Small target detection method and system based on intelligent decoupling network

Abstract

The invention discloses a small target detection method and system based on an intelligent decoupling network, and the method comprises the steps: carrying out the preprocessing operation of an input to-be-detected image, and obtaining standardized image data; inputting the standardized image data into an improved YOLOv10n backbone network to extract features; based on an intelligent decoupling network, the extracted features are decoupled into basic structure features and detail texture features, and the basic structure features and the detail texture features are fused and output as enhanced features; inputting the enhanced features into a path aggregation network, carrying out cross-scale feature deep fusion, and outputting a multi-scale enhanced feature map; inputting the multi-scale enhanced feature map into a detection head, and outputting target frame normalized coordinates, category probability and confidence as detection results through a regression branch and a classification branch; performing post-processing on the detection result; according to the method, the detection precision and recall rate of the small target are remarkably improved, and the method has high robustness and practical value in unmanned aerial vehicle aerial photography, industrial vision and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of computer vision target detection technology, and in particular relates to a method and system for small target detection based on intelligent decoupling networks. Background Technology

[0002] With the rapid development of deep learning technology, object detection algorithms based on convolutional neural networks have achieved remarkable success in many fields such as security monitoring, autonomous driving, medical image analysis, and industrial quality inspection. The YOLO series of algorithms, due to their good balance between accuracy and speed, has become a representative framework in the field of single-stage object detection. Among them, the YOLOv10n model further improves the inference efficiency and accuracy of the basic architecture through structural reparameterization and optimization of label allocation strategies, providing a feasible solution for real-time detection in resource-constrained environments.

[0003] However, despite significant improvements in general object detection performance, small object detection remains a long-standing and highly challenging problem in this field. Small objects typically refer to objects that occupy a small pixel area in an image, have weak appearance features, and lack detailed information, easily leading to missed detections and false detections by existing detectors. In practical applications such as UAV aerial imagery, remote sensing images, and large-scale scene monitoring, the targets to be detected often exist in a small-scale form, posing a severe challenge to the performance of existing detectors. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a small target detection method and system based on intelligent decoupling networks that can improve detection accuracy.

[0005] Technical solution: The present invention provides a small target detection method based on an intelligent decoupling network, comprising:

[0006] S1: Perform preprocessing operations on the input image to be detected to obtain standardized image data;

[0007] S2: Input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module.

[0008] S3: Based on an intelligent decoupling network, the extracted features are decoupled into basic structural features and detailed texture features, and the basic structural features and detailed texture features are fused together to output an enhanced feature; the intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module and an intelligent gating fusion module;

[0009] S4: The enhanced feature input path aggregation network is used to perform cross-scale feature deep fusion and output multi-scale enhanced feature maps.

[0010] S5: Input the multi-scale enhanced feature map into the detection head, and output the normalized coordinates of the target box, the class probability and the confidence as the detection result through the regression branch and the classification branch respectively;

[0011] S6: Perform post-processing on the detection results. Specifically, the post-processing involves filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

[0012] Furthermore, the improved YOLOv10n backbone network described in step S2 includes four downsampling stages, which are sequentially passed through a cascaded structure of the Conv module, C2f module and SCDown module to output four sets of multi-scale feature maps.

[0013] The SCDown module halves the feature size through channel-separated convolution, while expanding the number of channels to twice the input.

[0014] The improved YOLOv10n backbone network has a spatial attention module embedded at its end.

[0015] Furthermore, the feature decoupling module in step S3 expands the number of channels of the input feature to twice the original number of channels through 1×1 convolution, and evenly divides it along the channel dimension into a basic structure feature branch and a detail texture feature branch, which are used to extract basic structure features and detail texture features, respectively.

[0016] The dual-branch enhancement module employs a C2f-CS module, which integrates channel attention and spatial attention mechanisms to enhance the basic structural feature branches; and employs a Gated Swin Transformer module, which enhances the detailed texture feature branches through a window attention mechanism.

[0017] The intelligent gating fusion module dynamically learns the fusion weights of the enhanced basic structural features and the enhanced detailed texture features through a global gating weight generator and a spatial coordination weight graph generator, thereby achieving adaptive fusion of the enhanced basic structural features and the enhanced detailed texture features.

