A multi-scale target feature enhancement method based on frequency domain-space domain cooperation

Abstract

The application discloses a multi-scale target feature enhancement method based on frequency domain-space domain cooperation, and proposes a double-domain cooperative enhancement residual connection module (DDARB), which adopts a parallel double-branch residual connection structure, including a wavelet transform convolution branch (WT-Conv) and a multi-receptive field spatial aggregation branch (MRF-SAM). The wavelet transform convolution branch enhances the contour features of large targets by expanding the receptive field, and the multi-receptive field spatial aggregation branch improves the response ability of small targets by retaining multi-scale details. The output feature maps of the two parallel branches are added to the original input feature map in residual, and an enhanced feature map is obtained. The module can be embedded into various target detection networks, including but not limited to YOLO, DETR and R-CNN networks, and is especially suitable for robot vision grabbing, automatic driving, unmanned aerial vehicle inspection and other task scenarios that need to cope with the dynamic changes of the distance between the camera and the target object.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and deep learning technology, and in particular, it is a multi-scale target feature enhancement method based on frequency domain-spatial domain collaboration. It is especially suitable for task scenarios such as robot vision grasping, autonomous driving, and drone inspection that require dealing with dynamic changes in the distance between the vision sensor and the target object. Background Technology

[0002] Object detection is one of the core tasks in computer vision, with wide applications in robot vision grasping, autonomous driving, and intelligent monitoring. Single-stage object detection algorithms, such as the YOLO series, have become core components of visual perception systems due to their excellent balance between accuracy and speed. However, in real-world applications, such as robot vision grasping tasks, the distance between the visual sensor and the target object is continuously and dynamically changing. When the end-effector camera approaches the target, the object occupies a large area in the image, its local details are clear, but its overall structure exceeds the receptive field of a conventional convolutional kernel; when the camera moves away from the target, the object's pixel area is too small, and its key features are easily lost during the downsampling process of the network. This loss of features at different scales caused by dynamic distance changes is one of the main reasons for the performance degradation of existing object detection networks in practical applications.

[0003] Researchers have proposed various solutions to address the challenges of feature extraction and multi-scale perception in target detection. Chinese patent document CN118154840A discloses an infrared small target detection method based on a cross-channel fusion attention mechanism. This method refines and fuses shallow and deep features by designing a cross-channel fusion attention module, reducing information loss during feature extraction. Chinese patent document CN118762172B discloses an infrared target detection method based on group feature interaction and assigned feature fusion. This method uses a multi-level feature fusion module to perform multiple cross-fusions and storage reuse of extracted features from different levels, achieving higher detection accuracy with a lower parameter count. While these methods have made valuable explorations in feature fusion, they are essentially still post-optimizations of extracted features and do not enhance the network's ability to represent multi-scale information from the source of feature extraction.

[0004] To address the issue of limited receptive field in convolutional kernels, Chinese patent document CN118691645A discloses a vehicle tracking method that replaces the convolutional layers in the YOLO network with variable kernel convolutional layers, improving adaptability to targets of different shapes by changing the kernel size. Chinese patent document CN120807587A discloses an open-vocabulary multi-target tracking method based on confidence adjustment and wavelet convolution, introducing Haar wavelet transform to expand the receptive field of the convolutional operation, improving the ability to capture high-frequency motion details and preserve low-frequency contour information. Chinese patent document CN110751160B discloses a method for detecting objects in images, improving detection accuracy in complex environments through focus stacking and multi-level feature fusion.

[0005] However, the above methods still have the following technical bottlenecks:

[0006] (i) Insufficient single-dimensional enhancement. Existing improvement schemes mostly focus on a single dimension: expanding the receptive field only in the spatial domain through variable kernel convolution, or only in the frequency domain through wavelet transform, Fourier transform, and other means. However, the scale problem caused by changes in camera distance is a complex multi-dimensional problem—long distances rely on high-frequency detail preservation, while close distances rely on low-frequency shape understanding. Therefore, a collaborative enhancement mechanism is needed in both the frequency and spatial domains.

