A lightweight target detection method based on feature enhancement and multi-branch fusion

Abstract

The present disclosure is a lightweight target detection method based on feature enhancement and multi-branch fusion, comprising: feature extraction and enhancement: extracting features through serial heterogeneous convolution and wavelet convolution, after splicing, compression and attention screening, outputting through residual connection; multi-scale feature fusion; small target path feature enhancement fusion: extracting detail and context information through channel segmentation and double-branch parallel convolution; target classification and positioning. Through serial heterogeneous convolution and wavelet convolution and active fusion screening mechanism, the present embodiment significantly improves the diversity of basic features while reducing the amount of calculation; by increasing an independent pooling branch, global statistics and spatial guidance information are explicitly introduced, breaking through the limitation of traditional double-branch FPN / PAN or BiFPN which only optimizes the fusion weight between convolution features, achieving more sufficient and robust feature fusion with low computational cost; adopting a branch processing strategy, the quality of the input features of the detection head is strengthened, achieving a balance between model precision and speed.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to a lightweight target detection method based on feature enhancement and multi-branch fusion. Background Technology

[0002] Object detection is a core task in computer vision. With the development of deep learning, detection models represented by the YOLO series have made significant progress in accuracy. However, the continuous deepening and widening of network structures in pursuit of higher accuracy has led to problems such as a large number of parameters, high computational complexity, and slow inference speed, making it difficult to deploy directly on mobile devices, embedded terminals, and edge computing devices with limited computing power, memory, and power consumption.

[0003] To adapt to resource-constrained environments, researchers have proposed various lightweight solutions, but the following systemic shortcomings still exist: 1. The contradiction between lightweight model and feature representation capability: Mainstream solutions often use depthwise separable convolutions (such as the MobileNet series) or directly replace the entire backbone network. Although such methods effectively reduce the amount of computation, they severely weaken the network's feature extraction and representation capabilities, resulting in poor basic feature map information in the output and a significant decrease in detection accuracy in complex scenes.

[0004] 2. Information Interaction Bottleneck in Feature Fusion Structures: Existing models generally employ a two-branch structure combining Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) for multi-scale feature fusion. However, in lightweight models, the fixed and limited information flow paths of the two-branch structure make it difficult to achieve sufficient complementarity and integration between deep semantic features and shallow detailed features. Although some studies have introduced Bidirectional Feature Pyramid Network (BiFPN) to optimize connections and weights, it is essentially still an improvement within the two-branch framework and fails to fundamentally overcome the limitations of information sufficiency in fusion.

[0005] 3. Feature degradation problem in lightweight models for small object detection: While reducing model complexity, lightweight operations inevitably weaken the network's feature extraction and preservation capabilities. For small objects, the effective information region they occupy in the feature map is already extremely limited. Lightweighting further exacerbates the loss of detailed information and the inadequacy of semantic representation, resulting in features used for detection that are both "sparse" and "weak".

[0006] Existing lightweight target detection technologies mostly focus on optimizing a single step or replacing existing modules, lacking a systematic design across the entire chain from feature extraction and multi-scale fusion to enhancing specific performance weaknesses, making it difficult to achieve the best balance between accuracy and speed under stringent resource constraints.

[0007] Therefore, an innovative and systematic approach to lightweight architecture design is needed.

[0008] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0009] The purpose of this invention is to provide a lightweight target detection method based on feature enhancement and multi-branch fusion, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.

[0010] This invention first provides a lightweight target detection method based on feature enhancement and multi-branch fusion, comprising the following steps: S1, Feature Extraction and Enhancement: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After splicing, compression and attention filtering, the features are output through residual connection to generate a multi-scale primary feature map. S2, Multi-scale feature fusion: Input the multi-scale primary feature map into the neck network, and perform information interaction and fusion through the top-down path, the bottom-up path and the parallel pooling enhancement branch to output the enhanced multi-scale fused features. S3, Small Target Path Feature Enhancement and Fusion: The feature enhancement and fusion module is used to extract details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and then outputs optimized features after fusion and residual connection. S4, Target Classification and Localization: The optimized features and multi-scale fusion features are input into the corresponding detection head for processing to complete the classification and localization of the target.