[0018] Furthermore, the C2f-CS module includes: a channel attention mechanism, a spatial attention mechanism, a feature recalibration unit, and a residual connection structure;

[0019] The channel attention mechanism generates channel attention weights using the following formula. :

[0020] ,

[0021] in, For the Sigmoid function, For ReLU functions, For the input features, Indicates global average pooling. This indicates the first fully connected layer. This indicates the second fully connected layer;

[0022] The spatial attention mechanism generates a spatial attention weight map by performing a 7×7 depthwise separable convolution using the following formula. :

[0023] ,

[0024] in, Indicates global max pooling. For splicing operations; for Convolution operation;

[0025] The feature recalibration unit performs element-wise multiplication of the channel attention weights and the spatial attention weight map to obtain double-weighted features;

[0026] The residual connection structure includes two Bottleneck modules. Each Bottleneck includes 1×1 convolution dimensionality reduction, 3×3 convolution feature extraction, and 1×1 convolution dimensionality increase operations, which are used to output enhanced basic structural features from the double-weighted features through residual connections.

[0027] Furthermore, the Gated Swin Transformer module includes: a local window self-attention calculation unit, a shift window mechanism, a gated feedforward network, and a relative position bias table;

[0028] The local window self-attention calculation unit divides the input features into 7×7 local windows, and calculates the self-attention within each window using the following formula. :

[0029] ,

[0030] in, It is a query matrix. It is a key matrix. It is a value matrix. Indicates transpose. It is the dimension of the key vector. It is a normalized exponential function;

[0031] The shifting window mechanism alternately employs conventional window partitioning and shifting window partitioning strategies to promote information interaction between non-overlapping windows;

[0032] The gated feedforward network, by introducing a gating mechanism into the feedforward network, outputs features through the following formula:

[0033] ,

[0034] in, It is a feedforward network. For gated feedforward networks, For the input features, This indicates element-wise multiplication. The gating weights are generated through a two-layer fully connected network.

[0035] The relative position bias table assigns adaptive weights to features at different relative positions.

[0036] Furthermore, the implementation process of the intelligent gating fusion module described in step S3 includes:

[0037] The enhanced basic structural features and enhanced detail texture features are subjected to global average pooling by a global gate weight generator, and then input into a two-layer fully connected network to generate channel-dimensional gate weights.

[0038] Then, a spatially adaptive fusion weighted graph is generated using a spatially coordinated weighted graph generator through feature similarity analysis.

[0039] Next, the fused features are obtained by using the following formula through the gated fusion unit to achieve feature fusion. :

[0040] ,

[0041] in For gating weights, The residual coefficient is... For the enhanced basic features, Enhanced detail texture features;

[0042] Finally, a 1×1 convolution is used to restore the number of channels of the fused features to the original dimensions.

[0043] Further, the path aggregation network described in step S4 is implemented as follows: High-level strong semantic features are upsampled to the same size as shallow high-resolution features through bilinear interpolation via a bidirectional fusion process. After channel concatenation, the features are input into the C2f-CS module for feature enhancement, sequentially completing top-down fusion from 20×20 to 160×160. The enhanced shallow features are downsampled to the same size as mid-level features through the Conv module. After channel concatenation, the features are input into the C2f-CS module for feature enhancement, sequentially completing bottom-up fusion from 160×160 to 20×20. A small target detection layer is used to enhance the 160×160 scale feature map. Fine-grained detail features are extracted through 7×7 depthwise separable convolution. A spatial attention mechanism is used to weight the small target candidate region features and suppress background noise, ultimately outputting enhanced feature maps at scales of 160×160, 80×80, 40×40, and 20×20.

[0044] Furthermore, the detection head described in step S5 adopts an anchor-free design. Specifically, it outputs the normalized coordinates of the target box through the regression branch and optimizes the positioning accuracy using the complete intersection-union loss; it outputs the class probability and confidence through the classification branch and uses the focus loss to solve the class imbalance problem.

[0045] Further, step S6 specifically involves: reducing the confidence threshold of target boxes smaller than 32×32 pixels to 0.15 to improve recall by using a small target compensation mechanism; filtering target boxes with a size not smaller than 32×32 pixels using a default confidence threshold; calculating the intersection-union ratio (IU-NMS) of the predicted boxes using the DIoU-NMS algorithm and removing duplicate boxes with an overlap exceeding a set threshold; restoring the normalized coordinates to the original image size using the inverse transformation formula; drawing the predicted boxes on the original image and labeling them with corresponding category labels and confidence levels to complete the visualization output.