[0007] (ii) The challenge of balancing receptive field and efficiency. While using a single ultra-large convolutional kernel can expand the receptive field, the number of parameters and computational cost increases quadratically with the kernel size, and it is prone to boundary effects. Recent studies have shown that by appropriately combining small and medium-sized convolutional kernels, the receptive field can be effectively expanded while maintaining efficiency, but a plug-and-play modular solution has not yet been formed, and there is a lack of specific optimization for dynamic distance scenarios.

[0008] (iii) The lag in multi-scale feature fusion. Although feature pyramid structures, such as FPN and PANet, have alleviated the scale problem to some extent, they are essentially a late-stage fusion of extracted features. When small target features are lost during the front-end downsampling process, subsequent fusion is difficult to recover effectively.

[0009] To address the aforementioned issues, there is an urgent need to design a modular feature enhancement method that can simultaneously address scale variation problems from both the frequency and spatial domains using a lightweight pure convolution approach, thereby improving the network's multi-scale target detection capabilities in dynamic distance scenarios. Summary of the Invention

[0010] The purpose of this invention is to provide a multi-scale target feature enhancement method based on frequency domain-spatial domain synergy, in order to solve the problem of drastic fluctuations in object scale caused by continuous changes in the distance between the camera and the target in dynamic distance scenarios, and to overcome the technical defects of existing methods such as insufficient single-dimensional enhancement, difficulty in balancing receptive field and efficiency, and lag in multi-scale feature fusion.

[0011] The technical solution to achieve the purpose of this invention is as follows:

[0012] A multi-scale target feature enhancement method based on frequency-spatial domain collaboration is proposed. The feature map initially extracted by the target detection network backbone is input into a dual-domain collaborative enhancement residual connection module (DDARB). The DDARB module adopts a residual connection structure, with the main path being an identity mapping. The enhancement path consists of parallel wavelet transform convolution branches and multi-receptive-field spatial aggregation branches. Specifically, the wavelet transform convolution branch uses a Haar wavelet basis to perform multi-level two-dimensional discrete wavelet transform on the input feature map, decomposing it into low-frequency sub-bands and high-frequency sub-bands in three directions. Each sub-band is then depthwise convolved and reconstructed using inverse wavelet transform to expand the receptive field and enhance the contour features of large targets. The multi-receptive-field spatial aggregation branch simultaneously feeds the input feature map into multiple parallel convolutional layers of different sizes and depths, and adaptively weights them using a channel attention mechanism to preserve multi-scale details and improve the responsiveness of small targets. The outputs of the two branches are summed, and then added to the residual of the input feature map to obtain the enhanced feature map.

[0013] The significant advantages of this invention compared to existing technologies are:

[0014] (1) Frequency-spatial domain synergistic enhancement to specifically address the problem of object scale fluctuations caused by dynamic distance. This invention proposes a residual connection structure with dual-path synergistic enhancement in the frequency and spatial domains. The WT-Conv branch explicitly expands the receptive field through multi-level wavelet decomposition, with the low-frequency subband enhancing the overall shape perception of large targets at close range and the high-frequency subband preserving the edge details of small targets at distant range. The MRF-SAM branch enables the network to adaptively call features from different receptive fields to match the target scale at different distances through parallel multi-size large kernel convolution. The synergistic effect of the two solves the multi-scale perception problem in dynamic distance scenarios from the source of feature extraction.

[0015] (2) Pure CNN architecture, engineering-friendly. This invention is built entirely on the basic components of convolutional neural networks, without introducing any Transformer architecture or complex self-attention mechanisms. Although the SE module used implements channel attention, it is essentially a lightweight combination of global average pooling and two fully connected layers, with minimal computational overhead. Therefore, while achieving significant performance gains, this invention maintains the high computational efficiency, low memory footprint, and excellent hardware compatibility of classic CNNs. It can be directly deployed on edge computing platforms such as NVIDIA Jetson series and Huawei Ascend Atlas 800, meeting the requirements of real-time performance and low power consumption for engineering tasks such as robot vision grasping.

[0016] (3) Plug and play, with extremely low modification costs. The input and output dimensions of the dual-domain collaborative enhancement residual connection module are completely consistent. When integrated into the target detection network, there is no need to add an additional dimension matching layer or modify the rest of the network. Only a small number of feature extraction modules need to be replaced in the backbone network to improve the overall network's multi-scale target recognition and detection performance. The amount of code modification is minimal, and there is no need to redesign the training strategy or adjust the hyperparameters, which greatly reduces the cost of technology migration.