[0011] In this invention, S1 specifically includes the following steps: S11, Serial heterogeneous feature extraction: The input features are processed sequentially through a heterogeneous convolutional layer and a wavelet convolutional layer to obtain the first-level features and the second-level features respectively, and the first-level features are retained. S12, Feature concatenation: The first-level feature and the second-level feature are concatenated along the channel dimension to obtain the fused feature; S13, Channel compression and attention filtering: The fused features are reduced in dimensionality and fused across channels using a 1×1 convolutional layer, and then the fused features are calibrated using a channel attention module; S14, Residual Output: The calibrated features are added to the original input features by residual addition to obtain the output features, i.e., the multi-scale primary feature map.

[0012] In this invention, the heterogeneous convolutional layer is a HetConv layer, which uses a hybrid grouped convolutional kernel to reduce the number of parameters while maintaining feature extraction capability; the wavelet convolutional layer is a WTConv layer, which uses wavelet transform basis to extract frequency domain information of input features and capture feature patterns complementary to standard convolution.

[0013] In this invention, the channel attention module is a SENetV2 module, which is used to generate channel weight vectors by modeling the dependencies between channels, and to reweight the input features at the channel level to enhance important features and suppress redundant features.

[0014] In this invention, S2 includes the following steps: S21, Obtain multi-scale input: Receive multi-scale primary feature maps from the backbone network, including deep features, mid-level features and shallow features; S22, Construct parallel pooling enhanced features: Perform global average pooling and global max pooling on the shallow features simultaneously, add and fuse the results to form parallel pooling enhanced features containing global context and spatial saliency information; S23, Three-branch feature fusion: The deep features, mid-level features and parallel pooling enhancement features are spliced ​​together at the same resolution scale to achieve deep fusion of multi-source information and obtain fused features.

[0015] In this invention, step S3 includes the following steps: S31, Channel Segmentation: The shallow features of the backbone network are segmented into a first sub-feature and a second sub-feature along the channel dimension; S32, dual-branch parallel processing: the first sub-feature is processed by a standard convolutional layer to enhance local details; the second sub-feature is processed by a dilated convolutional layer to expand the receptive field and obtain multi-scale contextual information without increasing the number of parameters. S33, Feature Fusion: The features output from the two branches above are concatenated by channels and integrated through a 1×1 convolutional layer; S34, Residual Enhancement: The fused and integrated features are added to the original input features by residual addition to obtain the enhanced optimized features, which are then output to the corresponding small target detection head.

[0016] In this invention, the hollow convolutional layer is a 3×3 hollow convolutional layer with an expansion rate of 1.

[0017] The present invention further provides a lightweight target detection device based on feature enhancement and multi-branch fusion, comprising: Backbone Network Module: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After concatenation, compression and attention filtering, the features are output through residual connections to generate multi-scale primary feature maps. Neck network module: Input multi-scale primary feature maps into the neck network, and perform information interaction and fusion through top-down path, bottom-up path and parallel pooling enhancement branch, and output enhanced multi-scale fused features; Feature enhancement and fusion module: The feature enhancement and fusion module extracts details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and outputs optimized features after fusion and residual connection; Detection head module: Inputs optimized features and multi-scale fusion features into the corresponding detection head for processing to complete the classification and localization of the target.