[0046] Based on the same inventive concept, the present invention also provides a small target detection system based on an intelligent decoupling network, comprising:

[0047] The preprocessing module is used to perform preprocessing operations on the input image to be detected to obtain standardized image data;

[0048] The extraction module is used to input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module.

[0049] The decoupling module is used to decouple the extracted features from basic structural features and detailed texture features based on an intelligent decoupling network, and then fuse the basic structural features and detailed texture features to output an enhanced feature; the intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module, and an intelligent gating fusion module;

[0050] The fusion module is used to aggregate the enhanced feature input path network, perform cross-scale feature deep fusion, and output multi-scale enhanced feature maps.

[0051] The detection module is used to input multi-scale enhanced feature maps into the detection head, and output the normalized coordinates of the target box, the class probability, and the confidence score as the detection results through the regression branch and the classification branch respectively.

[0052] The post-processing module is used to perform post-processing on the detection results. Specifically, the post-processing includes filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

[0053] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. This invention uses an intelligent decoupling network to accurately decompose the features extracted by the backbone network into basic structural features and detailed texture features, and uses C2f-CS and Gated Swin Transformer modules for targeted optimization respectively; combined with a gated residual fusion mechanism, it dynamically balances the contribution weights of the two types of features, preserving the complete contour information of the target while strengthening the expression of detailed features such as edges and textures, effectively solving the problem of insufficient feature extraction for small targets and occluded targets in traditional models, and significantly improving the mAP and recall rate of small target detection; 2. This invention uses a bidirectional path aggregation network to achieve deep fusion of multi-scale features, supplementing the semantic information of low-level features and the detailed information of high-level features, and with the special enhancement strategy of the small target detection layer, it further improves the discriminative power of small target features and reduces the probability of missed detection and false detection; 3. The core improved modules of this invention (intelligent decoupling network, C2f-CS, Gated Swin) The Transformer series adopts a modular design, which can be flexibly ported to other YOLO series models and adapted to different detection tasks and datasets. The gate weight generation mechanism supports dynamic adaptation to different target types without the need to manually adjust the feature fusion ratio. It controls the computational complexity while ensuring detection performance and has good engineering applicability. Attached Figure Description

[0054] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the overall network structure according to an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the feature processing procedure according to an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the C2f-CS module structure according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the Gated Swin Transformer module according to an embodiment of the present invention;

[0059] Figure 6 This is a diagram of the intelligent decoupling network structure according to an embodiment of the present invention;

[0060] Figure 7 This is a schematic diagram of the gated residual fusion module structure according to an embodiment of the present invention;

[0061] Figure 8 This is a visual comparison chart of the detection results in an embodiment of the present invention. Detailed Implementation

[0062] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] Example 1, as shown in the attached document Figure 1 As shown, the small target detection method based on intelligent decoupling network in this embodiment includes:

[0064] S1: Perform preprocessing operations on the input image to be detected to obtain standardized image data;

[0065] S2: Input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module.

[0066] S3: Based on an intelligent decoupling network, the extracted features are decoupled into basic structural features and detailed texture features, and the basic structural features and detailed texture features are fused together to output an enhanced feature; the intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module and an intelligent gating fusion module;

[0067] S4: The enhanced feature input path aggregation network is used to perform cross-scale feature deep fusion and output multi-scale enhanced feature maps.

[0068] S5: Input the multi-scale enhanced feature map into the detection head, and output the normalized coordinates of the target box, the class probability and the confidence as the detection result through the regression branch and the classification branch respectively;

[0069] S6: Perform post-processing on the detection results. Specifically, the post-processing involves filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

[0070] Specifically, this embodiment constructs a YOLOv10n-SCS network, with the overall structure as follows: Figure 2 As shown, it mainly consists of five core parts: an improved backbone network, an intelligent decoupling network, a path aggregation network, a small target detection layer, and a detection head. The improved backbone network achieves efficient extraction of multi-scale features through a cascaded structure of Conv, C2f, and SCDown modules. The intelligent decoupling network is embedded in the output of the backbone network to complete feature decoupling, branch enhancement, and dynamic fusion. The path aggregation network adopts a bidirectional path aggregation (BiFPN) structure to achieve deep fusion of cross-scale features. The small target detection layer performs specific enhancement on fine-scale feature maps to improve the feature response of small targets. The detection head achieves accurate prediction of target coordinates and class probabilities through independent regression and classification branches.