[0017] (4) Flexible deployment location and wide applicability. To avoid the loss of feature information due to the boundary effect of large convolutional kernels in the MRF-SAM branch caused by the small feature map size, the DDARB module should be integrated into a stage with moderate feature map resolution in the backbone network. This stage needs to satisfy the requirement that the ratio of the feature map spatial size (H×W) to the maximum convolutional kernel (11×11) is large enough. Specifically, the minimum side length should not be less than twice the size of the maximum convolutional kernel, i.e., H ≥ 22 and W ≥ 22, to avoid the loss of feature information caused by filling in the edge region of the convolutional kernel and to ensure the effectiveness of the convolution operation. Taking the YOLOv11 object detection network architecture as an example, when integrating the DDARB module into the YOLOv11 backbone network, Stage 2 (80×80) and Stage 3 (40×40) levels are selected for integration. For other object detection network architectures, this principle can be used as a reference to flexibly adjust the feature map size according to the specific network architecture.

[0018] (5) Strong generalization ability and broad application prospects. In addition to robot vision grasping, this invention can also be directly transferred to other task scenarios that require dealing with the dynamic changes in the distance between the vision sensor and the target object, such as autonomous driving, drone inspection, and intelligent monitoring. It has significant academic value and industrialization potential. Attached Figure Description

[0019] Figure 1 This is a data processing flowchart of the Dual-Domain Collaborative Enhanced Residual Connection Module (DDARB) proposed in this invention.

[0020] Figure 2 A flowchart outlining the general steps for integrating the proposed Dual-Domain Collaborative Enhanced Residual Connection Module (DDARB) into a target detection network.

[0021] Figure 3 This is a schematic diagram of a general architecture for an object detection network.

[0022] Figure 4 for Figure 3 The diagram shows the structure of some basic modules in the target detection network.

[0023] Figure 5 To integrate the proposed Dual-Domain Collaborative Enhanced Residual Connection Module (DDARB) into... Figure 3 The diagram shows the network architecture after the backbone of the target detection network.

[0024] Figure 6 This is a schematic diagram of the structure of the Dual Domain Collaborative Enhancement Residual Module (DDARB) proposed in this invention.

[0025] Figure 7 This is a schematic diagram of the wavelet transform convolution branch (WT-Conv) in this invention.

[0026] Figure 8 This is a schematic diagram illustrating the principle of wavelet transform convolution to expand the receptive field in this invention.

[0027] Figure 9 This is a schematic diagram of the structure of the Multi-Receptive Field Spatial Aggregation Branch (MRF-SAM) in this invention.

[0028] Figure 10 This is a schematic diagram of the structure and data flow of the channel attention mechanism unit (SE Block) in this invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, the present invention proposes a multi-scale target feature enhancement method based on frequency domain-spatial domain synergy, which includes the following steps:

[0031] S1 acquires the feature map of the image after initial extraction by the backbone network of the object detection network;

[0032] S2 inputs the feature map into a feature enhancement module called "Dual-Domain Co-Augmented Residual Block (DDARB)". This module enhances the feature extraction capability of the target detection network for multi-scale targets from the source of feature extraction through the synergistic effect of two orthogonal dimensions, the frequency domain and the spatial domain. It obtains a feature map after processing by the DDARB module, which integrates the contour semantics enhanced by the frequency domain and the detailed information aggregated by the spatial domain. This significantly improves the richness of feature expression and enhances the target detection network's ability to recognize and detect multi-scale targets, thereby solving the problem of feature loss caused by drastic changes in object scale in dynamic distance scenarios such as robot vision grasping.

[0033] (I) Design of the DDARB module:

[0034] like Figure 6 As shown, this module employs a residual connection structure. The main path is an identity mapping, directly passing the input feature map to the output. In the enhancement path, the first branch is a wavelet transform convolution (WT-Conv) branch, used to perform multi-level wavelet decomposition and reconstruction on the input feature map to extract frequency domain enhancement features. The second branch is a multi-receptive field spatial aggregation module (MRF-SAM) branch, which extracts spatial domain enhancement features in parallel using multiple convolutional kernels of different sizes and adaptively weights them using channel attention. The outputs of the first and second branches are added together, and then element-wise added to the input feature map of the main path to form a complete residual connection, resulting in the output enhanced feature map.