[0018] The technical solution provided by this invention may include the following beneficial effects: This invention presents a lightweight target detection method based on feature enhancement and multi-branch fusion. By employing serial heterogeneous convolution and wavelet convolution with an active fusion and filtering mechanism, it significantly improves the diversity of basic features while reducing computational load, fundamentally resolving the contradiction of feature degradation caused by lightweight operations. By adding an independent pooling branch, it explicitly introduces global statistics and spatial guidance information, breaking through the limitations of traditional dual-branch FPN / PAN or BiFPN that only optimize the fusion weights between convolutional features, achieving more thorough and robust feature fusion with low computational cost. By adopting a branching processing strategy, it specifically enhances the quality of the input features of the detection head without introducing additional detection branches or significantly increasing computational cost, achieving a balance between model accuracy and speed. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0020] Figure 1 A flowchart illustrating a lightweight target detection method based on feature enhancement and multi-branch fusion in an exemplary embodiment of this disclosure is shown. Figure 2 This diagram illustrates the structural block diagram of a lightweight target detection device based on feature enhancement and multi-branch fusion in an exemplary embodiment of this disclosure. Figure 3 This diagram illustrates the structure of the lightweight feature enhancement fusion module for the backbone network in an exemplary embodiment of this disclosure. Figure 4 This diagram illustrates a three-branch structure of the neck network in an exemplary embodiment of this disclosure. Figure 5 A structural diagram of the feature enhancement fusion module in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0022] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0023] This example implementation first provides a lightweight target detection method based on feature enhancement and multi-branch fusion. Please refer to [link / reference]. Figure 1 This method may include: S1-S4, as follows: S1, Feature Extraction and Enhancement: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After splicing, compression and attention filtering, the features are output through residual connection to generate a multi-scale primary feature map. S2, Multi-scale feature fusion: Input the multi-scale primary feature map into the neck network, and perform information interaction and fusion through the top-down path, the bottom-up path and the parallel pooling enhancement branch to output the enhanced multi-scale fused features. S3, Small Target Path Feature Enhancement: The feature enhancement and fusion module is used to extract details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and then outputs optimized features after fusion and residual connection. S4, Target Classification and Localization: The optimized features and multi-scale fusion features are input into the corresponding detection head for processing to complete the classification and localization of the target.

[0024] In this embodiment, by combining serial heterogeneous convolution and wavelet convolution with an active fusion and filtering mechanism, the diversity of basic features is significantly improved while reducing computational load, fundamentally resolving the contradiction of feature degradation caused by lightweight operations. By adding an independent pooling branch, global statistics and spatial guidance information are explicitly introduced, breaking through the limitation of traditional dual-branch FPN / PAN or BiFPN that only optimizes the fusion weights between convolutional features, achieving more thorough and robust feature fusion with low computational cost. By adopting a split-processing strategy, the quality of the input features of the detection head is specifically enhanced without introducing additional detection branches or significantly increasing computational cost, achieving a balance between model accuracy and speed.

[0025] The specific process of each step in the above embodiments will be described below.

[0026] S1 is the feature extraction and enhancement process, which specifically includes the following steps: S11, Serial heterogeneous feature extraction: The input features are processed sequentially through a heterogeneous convolutional layer and a wavelet convolutional layer to obtain the first-level features and the second-level features respectively, and the first-level features are retained. S12, Feature concatenation: The first-level feature and the second-level feature are concatenated along the channel dimension to obtain the fused feature; S13, Channel compression and attention filtering: The fused features are reduced in dimensionality and fused across channels using a 1×1 convolutional layer, and then the fused features are calibrated using a channel attention module; S14, Residual Output: The calibrated features are added to the original input features by residual addition to obtain the output features, i.e., the multi-scale primary feature map.

[0027] In this step, the input image is processed through a backbone network improved with heterogeneous convolutional layers and wavelet convolutional layers. Each lightweight feature enhancement and fusion module in the network extracts complementary features through serial heterogeneous convolution and wavelet convolution. After fusion, compression, and attention filtering, the features are output through residual connections, ultimately generating a set of multi-scale and information-rich primary feature maps.

[0028] The heterogeneous convolutional layer is a HetConv layer, which uses a hybrid grouped convolutional kernel to reduce the number of parameters while maintaining feature extraction capability; the wavelet convolutional layer is a WTConv layer, which uses wavelet transform basis to extract frequency domain information of input features and capture feature patterns complementary to standard convolution.