[0071] In step S2, the improved YOLOv10n backbone network adopts the CSPDarknet structure, which includes four downsampling stages. These stages are sequentially passed through a cascaded structure of Conv, C2f, and SCDown modules, outputting four sets of multi-scale feature maps (scales of 160×160, 80×80, 40×40, and 20×20), adapting to the detection needs of targets of different sizes. Specifically, the C2f module retains rich feature information while maintaining lightweight design through residual connections and grouped convolutions; the SCDown module halves the feature size through channel-separating convolutions while doubling the number of channels to the input, achieving efficient downsampling while ensuring feature expressiveness; and a spatial attention module is embedded at the end of the backbone network to enhance the feature response of the target region through spatial attention weighting, providing high-quality basic features for subsequent feature decoupling.

[0072] In step S3, the structure of the intelligent decoupling network is as follows: Figure 6 As shown, the core function is to decouple the high-dimensional features output by the backbone network from the basic structural features and the detailed texture features, and then dynamically fuse them after targeted enhancement. Specifically, it includes three parts: a feature decoupling module, a dual-branch enhancement module, and an intelligent gating fusion module.

[0073] The Feature Decoupling Module (FDM) expands the number of input feature channels to twice the original number of channels through 1×1 convolution, and evenly divides them into two branches along the channel dimension, which are used to extract basic structural features and detailed texture features respectively.

[0074] The basic structural feature branch in the dual-branch enhancement module adopts the C2f-CS module, which integrates the channel attention mechanism (CAM) and the spatial attention mechanism (SAM) to enhance the expression of macro features such as target contours. The detailed texture feature branch adopts the Gated Swin Transformer module (GST), which captures micro-detail features such as target edges and textures and global context information through the window attention mechanism.

[0075] The intelligent gating fusion module dynamically learns the fusion weights of two types of features through a global gating weight generator and a spatially coordinated weight graph generator, thereby achieving adaptive fusion of basic and detailed features.

[0076] The structure of the C2f-CS module is as follows: Figure 4 As shown, the core is to enhance the expressive power of basic structural features through a dual attention mechanism and residual connections. The specific structure and working principle are as follows: The channel attention mechanism (CAM) uses the formula... Generate channel weights, where For the Sigmoid function, The ReLU function is used, with a compression ratio set to 16, to achieve adaptive weighting of key feature channels. For the input features, Indicates global average pooling. This indicates the first fully connected layer. This represents the second fully connected layer; the Spatial Attention (SAM) mechanism is implemented through the formula... Perform a 7×7 depthwise separable convolution to generate a spatial weight map, focusing on the spatial region where the target is located. Indicates global max pooling. For splicing operations; for Convolutional operations; the feature recalibration unit performs element-wise multiplication of the channel attention weights with the spatial attention map to obtain double-weighted features, while suppressing background noise and redundant information; the residual connection structure (RCS) contains two Bottleneck modules, each of which contains 1×1 convolutional dimensionality reduction, 3×3 convolutional feature extraction and 1×1 convolutional dimensionality increase operations. The residual connection avoids feature degradation and outputs enhanced basic structural features.

[0077] The window attention calculation process of the Gated Swin Transformer module is as follows: Figure 5As shown, the core is to efficiently capture the local correlation and global context of detailed features through windowed attention and gating mechanisms. The specific structure and working principle are as follows: The Local Window Self-Attention Calculation Unit (LWSAU) divides the input features into 7×7 local windows, and within each window, the formula is used to... Self-attention computation enhances local feature interactions while reducing computational complexity. It is a query matrix. It is a key matrix. It is a value matrix. Indicates transpose. It is the dimension of the key vector. It is a normalized exponential function; the shifted window mechanism (SWM) alternates between conventional window partitioning and shifted window partitioning strategies to promote information interaction between non-overlapping windows and achieve global context modeling; the gated feedforward network introduces a gating mechanism into the feedforward network, through the formula Output features, where, It is a feedforward network. For gated feedforward networks, For the input features, This indicates element-wise multiplication. For gating weights, where the gating weights are... The feature is generated through a two-layer fully connected network, enabling adaptive filtering and enhancement. The Relative Position Bias Table (RPBT) assigns adaptive weights to features at different relative positions, improving the ability to capture position-sensitive detail features.