[0035]

[0036] in For the input feature map, This is the first enhanced feature map output by the wavelet transform convolution branch. The second enhanced feature map is the output of the multi-receptive field spatial aggregation branch. For the final enhanced feature map, H is the feature map height, W is the feature map width, and C is the number of feature map channels.

[0037] (II) Design of the wavelet transform convolution branch WT-Conv:

[0038] like Figure 7As shown, this branch uses the Haar wavelet basis to perform an L-level two-dimensional discrete wavelet transform on the input feature map. The wavelet transform decomposes the feature map into sub-bands of different frequencies through low-pass and high-pass filters. Its core advantage is that by performing a 3×3 depth-small kernel convolution operation on the low-frequency sub-bands, a large receptive field on the input image can be obtained equivalently, and the receptive field grows exponentially with the decomposition level, while the number of parameters grows only linearly.

[0039] Each level of transformation decomposes the feature map into a low-frequency subband. and three high-frequency subbands Low-frequency information is extracted from the feature maps respectively. ) and horizontal direction ( ), vertical direction ( ) and diagonal direction ( High-frequency information on ) is specifically represented as follows:

[0040] i=1,2,...,L

[0041] Where i represents the wavelet convolution operation level, the entire wavelet convolution process is recursively derived from the shallowest level i=1 to i=L. For convenience, let X represent the input feature map of the first-level decomposition, i.e. , This represents the wavelet transform convolution operation.

[0042] The wavelet transform uses the Haar wavelet basis:

[0043] , , ,

[0044] in These are convolution kernels that extract low-frequency information, high-frequency information in the horizontal direction, high-frequency information in the vertical direction, and high-frequency information in the diagonal direction from the feature map using wavelet transform convolution operations based on the Haar wavelet basis.

[0045] The specific operation of wavelet transform convolution is as follows:

[0046]

[0047] . . .

[0051]

[0052] in This represents a convolution operation on the feature map. The convolution operation uses the Haar wavelet basis to decompose each channel of the feature map. The total number of channels in the output subband of each operation is four times the number of input channels.

[0053] For each frequency sub-band of each level, perform depthwise convolution operations using independent 3×3 depthwise convolution kernels:

[0054] ,

[0055] in for The feature map obtained after a 3×3 depthwise convolution operation has subscripts LL, LH, HL, and HH, which represent the low-frequency information subband, horizontal high-frequency information subband, vertical high-frequency information subband, and diagonal high-frequency information subband obtained from the wavelet transform convolution operation, respectively. This represents performing an M×M depthwise convolution operation on the feature map.

[0056] The convolution results of each sub-band are reconstructed and fused step by step using inverse wavelet transform:

[0057] j=L,L-1,...,1

[0058] in The feature map is obtained from the j-th level inverse wavelet transform process. Here, j represents the inverse wavelet transform series. The entire inverse wavelet transform process is recursively derived from the deepest level j=L down to j=1. for The feature map obtained after a 3×3 depthwise convolution operation The low-frequency subband is obtained by performing a j-level wavelet transform convolution operation on X. represent The set of values, the inverse wavelet transform takes them as independent parameter inputs, i.e. , This represents the inverse wavelet transform operation.

[0059] The final result is:

[0060]

[0061] Example of receptive field expansion: such as Figure 8As shown, for a 3×3 convolution kernel, the equivalent receptive field in the low-frequency subband after L-level wavelet decomposition is (3×2L)×(3×2L). When L=2, the equivalent receptive field is 12×12; when L=3, the equivalent receptive field is 24×24. This indicates that through wavelet transform, small-sized convolution kernels can obtain a receptive field comparable to large-sized convolution kernels at low cost, while preserving multi-frequency features. The low-frequency subband enhances the perception of the overall shape of large-scale targets, while the high-frequency subband preserves the edge details of small-scale targets, avoiding feature loss during downsampling.