[0029] The channel attention module is a SENetV2 module, which is used to generate channel weight vectors by modeling the dependencies between channels, and to reweight the input features at the channel level to enhance important features and suppress redundant features.

[0030] S2 is a multi-scale feature fusion process, which specifically includes the following steps: S21, Obtain multi-scale input: Receive multi-scale primary feature maps from the backbone network, including deep features, mid-level features and shallow features; S22, Construct parallel pooling enhanced features: Perform global average pooling and global max pooling on the shallow features simultaneously, add and fuse the results to form parallel pooling enhanced features containing global context and spatial saliency information; S23, Three-branch feature fusion: The deep features, mid-level features and parallel pooling enhancement features are spliced ​​together at the same resolution scale to achieve deep fusion of multi-source information and obtain fused features.

[0031] In this step, the multi-scale features generated by S1 are input into the improved neck network. First, semantic information is passed through the top-down FPN path; second, global and saliency information of each layer's features is extracted through the newly added parallel pooling branch; finally, combined with the detailed information of the bottom-up PAN path, the features from the three paths are concatenated and fused at the corresponding scales to output the enhanced multi-scale fused features.

[0032] S3 is the process of enhancing and fusing path features for small targets, specifically including the following steps: S31, Channel Segmentation: The shallow features of the backbone network are segmented into a first sub-feature and a second sub-feature along the channel dimension; S32, dual-branch parallel processing: the first sub-feature is processed by a standard convolutional layer to enhance local details; the second sub-feature is processed by a dilated convolutional layer to expand the receptive field and obtain multi-scale contextual information without increasing the number of parameters. S33, Feature Fusion: The features output from the two branches above are concatenated by channels and integrated through a 1×1 convolutional layer; S34, Residual Enhancement: The fused and integrated features are added to the original input features by residual addition to obtain the enhanced optimized features, which are then output to the corresponding small target detection head.

[0033] The hollow convolutional layer is a 3×3 hollow convolutional layer with an expansion rate of 1.

[0034] In this step, the input features are split into two parallel processing paths using a channel splitting (Split) operation. One path is processed through a standard convolutional layer to enhance local details, while the other path uses a dilated convolutional layer to efficiently obtain multi-scale contextual information, so as to preserve spatial details and supplement contextual semantics at the same time. The results are then fused and output to enhance the feature quality of the input detection head and generate optimized features specifically for small object detection.

[0035] S4 takes the final feature maps at each scale after processing by S2 and S3, inputs them into the corresponding detection head, completes the target category classification and bounding box coordinate regression, and outputs the final detection result.

[0036] The lightweight target detection method based on feature enhancement and multi-branch fusion of this application will be further illustrated below through specific embodiments.

[0037] This embodiment presents a lightweight target detection method based on an improved YOLOv8, with the overall network architecture as follows: Figure 2 As shown, the overall network architecture includes a backbone network, a neck network, and a detection head connected in sequence. The target detection method of this application will be described below in conjunction with each component of the overall network architecture.

[0038] This application reconstructs the backbone, neck, and head of the YOLOv8 model, rather than simply replacing or stacking existing modules. Specifically, firstly, a lightweight enhancement module integrating heterogeneous feature extraction, wavelet feature extraction, and attention filtering is designed in the backbone to improve feature quality from the source; secondly, the traditional two-branch fusion paradigm is broken in the neck to construct a three-branch enhancement topology that incorporates global contextual information, achieving more comprehensive multi-scale feature fusion; finally, an efficient feature preprocessing unit is designed before the head for small target detection paths to specifically repair feature degradation caused by lightweighting.

[0039] 1. Implementation method of lightweight feature enhancement fusion module for backbone network See Figure 3 The lightweight feature enhancement and fusion module proposed in this invention replaces the DarknetBottleneck module in the original YOLOv8 network. The input feature map of this module is denoted as follows: ,in C 1 H, W These represent the number of channels, height, and width, respectively.

