and a stride of 1. The convolutions first halve the feature map channels and then double them, and finally add the original feature map to achieve residual links.

[0017] The SPPF module first uses a 1×1 convolution to halve the number of channels in the input feature map, then concatenates the outputs of three consecutive pooling operations with a pooling kernel size of 5×5, a stride of 1, and a padding factor of 2, and finally uses a 1×1 convolution to adjust the number of output channels.

[0018] The C3WT module adopts a parallel multi-branch architecture and extracts features through multi-order wavelet transform by introducing WTConv wavelet convolution.

[0019] The CBS modules are a first CBS module, a second CBS module, a third CBS module, and a fourth CBS module. The fourth CBS module includes a first C3k2 module, a second C3k2 module, and a third C3k2 module.

[0020] The Backbone module is formed by sequentially connecting the first CBS module, the second CBS module, the first C3k2 module, the third CBS module, the second C3k2 module, the fourth CBS module, the third C3k2 module, the SPPF module, and the C3WT module.

[0021] Preferably, in step S2, the preprocessed data is input into the backbone network module to extract multi-level features of the input image and obtain multi-scale feature maps, specifically:

[0022] Step S2-1: Input the 640×640 preprocessed image into the first CBS module. After downsampling, output a feature map P1 with a spatial resolution of 320×320, which is used to retain the rich texture and edge information in the original image and reduce the amount of subsequent calculations.

[0023] Step S2-2: Input feature map P1 into the second CBS module, and then into the first C3k2 module to complete the second downsampling and multi-branch feature enhancement, and output a mid-level feature map P2 with a spatial resolution of 160×160, which has both local detail expression ability and preliminary semantic abstraction ability.

[0024] Steps S2-3: The middle layer feature map P2 is input into the third CBS module, and then into the second C3k2 module, outputting a deep feature map P3 with a spatial resolution of 80×80, which serves as one of the main semantic sources for cross-scale fusion of the neck network;

[0025] Steps S2-4: The deep feature map P3 is input into the fourth CBS module, then into the third C3k2 module for further downsampling, and then into the SPPF module and C3WT module in sequence. The final output is a high-level feature map P4 with a spatial resolution of 40×40. The SPPF module is used to expand the receptive field and aggregate multi-scale context. The C3WT module uses learnable wavelet transform convolution WTConv instead of standard convolution to enhance the sensitivity to high-frequency textures and weak-contrast small targets.

[0026] Step S2-5: Use the middle layer feature map P2, the deep layer feature map P3, and the high layer feature map P4 as input feature maps of the BiCross structure in the neck network to perform bidirectional feature fusion, thereby achieving complementary enhancement of cross-scale features and noise suppression.

[0027] Preferably, the neck network adopts a bidirectional feature pyramid structure, including a BiCross structure, a channel reweighting splicing mechanism, a C3WT module, and a sparse dynamic gating enhancement module SDGE. The BiCross structure includes cross-resolution connections and cross-stage skip connections; it is composed of a top-down semantic guidance path and a bottom-up detail enhancement path, and achieves semantic consistency and spatial accuracy optimization of the feature pyramid through cross-resolution feature alignment and cross-stage skip connections.

[0028] Cross-stage skip connections establish short-circuit paths between different stages of the network, facilitating the direct transmission of deep semantic information and shallow detailed features.

[0029] The channel reweighting concatenation mechanism designs a learnable weight for each channel and outputs features. The expression is:

[0030] ;

[0031] in, These are the feature maps at three different resolutions after stitching. It is the output feature map; Represents the learnable weights of the channel. It is a parameter; This represents channel-by-channel multiplication; This represents the weight corresponding to each feature map; Indicates the number of feature channels.

[0032] Preferably, the C3WT module is used to replace the original C2PSA structure and is deployed at the end or neck of the backbone network.

[0033] The C3WT module adopts a multi-branch parallel architecture, and the processing flow is as follows:

[0034] Input feature map Composed of two parallel The convolutional blocks undergo batch normalization and SiLU activation to generate two feature sets. and :

[0035] ;

[0036] in, , It is the output of two CBS blocks;

[0037] It is sent to the WTConv module and performs the following operations:

[0038] Input feature mapping is performed using Haar-based two-dimensional discrete wavelet transform. It is decomposed into four subbands: LL, LH, HL, and HH. Convolutional layers are applied to each subband and the original feature map;

[0039] The processed subbands are reconstructed to their original spatial size using inverse wavelet transform and then added to the output of the convolutional layer of the original feature map.

[0040] Output , This indicates the output of the WTConv block;

[0041] The WTConv block allows for multi-level wavelet transforms, enabling the input feature map to be decomposed into multiple levels of subbands;

[0042] Output of all branches , and The outputs are concatenated along the channel dimension and passed through the final... The convolutional blocks are fused to produce the output feature map O:

[0043] ;

[0044] in, This indicates splicing along the channel dimension.