[0062] Preferably, the present invention uses L=2. Under this configuration, the equivalent receptive field of the 3×3 convolution kernel in the lowest frequency subband is 12×12, which matches the 11×11 large kernel receptive field of the MRF-SAM branch, achieving a good frequency domain-spatial domain synergistic enhancement effect.

[0063] (III) Design of Multi-Receptive Field Spatial Convergence Branch MRF-SAM:

[0064] like Figure 9 As shown, this branch employs a design combining multi-size large-kernel depthwise convolution with a channel attention mechanism. The input feature map... Simultaneously sent in The system uses three parallel deep convolutional layers, preferably with K=3, and kernel sizes of 3×3, 7×7, and 11×11, with a stride of 1 for each layer and padding of 1, 3, and 5 respectively, to ensure consistency in the size of the input and output feature maps. The output feature map is represented as follows:

[0065]

[0066]

[0067]

[0068] in Input feature map The output feature map after passing through the depthwise convolutional unit of the k-th branch. .

[0069] Output feature map for each branch Channel attention recalibration is performed separately through the channel attention mechanism unit SE Block. For example... Figure 10 As shown, firstly... Perform global average pooling to generate channel description vectors. :

[0070]

[0071] in, represent The value of the c-th channel, represent Feature map of the c-th channel Representative to The average value is calculated from all pixels of the feature map of the c-th channel. represent The pixel value in the m-th row and n-th column. .

[0072] The non-linear dependencies between channels are learned through two fully connected layers, generating channel weight vectors. :

[0073] Where the weight vector Bias vector r is the compression ratio. For ReLU activation function, This is the Sigmoid activation function.

[0074] The feature map output by the depthwise convolutional unit of the k-th branch With channel weight vector By multiplying each path, we obtain the weighted feature map of the k-th branch. :

[0075]

[0076] in This represents multiplying each channel sequentially.

[0077] The three weighted feature maps are summed element-wise to obtain the final output of the MRF-SAM branch:

[0078]

[0079] Example 1: A multi-scale target recognition and detection method for dynamic distance scenarios based on DDARB module feature enhancement:

[0080] This embodiment addresses the problem of drastic fluctuations in object scale caused by continuous changes in the distance between the camera and the target in dynamic distance scenarios such as robot vision grasping and autonomous driving. It proposes a feature enhancement method based on the dual-domain collaborative enhancement residual connection module DDARB, and integrates it into the target detection network to improve the network's multi-scale target recognition and detection performance in dynamic distance scenarios. Figure 2 As shown, the method includes the following general steps:

[0081] S1 Analysis of the Backbone Network Architecture of the Object Detection Network:

[0082] We analyzed each stage (Stage 1-Stage 4) responsible for feature extraction in the backbone network and recorded the spatial resolution of the output feature maps at each stage. Typically, the feature map size decreases progressively with network depth. To ensure that the effective receptive field of the large convolutional kernel in the MRF-SAM branch can cover a sufficient spatial range and avoid feature degradation caused by boundary effects, we selected an intermediate feature extraction stage with moderate feature map resolution as the integration location for the DDARB module. This stage must satisfy a sufficiently large ratio between the feature map spatial size (H×W) and the largest convolutional kernel (11×11). Specifically, the minimum side length should not be less than twice the size of the largest convolutional kernel, i.e., H ≥ 22 and W ≥ 22, to avoid feature information loss due to padding in the edge regions of the convolutional kernel and ensure the effectiveness of the convolution operation.

[0083] S2 module replacement:

[0084] The original feature extraction module in the selected feature extraction stage is replaced with the DDARB module proposed in this invention. Since the number of input channels, output channels, feature map height, and width of the DDARB module are completely identical to those of the original module, no additional dimension matching layer is needed, and it can be directly embedded into the network. The structure of other parts of the network (including other feature extraction stages in the backbone network, the neck network, the head network, etc.) remains unchanged from the original design.

[0085] S3 Training and Deployment:

[0086] All convolutional kernel weight vectors within the DDARB module and the weight vectors of the fully connected layers in the SE module are learnable parameters. They can be jointly optimized with the network backbone using the backpropagation algorithm in deep learning frameworks such as PyTorch and TensorFlow. During training, end-to-end training can be performed using the original network's training parameters, loss function, and optimization strategies, without requiring special adjustments. During deployment, the trained weight file can be loaded onto edge computing devices (such as the NVIDIA Jetson series) and accelerated using inference engines such as TensorRT to meet real-time requirements.