[0040] (1) Serial heterogeneous feature extraction and intermediate feature retention.

[0041] First, Input a heterogeneous convolutional layer (HetConv), for example, using a 3×3 convolutional kernel, and obtain the first-level feature map through batch normalization (BatchNorm) and activation function (such as SiLU). Here This represents the number of intermediate channels.

[0042] Then Input a wavelet convolutional layer (WTConv), for example, using a 5×5 convolutional kernel, followed by batch normalization and activation functions, to obtain the second-level feature map. Here C 2 represents the preset number of output channels for the module.

[0043] (2) Heterogeneous feature splicing and fusion: The first-level features retained in step S11 are spliced ​​and fused together. With second-level features Concat along the channel dimension to get the number of channels. Fusion characteristics .

[0044] (3) Channel compression and attention filtering: First, a 1×1 standard convolutional layer is used to process the concatenated high-dimensional features. Dimensionality reduction and cross-channel information integration are performed to compress the number of channels to the target output channel number. C 2. Obtaining features Then, Input a SENetV2 attention module for feature calibration, and output the enhanced features. .

[0045] (4) Residual integration: Using residual join, then... With the module's original input Element-by-element addition is performed to obtain the final output feature map of this module. .

[0046] In actual experiments .

[0047] 2. For the implementation method of the neck network three-branch feature enhancement and fusion processing, please refer to [link / reference]. Figure 4 This invention reconstructs the neck network of YOLOv8. Taking one of the fusion nodes as an example, its input typically comes from shallow (high resolution), medium, and deep (low resolution) layers.

[0048] (1) Feature input: Let the features from the shallow layer be... The characteristics from the middle layer are From this floor (Deep) Features after upsampling are .

[0049] (2) Constructing parallel pooling enhancement branches: for Perform the following parallel operations: Perform global average pooling to obtain feature vectors that represent the global context.

[0050] Perform global max pooling to obtain eigenvectors that represent the saliency of the space.

[0051] Add the two pooling results to obtain the enhanced features. .

[0052] (3) Three-branch feature fusion: Deep features from the FPN path are fused together. Original features from the middle layer And enhanced features from the new pooling branch Perform channel splicing and pass it to the next layer.

[0053] 3. For the implementation method of the feature enhancement and fusion module, please refer to [link / reference]. Figure 5 This module is placed in front of the small target detection head.

[0054] (1) Channel segmentation: The feature map input to this module Divide into two parts along the channel dimension at a preset ratio (1:1): and .

[0055] (2) Parallel dual-branch convolution processing: Branch A (Local detail enhancement path): will The input is a standard 3×3 convolutional layer, followed by batch normalization and an activation function (such as SiLU). This path focuses on further refining local spatial details and texture features, and the output is denoted as... .

[0056] Branch B (Efficient Context Extension Path): will Input a dilated convolutional layer. This path, without increasing the number of parameters, fuses multi-scale contextual semantic information by expanding the receptive field of the convolutional kernel, and the output is labeled as follows. .

[0057] (3) Feature compression and fusion: The detailed features output by branch A are compressed and fused. Contextual features of branch B output The concatenated features are then concatenated along the channel dimension. Subsequently, a standard 1×1 convolutional layer is used to integrate cross-channel information and adjust the number of channels, outputting a feature map that blends details and context. .

[0058] (4) Residual integration and output: fusion features This is treated as a "feature enhancement amount" and added element-wise to the input of the original small object detection head. Through this residual connection, the enhancement information extracted by the dual-branch parallel processing is injected while fully preserving the original feature information. Finally, it is fed into the corresponding YOLO small object detection head to complete the object category classification and bounding box coordinate regression, outputting the final detection result.