[0045] Preferably, the sparse dynamic gating enhancement module SDGE is used to enhance the interaction capability between global and local information and is deployed at the end of the neck network or the front end of the detection head;

[0046] The Sparse Dynamic Gated Enhancement Module (SDGE) includes two parallel branches: a sparse dynamic enhancement branch and a fast branch. The sparse dynamic enhancement branch is used to sparsely extract multi-scale features and perform dynamic filtering. The sparse dynamic enhancement branch applies gated adaptive enhancement to improve feature quality.

[0047] The shortcut branch is used to process the raw input and match the channel size;

[0048] The calculation process for the Sparse Dynamic Gated Enhancement Module (SDGE) is as follows:

[0049] First, the input is processed by point-to-point convolution. Halving the number of channels reduces computational cost; the output is named... ;

[0050] To sparsely capture features at multiple scales, the feature maps are processed by three depth-first convolutional layers with different kernel sizes and different dilation rates, producing multi-scale feature maps. :

[0051] ;

[0052] Multi-scale feature maps The concatenation and fusion are performed using conditional convolution PWCondConv, with the kernel size set to [value missing]. Then, after batch normalization and SiLU activation:

[0053] ;

[0054] at the same time, Add the output of the shortcut branch to the residual join. To facilitate gradient flow and preserve original information, the first enhanced feature map is generated. ;

[0055] Dual-pooling spatial gate DPSG is used to enhance the feature map. The operation is as follows:

[0056] The dual-pooling spatial gate (DPSG) concatenates the outputs of max pooling and average pooling along the channel dimension, processes them with a convolutional layer, activates them with SiLU, and generates a spatial attention mask:

[0057] ;

[0058] in, for The number of channels in;

[0059] The outputs of the two branches are fused using residual gates, and the resulting feature map is as follows. The calculation is as follows:

[0060] ;

[0061] in, It is a balanced enhancement branch and original input Learnable weights.

[0062] Preferably, in step S3, the multi-scale feature maps are passed to the neck network for feature processing and fusion, specifically as follows:

[0063] Step S3-1: Receive the mid-level feature map P2, deep feature map P3, and high-level feature map P4 output from the Backbone module. Use the mid-level feature map P2, deep feature map P3, and high-level feature map P4 as initial input nodes, and input the bottom-up detail enhancement path and the top-down semantic guidance path to achieve multi-level interaction and fusion of cross-scale features.

[0064] Step S3-2: In the top-down semantic guidance path, upsample the high-level feature map P4 output by the backbone network to obtain a feature map P4 with a resolution of 80×80.

[0065] The mid-layer feature map P2 is upsampled by the BWCBS depth-first convolution module to obtain a feature map P2 with a resolution of 80×80.

[0066] P4, P2 and the original mid-level feature map P3 are input into the CRC channel reweighting and splicing mechanism to achieve cross-scale semantic distribution alignment. Then, they are input into the C3k2 module for feature fusion and output the enhanced mid-level feature map P3′, which is used to carry the guidance information of high-level semantics for mid-level localization.

[0067] Upsample P3′ to 160×160 and concatenate it with the original mid-layer feature map P2. Input the concatenation into the SDGE module to generate a sparse dynamic enhancement feature map P2′.

[0068] Step S3-3: In the bottom-up detail enhancement path, P2′ is sampled to 80×80 using the CBS module to obtain P2′′;

[0069] P2′′, P3′ and the original P3 are input into the CRC module again for secondary alignment, and then input into the C3k2 module for fusion, outputting the enhanced mid-layer feature map P3″;

[0070] P3″ is downsampled to 40×40 by the CBS module and concatenated with the original high-level feature map P4 output by the backbone network. Then it is fused by the C3k2 module to output the final high-level feature map P4′, completing the closed-loop feedback from low-level details to mid-level structure to high-level semantics.

[0071] Preferably, the detection head includes a decoupled head structure, a lightweight regression branch, and a dynamic loss allocation mechanism. The decoupled head structure separates the classification task from the localization task, extracts semantic features through independent neural network branches, trains the model using corresponding classification and regression loss functions, and finally obtains the total loss by weighted summation of various losses, thus eliminating interference between tasks. The lightweight regression branch is constructed using depthwise separable convolution, significantly reducing computational overhead and improving inference speed. The dynamic loss allocation mechanism adaptively adjusts the weight ratio of classification loss and localization loss according to the difficulty of the samples, optimizes the model convergence direction, and improves the detection sensitivity for small targets.