[0078] The structure of the intelligent gate control fusion module is as follows: Figure 7 As shown, the core is to dynamically balance the contribution weights of basic structural features and detailed texture features. The specific working principle is as follows: The Global Gated Weight Generator (GGWG) performs global average pooling on the enhanced features of the basic structural branch and the enhanced features of the detailed texture branch, respectively, and inputs them into a two-layer fully connected network to generate channel-dimensional gated weights; the Spatial Coordinating Weight Graph Generator (SCWMG) generates a spatially adaptive fusion weight graph through feature similarity analysis, achieving accurate fusion of spatial dimensions; the Gated Fusion Unit (GFU) uses a formula... To achieve feature fusion, among which For gating weights, The residual coefficient is... For the enhanced basic features, To enhance the detailed texture features and ensure that the basic structural features are not overwhelmed by the detailed features, while improving gradient flow, channel restoration is performed by using 1×1 convolutions to restore the number of channels of the fused features to the original dimensions, ensuring compatibility with the input of the subsequent path aggregation network.

[0079] In step S4, the path aggregation network adopts a bidirectional path aggregation (BiFPN) structure to achieve deep fusion of multi-scale features. The specific strategy is as follows: The bidirectional fusion process upsamples the high-level strong semantic features to the same size as the shallow high-resolution features through bilinear interpolation, performs channel concatenation (Concat), and then inputs them into the C2f-CS module to enhance the features, sequentially completing the top-down fusion from 20×20 to 160×160; the enhanced shallow features are downsampled to the same size as the mid-level features through the Conv module, and channel concatenation is performed. The C2f-CS module is then used to enhance features, sequentially performing bottom-up fusion from 160×160 to 20×20. For the small target detection layer, the enhancement targets the 160×160 scale feature map, extracting fine-grained details through 7×7 depthwise separable convolutions. A spatial attention mechanism is then used to weight and suppress background noise in the small target candidate region features, ultimately outputting enhanced feature maps for four detection scales (160×160, 80×80, 40×40, and 20×20), accurately matching the detection requirements of small targets of different sizes. The complete feature extraction, decoupling, and fusion process in this embodiment is as follows: Figure 3 As shown.

[0080] In step S5, the detection head adopts an anchor-free design to avoid redundant anchor box matching. Target localization and classification are achieved through independent regression and classification branches, specifically as follows: The regression branch (RB) contains four 3×3 convolutional layers, outputting the normalized coordinates of the bounding box. A complete intersection-union loss (CIoU Loss) is used to optimize localization accuracy, with the loss formula being... ,in , Let Euclidean distance be the center point of the predicted bounding box and the ground truth bounding box. The length of the diagonal of the smallest rectangle enclosing the two frames. These are the weighting coefficients. The aspect ratio consistency index is used; the classification branch (CB) contains 3 layers of 3×3 convolutions, outputting the class probability distribution (a combined result of class probability and confidence). Focal loss is used to solve the class imbalance problem, and the loss formula is as follows: ,in For column weights, To predict probabilities, For focusing parameters.

[0081] In step S6, the core of post-detection processing is to select the optimal detection results and balance recall and precision. The specific process is as follows: Confidence filtering uses a small target compensation mechanism (STCM) to reduce the confidence threshold of target boxes smaller than 32×32 pixels to 0.15 to improve recall. For regular target boxes, a default confidence threshold is used to filter low-confidence predicted boxes. Duplicate box removal uses the DIoU-NMS algorithm to calculate the intersection-union ratio between predicted boxes and removes duplicate boxes with an overlap exceeding a set threshold. Coordinate restoration uses an inverse transformation formula to restore the normalized coordinates to the original image size. Visualization output draws the predicted boxes on the original image and labels the corresponding category and confidence, outputting a visualized image and structured detection results to support subsequent analysis.

[0082] Example 2: The experimental setup and performance verification of the method in this example are as follows:

[0083] In the dataset and scene selection stage, typical datasets under drone aerial photography scenarios were selected, covering four types of challenging scenarios, including dense targets, complex backgrounds, motion blur, and small targets at a distance. Small targets accounted for no less than 60% of the dataset, and the target pixel area was mainly distributed between 10×10 and 32×32.

[0084] During the model training and configuration phase, a deep learning framework is used to build the model, an adaptive optimizer is used to adjust the training parameters, multiple rounds of warm-up training are set to stabilize the initial training process, and data augmentation methods are used to improve the model's generalization ability.