[0087] Example 2: Specific integration method of DDARB module taking YOLOv11 network architecture as an example:

[0088] To more intuitively demonstrate the general method described above, the integration process of the DDARB module is explained in detail below using the YOLOv11 object detection network as an example, in order to improve the network's multi-scale object recognition and detection performance in dynamic distance scenarios. The specific steps are as follows:

[0089] S1 analyzes the network architecture to locate the integration point of the DDARB module:

[0090] YOLOv11 was selected as the basic network, and its general network architecture diagram is as follows: Figure 3 As shown, it includes the backbone network, neck network, and head network connected in sequence. Figure 4 Showing Figure 3 The diagram shows the structure of some basic modules in the network architecture, including the C3K2 module, SPPF module, and C2PSA module.

[0091] The backbone network comprises four feature extraction stages (Stage 1-Stage 4), with feature map spatial resolutions of 160×160, 80×80, 40×40, and 20×20, respectively. Following the principle of general step S1, the original C3K2 modules in Stage 2 (80×80) and Stage 3 (40×40) with moderate feature map resolutions are selected as the integration locations for the DDARB module. This ensures that the enhancement module can capture multi-scale contour semantic information while preserving sufficient spatial details.

[0092] S2 module replacement:

[0093] like Figure 5 As shown, the original C3K2 modules (configured as C3K=False) in Stage 2 and the original C3K2 modules (configured as C3K=True) in Stage 3 are completely replaced with DDARB modules. Since the number of input channels, number of output channels, feature map height, and feature map width of the DDARB modules are exactly the same as those of the replaced C3K2 modules, no additional dimension matching layer is needed, and they can be directly embedded into the network. In addition, the original C3K2 modules (C3K=False) in Stage 1 and the C3K2 modules (C3K=True) in Stage 4 of the backbone network are retained unchanged, the SPPF module and C2PSA module retain their original structure, all C3K2 modules, upsampling module, and feature splicing module of the neck network are retained, and all detection modules of the head network retain their original YOLOv11 structure.

[0094] S3 Training and Deployment:

[0095] The YOLOv11 native training configuration was used: input resolution 640×640, optimizer AdamW, initial learning rate 4e-3, batch size 16, and training epochs 200. The DDARB module parameters were jointly optimized with the network backbone, requiring no special adjustments. During training, the network automatically learned how to utilize the enhanced features extracted by the DDARB module to improve the recognition and detection accuracy of multi-scale targets in dynamic distance scenarios. After training, the model can be exported to the standard ONNX format and deployed on edge computing devices (such as the NVIDIA Jetson series) using inference engines like TensorRT to meet real-time requirements.

[0096] This embodiment provides an optimal parameter configuration for integrating the Dual-Domain Collaborative Enhancement Residual Linking Module (DDARB) into the backbone network of a typical target detection network: Let the module input feature map... Where H and W are the height and width of the feature map, and C is the number of channels in the feature map.

[0097] (1) WT-Conv branch parameters:

[0098] 1) Number of decomposition levels: L=2 is preferred;

[0099] 2) Subband convolution kernel: 3×3 depthwise convolution, padding=1, stride=1;

[0100] Number of parameters: (Taking YOLOv11 as an example, C=128 in Stage 2 and Stage 3) = 9216.

[0101] (2) MRF-SAM branch parameters:

[0102] 1) Number of parallel convolution branches: K=3;

[0103] 2) Kernel size and padding:

[0104] 2.1) Branch 1: 3×3 depthwise convolution, padding=1, stride=1;

[0105] 2.2) Branch 2: 7×7 depthwise convolution, padding=3, stride=1;

[0106] 2.3) Branch 3: 11×11 depthwise convolution, padding=5, stride=1;

[0107] 3) SE module compression ratio: r=16;

[0108] Parameter calculation: Each SE module has an additional number of parameters.

[0109] The three SE modules bring an additional total number of parameters. When the number of feature map channels is very large (taking YOLOv11 as an example, C=128 in Stage 2 and Stage 3), the additional total number of parameters... .