[0072] A target detection system for small target detection in unmanned aerial vehicles, characterized in that:

[0073] The object detection system BiSD-YOLO includes a backbone network, a neck network, and a detection head. The backbone network is used to extract multi-level features from the input image; the neck network is used to fuse feature maps of different scales to achieve cross-layer and cross-resolution information interaction; and the detection head is used to complete object classification, bounding box regression, and confidence prediction.

[0074] Compared with the prior art, the present invention has the following improvements and advantages:

[0075] 1. By using the C3WT module to replace part of the standard convolution with wavelet transform, feature downsampling and filtering are achieved in the frequency domain. Its inherent sparsity greatly reduces the computational complexity. At the same time, the dynamic gating mechanism introduced by the SDGE module can adaptively activate important neurons and sparsify unimportant connections according to the input features, avoiding the fixed computation mode of traditional convolutional layers and realizing "on-demand" allocation of computational resources. This allows BiSD-YOLO to significantly reduce the number of model parameters while maintaining its powerful feature extraction capabilities.

[0076] 2. The BiSD-YOLO model designs a more streamlined cross-resolution interaction mechanism, which significantly reduces the number of channels and the amount of intermediate feature maps generated while ensuring effective fusion of multi-scale features. In practical applications, this high efficiency translates into faster detection frame rates and lower energy consumption, providing a reliable and efficient solution for scenarios such as real-time aerial video analysis and mobile intelligent perception. Attached Figure Description

[0077] Figure 1 This is a schematic diagram of the BiSD-YOLO structure of the present invention.

[0078] Figure 2 This is a schematic diagram of the BiCross structure.

[0079] Figure 3 This is a schematic diagram of the C3WT structure.

[0080] Figure 4 This is a schematic diagram of a dual-pooled space gate (DPSG) structure.

[0081] Figure 5 This is a schematic diagram of the sparse dynamic gating enhancement module SDGE.

[0082] Figure 6 This is a schematic diagram showing the comparison results on the TinyPerson dataset.

[0083] Figure 7This is a schematic diagram illustrating the detection results for images in the VisDrone2019 dataset.

[0084] Figure 8 This is a diagram illustrating the comparison results between the various models. Detailed Implementation

[0085] The invention will be further summarized below with reference to the accompanying drawings.

[0086] A bidirectional feature fusion and sparse dynamic enhancement method for small target detection in UAVs, comprising the following steps:

[0087] Step S1 involves acquiring drone aerial photography data and performing preprocessing, specifically as follows:

[0088] The VisDrone2019 dataset uses COCO format annotation and stores annotation information in JSON. The annotation information includes object category, top-left corner coordinates of bounding box, and width and height of bounding box. The YOLO method uses category, normalized bounding box center coordinates, and width and height of bounding box.

[0089] Raw image data typically has high resolution and rectangular dimensions, making direct input computationally intensive and unsuitable for convolution calculations. Therefore, images need to be scaled, cropped, and padded before input. In the training of the proposed method, images are resized to 640*640 pixels and processed using mosaic operations.

[0090] In step S2, the preprocessed data is input into the backbone network module to extract multi-level features from the input image and obtain multi-scale feature maps, specifically:

[0091] Step S2-1: Input the 640×640 preprocessed image into the first CBS module. After downsampling, output a feature map P1 with a spatial resolution of 320×320, which is used to retain the rich texture and edge information in the original image and reduce the amount of subsequent calculations.

[0092] Step S2-2: Input feature map P1 into the second CBS module, and then into the first C3k2 module to complete the second downsampling and multi-branch feature enhancement, and output a mid-level feature map P2 with a spatial resolution of 160×160, which has both local detail expression ability and preliminary semantic abstraction ability.

[0093] Steps S2-3: The middle layer feature map P2 is input into the third CBS module, and then into the second C3k2 module, outputting a deep feature map P3 with a spatial resolution of 80×80, which serves as one of the main semantic sources for cross-scale fusion of the neck network;

[0094] Steps S2-4: The deep feature map P3 is input into the fourth CBS module, then into the third C3k2 module for further downsampling, and then into the SPPF module and C3WT module in sequence. The final output is a high-level feature map P4 with a spatial resolution of 40×40. The SPPF module is used to expand the receptive field and aggregate multi-scale context. The C3WT module uses learnable wavelet transform convolution WTConv instead of standard convolution to enhance the sensitivity to high-frequency textures and weak-contrast small targets.

[0095] Step S2-5: Use the middle layer feature map P2, the deep layer feature map P3, and the high layer feature map P4 as input feature maps of the BiCross structure in the neck network to perform bidirectional feature fusion, thereby achieving complementary enhancement of cross-scale features and noise suppression.

[0096] To balance the need for lightweight model with the accuracy of small target localization, this method actively omits the fifth-level downsampling operation in the YOLO11 network, which is located before the SPPF module. This keeps the spatial resolution of P4 at 1 / 16 of the input image (instead of the usual 1 / 32), thereby significantly reducing computational overhead and effectively alleviating the problem of severe loss of spatial information of small targets in deep features due to oversampling.