[0085] A comparative experiment was conducted between the method of this invention and the benchmark model YOLOv10n. The results show that the method of this invention improves the detection box precision (BOX(P)) by 8.9%, mAP@50 and mAP@50-95 by 9.3% and 6.7% respectively, and the recall rate by 7.2%, achieving significant performance improvements in multiple challenging scenarios. A comparison of the detection performance of the method of this invention and the benchmark model is provided below. Figure 8 As shown.

[0086] Example 3, based on the same inventive concept, this example also provides a small target detection system based on an intelligent decoupling network, including:

[0087] The preprocessing module is used to perform preprocessing operations on the input image to be detected to obtain standardized image data;

[0088] The extraction module is used to input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module.

[0089] The decoupling module is used to decouple the extracted features from basic structural features and detailed texture features based on an intelligent decoupling network, and then fuse the basic structural features and detailed texture features to output an enhanced feature; the intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module, and an intelligent gating fusion module;

[0090] The fusion module is used to aggregate the enhanced feature input path network, perform cross-scale feature deep fusion, and output multi-scale enhanced feature maps.

[0091] The detection module is used to input multi-scale enhanced feature maps into the detection head, and output the normalized coordinates of the target box, the class probability, and the confidence score as the detection results through the regression branch and the classification branch respectively.

[0092] The post-processing module is used to perform post-processing on the detection results. Specifically, the post-processing includes filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

[0093] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A small target detection method based on intelligent decoupling networks, characterized in that, include: S1: Perform preprocessing operations on the input image to be detected to obtain standardized image data; S2: Input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module. S3: Based on the intelligent decoupling network, the extracted features are decoupled into basic structural features and detailed texture features, and the basic structural features and detailed texture features are fused together to output the enhanced features; The intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module, and an intelligent gating fusion module; S4: The enhanced feature input path aggregation network is used to perform cross-scale feature deep fusion and output multi-scale enhanced feature maps. S5: Input the multi-scale enhanced feature map into the detection head, and output the normalized coordinates of the target box, the class probability and the confidence as the detection result through the regression branch and the classification branch respectively; S6: Perform post-processing on the detection results. Specifically, the post-processing involves filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

2. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, The improved YOLOv10n backbone network described in step S2 includes four downsampling stages, which are sequentially passed through a cascaded structure of the Conv module, C2f module and SCDown module to output four sets of multi-scale feature maps. The SCDown module halves the feature size through channel-separated convolution, while expanding the number of channels to twice the input. The improved YOLOv10n backbone network has a spatial attention module embedded at its end.

3. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, The feature decoupling module described in step S3 expands the number of channels of the input feature to twice the original number of channels through 1×1 convolution, and evenly divides it along the channel dimension into a basic structure feature branch and a detailed texture feature branch, which are used to extract basic structure features and detailed texture features respectively. The dual-branch enhancement module employs a C2f-CS module, which integrates channel attention and spatial attention mechanisms to enhance the basic structural feature branches; and employs a Gated Swin Transformer module, which enhances the detailed texture feature branches through a window attention mechanism. The intelligent gating fusion module dynamically learns the fusion weights of the enhanced basic structural features and the enhanced detailed texture features through a global gating weight generator and a spatial coordination weight graph generator, thereby achieving adaptive fusion of the enhanced basic structural features and the enhanced detailed texture features.

4. The small target detection method based on intelligent decoupling network according to claim 3, characterized in that, The C2f-CS module includes: a channel attention mechanism, a spatial attention mechanism, a feature recalibration unit, and a residual connection structure; The channel attention mechanism generates channel attention weights using the following formula. : ， in, For the Sigmoid function, For ReLU functions, For the input features, Indicates global average pooling. This indicates the first fully connected layer. This indicates the second fully connected layer; The spatial attention mechanism generates a spatial attention weight map by performing a 7×7 depthwise separable convolution using the following formula. : ， in, Indicates global max pooling. For splicing operations; for Convolution operation; The feature recalibration unit performs element-wise multiplication of the channel attention weights and the spatial attention weight map to obtain double-weighted features; The residual connection structure includes two Bottleneck modules. Each Bottleneck includes 1×1 convolution dimensionality reduction, 3×3 convolution feature extraction, and 1×1 convolution dimensionality increase operations, which are used to output enhanced basic structural features from the double-weighted features through residual connections.