[0110] The sum of the parameters in the two branches = If the base network is chosen to be the lightest n-type (YOLOv11-n) of the YOLOv11 series, its parameter count is approximately 2.6 million. Introducing two DDARB modules only increases the number of parameters. The parameter count is approximately 9.4 million if the base network is a lightweight S-type YOLOv11 (YOLOv11-s). Introducing two DDARB modules only increases the parameter count by a small percentage. The percentage of parameters. It can be seen that introducing two DDARB modules into the backbone network only brings a slight increase in parameters, but it brings a significant increase in the receptive field, which is quite considerable.

[0111] Generality Explanation: The integration method described above, using YOLOv11 as an example, essentially embeds the DDARB module as a feature extraction component into the backbone of the object detection network to enhance its feature extraction capability for multi-scale targets, ultimately achieving highly robust multi-scale target recognition and detection in dynamic distance scenarios. Given that the DDARB module has consistent input and output dimensions and does not depend on a specific network structure, this integration template can be extended to other object detection networks. In practical applications, it is only necessary to locate the feature extraction stage with an appropriate feature map resolution based on the backbone structure of the target network, replace the original feature extraction module with the DDARB module, and keep the rest of the network architecture unchanged. The lightweight, easy-to-deploy, and plug-and-play characteristics of the DDARB module enable the method described in this invention to flexibly adapt to various object detection networks, providing a general and efficient feature enhancement solution for solving the problem of multi-scale target recognition and detection in dynamic distance scenarios.

[0112] This embodiment proposes a multi-scale target recognition and detection method for dynamic distance scenarios based on DDARB module feature enhancement. By integrating a DDARB module into the backbone of the target detection network, it provides a general and efficient feature enhancement scheme for solving the multi-scale target recognition and detection problem in dynamic distance scenarios. Any technical solution that integrates a DDARB module into the backbone of the target detection network to enhance multi-scale feature extraction capabilities based on the method described in this embodiment, as well as any modifications, equivalent substitutions, or improvements made by those skilled in the art based on the description of this embodiment without departing from the spirit and principles of this invention, should fall within the protection scope of this invention.

Claims

1. A multi-scale target feature enhancement method based on frequency domain-spatial domain synergy, characterized in that, By replacing the original feature extraction module of the target detection network with the DDARB module, the ability of the target detection network to identify and detect targets at multiple scales is enhanced. The DDARB module adopts a residual connection structure. The main path is an identity mapping, which directly transmits the input feature map to the output. In the enhancement path, the first branch is a wavelet transform convolution branch, which is used to perform multi-level wavelet decomposition and reconstruction on the input feature map to extract frequency domain enhancement features. The second branch is the multi-receptive field spatial aggregation branch, which is used to extract spatial domain enhancement features in parallel through multiple deep convolution kernels of different sizes and to perform adaptive weighting using channel attention; The outputs of the first and second branches are added together, and then element-wise added to the input feature map of the main path to form a complete residual connection, resulting in the output enhanced feature map.

2. The multi-scale target feature enhancement method based on frequency domain-spatial domain synergy according to claim 1, characterized in that, The wavelet transform convolution branch uses the Haar wavelet basis to perform an L-level two-dimensional discrete wavelet transform on the input feature map. Each level of the transform decomposes the feature map into a low-frequency sub-band and three high-frequency sub-bands, respectively extracting low-frequency information and high-frequency information in the horizontal, vertical and diagonal directions of the feature map.

3. The multi-scale target feature enhancement method based on frequency domain-spatial domain synergy according to claim 1, characterized in that, The multi-receptive field spatial aggregation branch combines multi-size large-kernel depthwise convolution with channel attention: it simultaneously feeds the input feature maps into the network. Each branch has a parallel deep convolutional layer, and the output feature map of each branch is recalibrated with channel attention through the channel attention mechanism unit SE Block.

4. The multi-scale target feature enhancement method based on frequency domain-spatial domain synergy according to claim 1, characterized in that, The output enhanced feature map is: in For the input feature map, This is the first enhanced feature map output by the wavelet transform convolution branch. The second enhanced feature map is the output of the multi-receptive field spatial aggregation branch. For the final enhanced feature map, H is the feature map height, W is the feature map width, and C is the number of feature map channels.

Images