[0097] In step S3, the multi-scale feature maps are passed to the neck network for feature processing and fusion, specifically as follows:

[0098] Step S3-1: Receive the mid-level feature map P2, deep feature map P3, and high-level feature map P4 output from the Backbone module. Use the mid-level feature map P2, deep feature map P3, and high-level feature map P4 as initial input nodes, and input the bottom-up detail enhancement path and the top-down semantic guidance path to achieve multi-level interaction and fusion of cross-scale features.

[0099] Step S3-2: In the top-down semantic guidance path, the high-level feature map P4 (40×40) output by the backbone network is upsampled to obtain a feature map P4 with a resolution of 80×80.

[0100] The mid-layer feature map P2 (160×160) is upsampled by the BWCBS depth-first convolution module (which includes 1×1 convolution with stride of 1, batch normalization, and SiLU activation operation) to obtain a feature map P2 with a resolution of 80×80.

[0101] P4, P2 and the original mid-level feature map P3 (80×80) are input into the CRC channel reweighting and splicing mechanism to achieve cross-scale semantic distribution alignment. Then, they are input into the C3k2 module for feature fusion and output the enhanced mid-level feature map P3′ (80×80), which is used to carry the guidance information of high-level semantics for mid-level localization.

[0102] Upsample P3′ to 160×160 and concatenate it with the original P2 (160×160). Input the concatenation into the SDGE module to generate a sparse dynamic enhancement feature map P2′ (160×160).

[0103] Step S3-3: Downsample P2′ to 80×80 using the CBS module (3×3 convolution with stride of 2) to obtain P2′′;

[0104] P2′′ and P3′ (80×80) are then input into the CRC module for secondary alignment with the original P3 (80×80), and then input into the C3k2 module for fusion, outputting the enhanced mid-layer feature map P3″ (80×80).

[0105] P3″ is downsampled to 40×40 by the CBS module and concatenated with P4 (40×40) output by the backbone network. Then it is fused by the C3k2 module to output the final high-level feature map P4′ (40×40), completing the closed-loop feedback from low-level details to mid-level structure to high-level semantics.

[0106] A bidirectional feature pyramid network based on a BiCross structure is constructed, which receives multi-scale feature maps P2, P3, and P4 output from the Backbone module, corresponding to different spatial resolutions and semantic levels, respectively. Among them, P2 and P3 are high-resolution low-level features, and P4 is a low-resolution high-level semantic feature. The above feature maps are used as initial input nodes to construct bidirectional information flow paths from bottom to top and from top to bottom, so as to realize multi-level interaction and fusion of cross-scale features.

[0107] Multi-round feature fusion operation: In the top-down path, the P4 feature map is upsampled, and the P2 feature map is upsampled using the BWCBS depth-first convolution module. The two are semantically aligned with the P3 feature map using the CRC module, and the CRC weights are automatically optimized using learnable parameters. Then, the features are fused using the C3K2 module to enhance the guidance of high-level semantics on mid-level features. Subsequently, the fused P3' feature map is upsampled again and combined with the P2 feature map to generate an enhanced high-resolution feature map P2' using SDGE. In the bottom-up path, the fused P2' feature map is downsampled using the CBS module and then fused with the P3' feature map and the original P3 feature map using CRC alignment and C3K2 to further enhance the detailed expression ability of mid-level features. Finally, the output feature map is downsampled using CBS and concatenated with the P5 feature map generated and extracted by the backbone network and fused again using C3K2 to form a closed-loop bidirectional feature transfer mechanism, which improves the semantic consistency and spatial accuracy of the feature pyramid.

[0108] Step S4: Input the fused feature map into the detection head to complete target classification, bounding box regression and confidence prediction.

[0109] like Figure 1 As shown, a target detection system for small target detection in UAVs, the BiSD-YOLO target detection system includes a backbone network, a neck network, and a detection head. The backbone network is used to extract multi-level features of the input image; the neck network is used to fuse feature maps of different scales to achieve cross-layer and cross-resolution information interaction; and the detection head is used to complete target classification, bounding box regression, and confidence prediction.

[0110] The backbone network uses a Backbone module, which includes a CBS module, a C3k2 module, an SPPF module, and a C3WT module. The CBS module consists of a 3×3 convolution with a stride of 2, batch normalization, and a SiLU activation function to achieve downsampling of the feature map.

[0111] The C3k2 module combines a multi-branch convolutional structure and the C3k sub-module with a shortcut connection mechanism. The input feature map is first adjusted to twice the number of hidden channels by a 3×3 standard convolution. Then, it is divided into two parts along the channel dimension. The second half is taken and processed by the C3k sub-module for feature extraction. Finally, the first half is concatenated and output by a 3×3 standard convolution.