5. The small target detection method based on intelligent decoupling network according to claim 3, characterized in that, The Gated Swin Transformer module includes: a local window self-attention calculation unit, a shift window mechanism, a gated feedforward network, and a relative position offset table; The local window self-attention calculation unit divides the input features into 7×7 local windows, and calculates the self-attention within each window using the following formula. : ， in, It is a query matrix. It is a key matrix. It is a value matrix. Indicates transpose. It is the dimension of the key vector. It is a normalized exponential function; The shifting window mechanism alternately employs conventional window partitioning and shifting window partitioning strategies to promote information interaction between non-overlapping windows; The gated feedforward network, by introducing a gating mechanism into the feedforward network, outputs features through the following formula: ， in, It is a feedforward network. For gated feedforward networks, For the input features, This indicates element-wise multiplication. The gating weights are generated through a two-layer fully connected network. The relative position bias table assigns adaptive weights to features at different relative positions.

6. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, The implementation process of the intelligent gating fusion module in step S3 includes: The enhanced basic structural features and enhanced detail texture features are subjected to global average pooling by a global gate weight generator, and then input into a two-layer fully connected network to generate channel-dimensional gate weights. Then, a spatially adaptive fusion weighted graph is generated using a spatially coordinated weighted graph generator through feature similarity analysis. Next, the fused features are obtained by using the following formula through the gated fusion unit to achieve feature fusion. : ， in For gating weights, The residual coefficient is... For the enhanced basic features, Enhanced detail texture features; Finally, a 1×1 convolution is used to restore the number of channels of the fused features to the original dimensions.

7. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, The path aggregation network described in step S4 is implemented as follows: High-level strong semantic features are upsampled to the same size as shallow high-resolution features through bilinear interpolation via a bidirectional fusion process. After channel concatenation, the features are input into the C2f-CS module for enhancement, completing top-down fusion from 20×20 to 160×160 sequentially. The enhanced shallow features are downsampled to the same size as mid-level features through the Conv module. After channel concatenation, the features are input into the C2f-CS module for enhancement, completing bottom-up fusion from 160×160 to 20×20 sequentially. A small object detection layer is used to enhance the 160×160 scale feature map. Fine-grained detail features are extracted through 7×7 depthwise separable convolution. Spatial attention mechanisms are used to weight the small object candidate region features and suppress background noise, ultimately outputting enhanced feature maps at scales of 160×160, 80×80, 40×40, and 20×20.

8. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, The detection head described in step S5 adopts an Anchor-Free design. Specifically, it outputs the normalized coordinates of the target box through a regression branch and uses full intersection-union loss to optimize the positioning accuracy; it outputs the class probability and confidence through a classification branch and uses focus loss to solve the class imbalance problem.

9. The small target detection method based on intelligent decoupling network according to claim 1, characterized in that, Step S6 specifically involves: using a small target compensation mechanism to reduce the confidence threshold of target boxes smaller than 32×32 pixels to 0.15 to improve recall; using a default confidence threshold to filter target boxes with a size not smaller than 32×32 pixels; using the DIoU-NMS algorithm to calculate the intersection-union ratio (IU / NMS) between predicted boxes and removing duplicate boxes with an overlap exceeding a set threshold; restoring the normalized coordinates to the original image size using the inverse transformation formula; drawing the predicted boxes on the original image and labeling them with the corresponding category labels and confidence levels to complete the visualization output.

10. A small target detection system based on an intelligent decoupling network, characterized in that, include: The preprocessing module is used to perform preprocessing operations on the input image to be detected to obtain standardized image data; The extraction module is used to input the standardized image data into the improved YOLOv10n backbone network to extract features. The improved YOLOv10n backbone network contains a cascaded structure of Conv module, C2f module and SCDown module. The decoupling module is used to decouple the extracted features from basic structural features and detailed texture features based on the intelligent decoupling network, and then fuse the basic structural features and detailed texture features to output the enhanced features. The intelligent decoupling network includes: a feature decoupling module, a dual-branch enhancement module, and an intelligent gating fusion module; The fusion module is used to aggregate the enhanced feature input path network, perform cross-scale feature deep fusion, and output multi-scale enhanced feature maps. The detection module is used to input multi-scale enhanced feature maps into the detection head, and output the normalized coordinates of the target box, the class probability, and the confidence score as the detection results through the regression branch and the classification branch respectively. The post-processing module is used to perform post-processing on the detection results. Specifically, the post-processing includes filtering low-confidence prediction boxes and removing duplicate boxes, restoring the target box coordinates to the original image size, and outputting the visualized detection results.

Images