[0059] This application also provides a lightweight target detection device based on feature enhancement and multi-branch fusion, comprising: Backbone Network Module: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After concatenation, compression and attention filtering, the features are output through residual connections to generate multi-scale primary feature maps. Neck network module: Input multi-scale primary feature maps into the neck network, and perform information interaction and fusion through top-down path, bottom-up path and parallel pooling enhancement branch, and output enhanced multi-scale fused features; Feature enhancement and fusion module: The feature enhancement and fusion module extracts details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and outputs optimized features after fusion and residual connection; Detection head module: Inputs optimized features and multi-scale fusion features into the corresponding detection head for processing to complete the classification and localization of the target.

[0060] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0061] The effectiveness of the target detection method of this application will be explained below through ablation experiments.

[0062] The core objective of the ablation experiment is to verify the independent effectiveness of the three improved modules: LFEM (Lightweight Feature Enhancement Fusion Module), PET-FPN (Pooling Enhancement Three-Branch FPN), and LDC-EM (Lightweight Dilated Convolutional Context Enhancement Module), to clarify the impact of each module on the model's "detection accuracy - lightweighting level," and to determine the optimal module combination scheme. This experiment uses the original YOLOv8n model as a baseline and employs a controlled variable method to design six experimental groups. The specific experimental group settings are as follows: Experimental group 1 (baseline group): Original YOLOv8n model, without any improved modules; Experimental Group 2: YOLOv8n + LFEM (only the backbone network was improved); Experimental group 3: YOLOv8n + PET-FPN (only the neck network was improved); Experimental group 4: YOLOv8n + LDC-EM (only improving the enhancement mechanism for small targets in front of the detection head); Experimental group 5: YOLOv8n + LFEM + PET-FPN module (improved backbone and neck network); Experimental group 6: YOLOv8n + LFEM + PET-FPN + LDC-EM (three-module integrated final improved model).

[0063] The experiment used the number of parameters, GFLOPs, and Model Size to characterize the lightweight nature of the model, and mAP50 and mAP50:95 to characterize the detection accuracy. The effectiveness of each module was validated by comparing the core indicator data from six sets of experiments. The ablation experiment results are shown in Table 1.

[0064] Table 1 Ablation Experiment Results

[0065] As shown in Table 1, the LFEM module, employing a combination of heterogeneous convolution and wavelet convolution, reduced the number of model parameters and computational cost by 15.0% and 14.8% respectively, while maintaining the detection accuracy of the benchmark model (mAP50:95, even improving by 0.1%), successfully achieving lightweighting in the feature extraction stage. The PET-FPN module, by adding a shallow pooling enhancement branch, improved the overall detection accuracy (mAP50:95) by 0.8% at the cost of less than 0.3% additional parameters, effectively solving the information bottleneck problem in multi-scale feature fusion. The LDC-EM module utilizes dilated convolution to expand the receptive field, focusing on enhancing the feature representation ability of small targets, achieving a significant improvement of 1.0% in overall detection accuracy (mAP50:95) at a minimal cost of only 2.3% increase in parameters.

[0066] It should be noted that although several modules of the system for executing actions are mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into multiple modules for embodiment. Components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0067] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0068] It should be noted that the installation of image acquisition and personal identification equipment in public places involved in this application is necessary for maintaining public safety, complies with relevant national regulations, and is accompanied by prominent warning signs. The collected personal images and identification information can only be used for the purpose of maintaining public safety and will not be used for other purposes; or the images, personal identification data, etc. in this application are all legally and compliantly obtained or collected with the individual's separate consent.

Claims

1. A lightweight target detection method based on feature enhancement and multi-branch fusion, characterized in that, Includes the following steps: S1, Feature Extraction and Enhancement: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After splicing, compression and attention filtering, the features are output through residual connection to generate a multi-scale primary feature map. S2, Multi-scale feature fusion: Input the multi-scale primary feature map into the neck network, and perform information interaction and fusion through the top-down path, the bottom-up path and the parallel pooling enhancement branch to output the enhanced multi-scale fused features. S3, Small Target Path Feature Enhancement and Fusion: The feature enhancement and fusion module is used to extract details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and then outputs optimized features after fusion and residual connection. S4, Target Classification and Localization: The optimized features and multi-scale fusion features are input into the corresponding detection head for processing to complete the classification and localization of the target.

2. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 1, characterized in that, S1 specifically includes the following steps: S11, Serial heterogeneous feature extraction: The input features are processed sequentially through a heterogeneous convolutional layer and a wavelet convolutional layer to obtain the first-level features and the second-level features respectively, and the first-level features are retained. S12, Feature concatenation: The first-level feature and the second-level feature are concatenated along the channel dimension to obtain the fused feature; S13, Channel compression and attention filtering: The fused features are reduced in dimensionality and fused across channels using a 1×1 convolutional layer, and then the fused features are calibrated using a channel attention module; S14, Residual Output: The calibrated features are added to the original input features by residual addition to obtain the output features, i.e., the multi-scale primary feature map.

3. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 2, characterized in that, The heterogeneous convolutional layer is a HetConv layer, which uses hybrid grouped convolutional kernels to reduce the number of parameters while maintaining feature extraction capability; the wavelet convolutional layer is a WTConv layer, which uses wavelet transform basis to extract frequency domain information of input features and capture feature patterns complementary to standard convolution.

4. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 2, characterized in that, The channel attention module is a SENetV2 module, which is used to generate channel weight vectors by modeling the dependencies between channels, and to reweight the input features at the channel level to enhance important features and suppress redundant features.

5. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 1, characterized in that, S2 includes the following steps: S21, Obtain multi-scale input: Receive multi-scale primary feature maps from the backbone network, including deep features, mid-level features and shallow features; S22, Construct parallel pooling enhanced features: Perform global average pooling and global max pooling on the shallow features simultaneously, add and fuse the results to form parallel pooling enhanced features containing global context and spatial saliency information; S23, Three-branch feature fusion: The deep features, mid-level features and parallel pooling enhancement features are spliced ​​together at the same resolution scale to achieve deep fusion of multi-source information and obtain fused features.

6. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 1, characterized in that, S3 includes the following steps: S31, Channel Segmentation: The shallow features of the backbone network are segmented into a first sub-feature and a second sub-feature along the channel dimension; S32, dual-branch parallel processing: the first sub-feature is processed by a standard convolutional layer to enhance local details; the second sub-feature is processed by a dilated convolutional layer to expand the receptive field and obtain multi-scale contextual information without increasing the number of parameters. S33, Feature Fusion: The features output from the two branches above are concatenated by channels and integrated through a 1×1 convolutional layer; S34, Residual Enhancement: The fused and integrated features are added to the original input features by residual addition to obtain the enhanced optimized features, which are then output to the corresponding small target detection head.

7. The lightweight target detection method based on feature enhancement and multi-branch fusion according to claim 6, characterized in that, The hollow convolutional layer is a 3×3 hollow convolutional layer with an expansion rate of 1.

8. A lightweight target detection device based on feature enhancement and multi-branch fusion, characterized in that, include: Backbone Network Module: The input image is processed by the lightweight feature enhancement and fusion module in the backbone network to extract features through serial heterogeneous convolution and wavelet convolution. After concatenation, compression and attention filtering, the features are output through residual connections to generate multi-scale primary feature maps. Neck network module: Input multi-scale primary feature maps into the neck network, and perform information interaction and fusion through top-down path, bottom-up path and parallel pooling enhancement branch, and output enhanced multi-scale fused features; Feature enhancement and fusion module: The feature enhancement and fusion module extracts details and contextual information from the shallow features of the backbone network through channel segmentation and dual-branch parallel convolution, and outputs optimized features after fusion and residual connection; Detection head module: Inputs optimized features and multi-scale fusion features into the corresponding detection head for processing to complete the classification and localization of the target.

Images