[0112] The C3k submodule consists of a single bottleneck block and residual connections. The bottleneck block consists of two convolutions with a kernel size of 3×3 and a stride of 1. The convolutions halve the feature map channels and then double them, and finally add the original feature map to achieve residual connections.

[0113] The SPPF module first uses a 1×1 convolution to halve the number of channels in the input feature map, then concatenates the outputs of three consecutive pooling operations with a pooling kernel size of 5×5, a stride of 1, and a padding factor of 2. Finally, it uses a 1×1 convolution to adjust the number of output channels.

[0114] The C3WT module adopts a parallel multi-branch architecture, similar in structure to C3k. It extracts features through multi-order wavelet transform by introducing WTConv wavelet convolution.

[0115] The C3WT module achieves a balance between a large receptive field and detail preservation through frequency domain decomposition and reconstruction mechanisms. It captures a wide range of contexts while avoiding the loss of local features, making it particularly suitable for small target detection in complex and dense scenes.

[0116] The Backbone module is formed by sequentially connecting the first CBS module, the second CBS module, the first C3k2 module, the third CBS module, the second C3k2 module, the fourth CBS module, the third C3k2 module, the SPPF module, and the C3WT module.

[0117] like Figure 2As shown, the neck network adopts a bidirectional feature pyramid structure, including a BiCross structure, a channel reweighting concatenation mechanism, a C3WT module, and a sparse dynamic gating enhancement module SDGE. The BiCross structure includes cross-resolution connections and cross-stage skip connections; it consists of a top-down semantic guidance path and a bottom-up detail enhancement path, and achieves semantic consistency and spatial accuracy optimization of the feature pyramid through cross-resolution feature alignment and cross-stage skip connections.

[0118] Cross-stage skip connections establish short-circuit paths between different stages of the network, facilitating the direct transmission of deep semantic information and shallow detailed features.

[0119] The channel reweighting concatenation mechanism designs a learnable weight for each channel and outputs features. The expression is:

[0120] ;

[0121] in, These are the feature maps at three different resolutions after stitching. It is the output feature map; Represents the learnable weights of the channel. It is a parameter; This represents channel-by-channel multiplication; This represents the weight corresponding to each feature map; Indicates the number of feature channels.

[0122] like Figure 3 As shown, the C3WT module is used to replace the original C2PSA structure and is deployed at the end or neck of the backbone network.

[0123] The C3WT module adopts a multi-branch parallel architecture, and the processing flow is as follows:

[0124] Input feature map Composed of two parallel The convolutional blocks undergo batch normalization and SiLU activation to generate two feature sets. and :

[0125] ;

[0126] in, , It is the output of two CBS blocks;

[0127] It is sent to the WTConv module and performs the following operations:

[0128] Input feature mapping is performed using Haar-based two-dimensional discrete wavelet transform. It is decomposed into four subbands: LL, LH, HL, and HH. Convolutional layers are applied to each subband and the original feature map;

[0129] The processed subbands are reconstructed to their original spatial size using inverse wavelet transform and then added to the output of the convolutional layer of the original feature map.

[0130] Output , This indicates the output of the WTConv block;

[0131] The WTConv block allows for multi-level wavelet transforms, enabling the input feature map to be decomposed into multiple levels of subbands;

[0132] Output of all branches , and The outputs are concatenated along the channel dimension and passed through the final... The convolutional blocks are fused to produce the output feature map O:

[0133] ;

[0134] in, This indicates splicing along the channel dimension.

[0135] like Figure 5 As shown, the Sparse Dynamic Gating Enhancement Module (SDGE) is used to enhance the interaction capability between global and local information and is deployed at the end of the neck network or the front end of the detection head.

[0136] The Sparse Dynamic Gated Enhancement (SDGE) module comprises two parallel branches: a sparse dynamic enhancement branch and a shortcut branch. The sparse dynamic enhancement branch is used to sparsely extract multi-scale features and perform dynamic filtering. The sparse dynamic enhancement branch applies gating to adaptively enhance feature quality. The shortcut branch is used to process the original input and match the channel size.

[0137] The calculation process for the Sparse Dynamic Gated Enhancement Module (SDGE) is as follows:

[0138] First, the input is processed by point-to-point convolution. Halving the number of channels reduces computational cost; the output is named... ;

[0139] To sparsely capture features at multiple scales, the feature maps are processed by three depth-first convolutional layers with different kernel sizes and different dilation rates, producing multi-scale feature maps. :

[0140] ;

[0141] Multi-scale feature maps The concatenation and fusion are performed using conditional convolution PWCondConv, with the kernel size set to [value missing]. Then, after batch normalization and SiLU activation:

[0142] ;

[0143] at the same time, Add the output of the shortcut branch to the residual join. To facilitate gradient flow and preserve original information, the first enhanced feature map is generated. ;

[0144] like Figure 4 As shown, dual-pooling spatial gate DPSG is used to enhance the feature map. The operation is as follows:

[0145] The dual-pooling spatial gate (DPSG) concatenates the outputs of max pooling and average pooling along the channel dimension, processes them with a convolutional layer, activates them with SiLU, and generates a spatial attention mask:

[0146] ;

[0147] in, for The number of channels in;

[0148] The outputs of the two branches are fused using residual gates, and the resulting feature map is as follows. The calculation is as follows:

[0149] ;

[0150] in, It is a balanced enhancement branch and original input Learnable weights.

[0151] The detection head includes a decoupled head structure, a lightweight regression branch, and a dynamic loss allocation mechanism. The decoupled head structure separates the classification and localization tasks, extracting semantic features through independent neural network branches, training them using corresponding classification and regression loss functions, and finally obtaining the total loss by weighted summation of various losses, thus eliminating interference between tasks. The lightweight regression branch is constructed using depthwise separable convolutions, significantly reducing computational overhead and improving inference speed. The dynamic loss allocation mechanism adaptively adjusts the weight ratio of classification and localization losses according to the difficulty of the samples, optimizing the model's convergence direction and improving the detection sensitivity for small targets.

[0152] The feasibility and effectiveness of this invention are verified through experiments as follows:

[0153] Table 1 shows the results of the model on the TinyPerson dataset.

[0154]

[0155] Table 2 shows the results of the model on the VisDrone2019 dataset.

[0156]

[0157] Table 3 shows the ablation experimental results for each module of the model.

[0158]

[0159] As shown in Tables 1-3, experiments on the highly challenging aerial small target detection datasets VisDrone2019 and TinyPerson demonstrate that, compared to the baseline model YOLO11n, BiSD-YOLO achieves a significant improvement in mean accuracy (mAP) while maintaining a substantial 58% reduction in parameters. This proves that the model achieves an excellent balance between accuracy and efficiency. The significant reduction in parameters means that the model requires less memory and storage resources, allowing it to be easily deployed on edge devices with limited computing power, memory, and power consumption (such as embedded platforms, UAV onboard computers, and mobile devices). In practical applications, this high efficiency translates to faster detection frame rates and lower energy consumption, providing a reliable and efficient solution for scenarios such as real-time aerial video analysis and mobile intelligent sensing.

[0160] from Figure 6 , 7 As can be seen, the BiSD-YOLO model achieves high detection accuracy with extremely low parameter scale, demonstrating excellent parameter efficiency and performance advantages.

[0161] from Figure 8 As can be seen from the above, the BiSD-YOLO model proposed in this invention application (marked with a red pentagram) has only about 2 × 10 parameters. 6 At that time, the mAP@0.5 metric reached 0.193, significantly outperforming existing comparative models: YOLOv5 (0.149), YOLOv6 (0.127), YOLOv8 (0.154), and YOLOv11 (0.154, with over 9×10^6 parameters). 6As can be seen, the BiSD-YOLO model achieves high detection accuracy with an extremely low parameter scale, demonstrating excellent parameter efficiency and performance advantages. Its mAP gain per unit parameter is significantly higher than existing methods, indicating that the BiSD-YOLO model structure has substantial improvements in feature extraction capability, information compression efficiency, and inference robustness. These results validate the outstanding advantages of the BiSD-YOLO model in balancing lightweight design and high accuracy, making it particularly suitable for real-time target detection tasks in resource-constrained environments.

[0162] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles, characterized in that: The method includes the following steps: Step S1: Acquire drone aerial photography data and perform preprocessing; Step S2: Input the preprocessed data into the backbone network module to extract multi-level features of the input image and obtain multi-scale feature maps; Step S3: Pass the multi-scale feature maps to the neck network for feature processing and fusion; Step S4: Input the fused feature map into the detection head to complete target classification, bounding box regression and confidence prediction.

2. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in UAVs according to claim 1, characterized in that: The Backbone module includes a CBS module, a C3k2 module, an SPPF module, and a C3WT module. The CBS module consists of a 3×3 convolution with a stride of 2, batch normalization, and a SiLU activation function to achieve downsampling of the feature map. The C3k2 module combines a multi-branch convolutional structure and a C3k sub-module with a shortcut connection mechanism. The input feature map is first adjusted to twice the number of hidden channels by a 3×3 standard convolution. Then, it is divided into two parts along the channel dimension. The second half is taken and processed by the C3k sub-module for feature extraction. Finally, the first half is concatenated and output by a 3×3 standard convolution. The C3k submodule consists of a single bottleneck block and residual links. The bottleneck block consists of two convolutions with a kernel size of 3×3 and a stride of 1. The convolutions first halve the feature map channels and then double them, and finally add the original feature map to achieve residual links. The SPPF module first uses a 1×1 convolution to halve the number of channels in the input feature map, then concatenates the outputs of three consecutive pooling operations with a pooling kernel size of 5×5, a stride of 1, and a padding factor of 2, and finally uses a 1×1 convolution to adjust the number of output channels. The C3WT module adopts a parallel multi-branch architecture and extracts features through multi-order wavelet transform by introducing WTConv wavelet convolution. The CBS modules are a first CBS module, a second CBS module, a third CBS module, and a fourth CBS module. The fourth CBS module includes a first C3k2 module, a second C3k2 module, and a third C3k2 module. The Backbone module is formed by sequentially connecting the first CBS module, the second CBS module, the first C3k2 module, the third CBS module, the second C3k2 module, the fourth CBS module, the third C3k2 module, the SPPF module, and the C3WT module.

3. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in UAVs according to claim 2, characterized in that: In step S2, the preprocessed data is input into the backbone network module to extract multi-level features from the input image and obtain multi-scale feature maps, specifically: Step S2-1: Input the 640×640 preprocessed image into the first CBS module. After downsampling, output a feature map P1 with a spatial resolution of 320×320, which is used to retain the rich texture and edge information in the original image and reduce the amount of subsequent calculations. Step S2-2: Input feature map P1 into the second CBS module, and then into the first C3k2 module to complete the second downsampling and multi-branch feature enhancement, and output a mid-level feature map P2 with a spatial resolution of 160×160, which has both local detail expression ability and preliminary semantic abstraction ability. Steps S2-3: The middle layer feature map P2 is input into the third CBS module, and then into the second C3k2 module, outputting a deep feature map P3 with a spatial resolution of 80×80, which serves as one of the main semantic sources for cross-scale fusion of the neck network; Steps S2-4: The deep feature map P3 is input into the fourth CBS module, then into the third C3k2 module for further downsampling, and then into the SPPF module and C3WT module in sequence. The final output is a high-level feature map P4 with a spatial resolution of 40×40. The SPPF module is used to expand the receptive field and aggregate multi-scale context. The C3WT module uses learnable wavelet transform convolution WTConv instead of standard convolution to enhance the sensitivity to high-frequency textures and weak-contrast small targets. Step S2-5: Use the middle layer feature map P2, the deep layer feature map P3, and the high layer feature map P4 as input feature maps of the BiCross structure in the neck network to perform bidirectional feature fusion, thereby achieving complementary enhancement of cross-scale features and noise suppression.

4. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in UAVs according to claim 1, characterized in that: The neck network adopts a bidirectional feature pyramid structure, including a BiCross structure, a channel reweighting concatenation mechanism, a C3WT module, and a sparse dynamic gating enhancement module SDGE. The BiCross structure includes cross-resolution connections and cross-stage skip connections; it consists of a top-down semantic guidance path and a bottom-up detail enhancement path, and achieves semantic consistency and spatial accuracy optimization of the feature pyramid through cross-resolution feature alignment and cross-stage skip connections. Cross-stage skip connections establish short-circuit paths between different stages of the network, facilitating the direct transmission of deep semantic information and shallow detailed features. The channel reweighting concatenation mechanism designs a learnable weight for each channel and outputs features. The expression is: ； in, These are the feature maps at three different resolutions after stitching. It is the output feature map; Represents the learnable weights of the channel. It is a parameter; This represents channel-by-channel multiplication; This represents the weight corresponding to each feature map; Indicates the number of feature channels.

5. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles according to claim 4, characterized in that: The C3WT module is used to replace the original C2PSA structure and is deployed at the end or critical neck position of the backbone network. The C3WT module adopts a multi-branch parallel architecture, and the processing flow is as follows: Input feature map Composed of two parallel The convolutional blocks undergo batch normalization and SiLU activation to generate two feature sets. and : ； in, , It is the output of two CBS blocks; It is sent to the WTConv module and performs the following operations: Input feature mapping is performed using Haar-based two-dimensional discrete wavelet transform. It is decomposed into four subbands: LL, LH, HL, and HH. Convolutional layers are applied to each subband and the original feature map; The processed subbands are reconstructed to their original spatial size using inverse wavelet transform and then added to the output of the convolutional layer of the original feature map. Output , This indicates the output of the WTConv block; The WTConv block allows for multi-level wavelet transforms, enabling the input feature map to be decomposed into multiple levels of subbands; Output of all branches , and The outputs are concatenated along the channel dimension and passed through the final... The convolutional blocks are fused to produce the output feature map O: ； in, This indicates splicing along the channel dimension.

6. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles according to claim 4, characterized in that: The Sparse Dynamic Gating Enhancement Module (SDGE) is used to enhance the interaction capability between global and local information and is deployed at the end of the neck network or the front end of the detection head. The Sparse Dynamic Gated Enhancement Module (SDGE) includes two parallel branches: a sparse dynamic enhancement branch and a shortcut branch. The sparse dynamic enhancement branch is used to sparsely extract multi-scale features and perform dynamic filtering; Sparse dynamic augmentation branching applies gated adaptive enhancement of feature quality; The shortcut branch is used to process the raw input and match the channel size; The calculation process for the Sparse Dynamic Gated Enhancement Module (SDGE) is as follows: First, the input is processed by point-to-point convolution. Halving the number of channels reduces computational cost; the output is named... ; To sparsely capture features at multiple scales, the feature maps are processed by three depth-first convolutional layers with different kernel sizes and different dilation rates, producing multi-scale feature maps. : ； Multi-scale feature maps The concatenation and fusion are performed using conditional convolution PWCondConv, with the kernel size set to [value missing]. Then, after batch normalization and SiLU activation: ； at the same time, Add the output of the shortcut branch to the residual join. To facilitate gradient flow and preserve original information, the first enhanced feature map is generated. ; Dual-pooling spatial gate DPSG is used to enhance the feature map. The operation is as follows: The dual-pooling spatial gate (DPSG) concatenates the outputs of max pooling and average pooling along the channel dimension, processes them with a convolutional layer, activates them with SiLU, and generates a spatial attention mask: ； in, for The number of channels in; The outputs of the two branches are fused using residual gates, and the resulting feature map is as follows. The calculation is as follows: ； in, It is a balanced enhancement branch and original input Learnable weights.

7. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles according to claim 4, characterized in that: In step S3, the multi-scale feature maps are passed to the neck network for feature processing and fusion, specifically as follows: Step S3-1: Receive the mid-level feature map P2, deep feature map P3, and high-level feature map P4 output from the Backbone module. Use the mid-level feature map P2, deep feature map P3, and high-level feature map P4 as initial input nodes, and input the bottom-up detail enhancement path and the top-down semantic guidance path to achieve multi-level interaction and fusion of cross-scale features. Step S3-2: In the top-down semantic guidance path, upsample the high-level feature map P4 output by the backbone network to obtain a feature map P4 with a resolution of 80×80. The mid-layer feature map P2 is upsampled by the BWCBS depth-first convolution module to obtain a feature map P2 with a resolution of 80×80. P4, P2 and the original mid-level feature map P3 are input into the CRC channel reweighting and splicing mechanism to achieve cross-scale semantic distribution alignment. Then, they are input into the C3k2 module for feature fusion and output the enhanced mid-level feature map P3′, which is used to carry the guidance information of high-level semantics for mid-level localization. Upsample P3′ to 160×160 and concatenate it with the original mid-layer feature map P2. Input the concatenation into the SDGE module to generate a sparse dynamic enhancement feature map P2′. Step S3-3: In the bottom-up detail enhancement path, P2′ is sampled to 80×80 using the CBS module to obtain P2′′; P2′′, P3′ and the original P3 are input into the CRC module again for secondary alignment, and then input into the C3k2 module for fusion, outputting the enhanced mid-layer feature map P3″; P3″ is downsampled to 40×40 by the CBS module and concatenated with the original high-level feature map P4 output by the backbone network. Then it is fused by the C3k2 module to output the final high-level feature map P4′, completing the closed-loop feedback from low-level details to mid-level structure to high-level semantics.

8. The bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles according to claim 1, characterized in that: The detection head includes a decoupled head structure, a lightweight regression branch, and a dynamic loss allocation mechanism. The decoupled head structure separates the classification task from the localization task, extracts semantic features through independent neural network branches, trains the model using corresponding classification and regression loss functions, and finally obtains the total loss by weighted summation of various losses, thus eliminating interference between tasks. The lightweight regression branch is constructed using depthwise separable convolution, significantly reducing computational overhead and improving inference speed. The dynamic loss allocation mechanism adaptively adjusts the weight ratio of classification loss and localization loss according to the difficulty of the samples, optimizes the model convergence direction, and improves the detection sensitivity for small targets.

9. A target detection system generated by a bidirectional feature fusion and sparse dynamic enhancement method for small target detection in unmanned aerial vehicles according to claims 1-8, characterized in that: The object detection system BiSD-YOLO includes a backbone network, a neck network, and a detection head. The backbone network is used to extract multi-level features from the input image; the neck network is used to fuse feature maps of different scales to achieve cross-layer and cross-resolution information interaction; and the detection head is used to complete object classification, bounding box regression, and confidence prediction.

Images