Target detection method and system based on feature aggregation

Abstract

The application discloses a target detection method and system based on feature aggregation, comprising the following steps: acquiring image data information corresponding to a to-be-detected image and pre-processing the image data information to construct a training data set; constructing an initial network for target detection based on feature aggregation based on a convolution scheme, a pooling scheme, a double-channel scheme, a spatial projection scheme and a channel projection scheme, and training the initial network to obtain a target detection network based on feature aggregation; and using the obtained target detection network based on feature aggregation to detect a target in the to-be-detected image. The application can not only realize target detection based on feature aggregation, but also has higher reliability and better accuracy.

Description

Technical Field

[0001] This invention belongs to the field of image processing, specifically relating to a target detection method and system based on feature aggregation. Background Technology

[0002] With the development of economy and technology, target detection solutions have been widely used in production and daily life. However, the increasing application of industrial IoT and edge computing scenarios has brought serious challenges to target detection solutions.

[0003] Traditional object detection schemes based on deep convolutional neural networks (CNNs), while achieving high detection accuracy, suffer from a large number of parameters and computational load that are difficult to adapt to the computational constraints of embedded chips (such as DSPs, FPGAs, or mobile GPUs). Researchers have proposed lightweight solutions for deep CNNs, including using depthwise separable convolutions or channel pruning strategies. While these lightweight solutions improve the speed of object detection, they also introduce other problems: First, to reduce computation, lightweight networks tend to use small convolutional kernels, which leads to a lack of macroscopic control over large-scale targets and difficulty in capturing long-distance pixel dependencies, resulting in decreased accuracy. Second, in the multi-scale feature fusion stage, lightweight networks ignore the differences in information content among features at different levels, causing target features to be overwhelmed by background noise, thus reducing the reliability of the detection results. Summary of the Invention

[0004] One of the objectives of this invention is to provide a highly reliable and accurate target detection method based on feature aggregation.

[0005] The second objective of this invention is to provide a system for implementing the target detection method based on feature aggregation.

[0006] The target detection method based on feature aggregation provided by this invention includes the following steps:

[0007] S1. Obtain image data information corresponding to the target image to be detected;

[0008] S2. Preprocess the image data information obtained in step S1 to construct a training dataset;

[0009] S3. Construct an initial network for target detection based on feature aggregation using convolution, pooling, dual-channel, spatial projection, and channel projection schemes;

[0010] The constructed target detection initial network based on feature aggregation includes an input processing module, a dual-domain rectification and aggregation module, a signal rectification module, and a target projection module connected in sequence.

[0011] An input processing module is constructed based on a convolution scheme; the input processing module is used to downsample the input image and upload the processing result to the dual-domain rectification and aggregation module;

[0012] A dual-domain rectification and aggregation module is constructed based on convolution, pooling, and dual-channel schemes. The dual-domain rectification and aggregation module extracts local features of the input data through micro-branch and global features of the input data through macro-branch. The feature extraction of the input data is achieved by aggregating local and global features.

[0013] A signal rectification module is constructed based on convolution and pooling schemes; the signal rectification module is used to implement boundary compensation and background suppression of input data;

[0014] A target projection module is constructed based on convolution, spatial projection, and channel projection schemes. The target projection module is used to classify targets and detect their bounding boxes through spatial and channel projection.

[0015] S4. Using the training dataset constructed in step S2, train the initial target detection network based on feature aggregation constructed in step S3 to obtain the target detection network based on feature aggregation.

[0016] S5. Use the feature aggregation-based target detection network obtained in step S4 to perform target detection on the image to be detected.

[0017] The input processing module's processing procedure includes the following steps:

[0018] The input processing module includes a first input downsampling layer and a second output downsampling layer connected in series.

[0019] The first input downsampling layer and the second output downsampling layer have the same structure; the input downsampling layer is a convolutional layer.

[0020] The input downsampling layer is used to downsample the input image.

[0021] The processing steps of the dual-domain rectification and aggregation module include the following:

[0022] The dual-domain rectification aggregation module includes micro-branch and macro-branch;

[0023] The micro-branch includes a first micro-downsampling layer, a first micro-dual-domain rectification and aggregation layer, a second micro-downsampling layer, a second micro-dual-domain rectification and aggregation layer, a third micro-downsampling layer, a third micro-dual-domain rectification and aggregation layer, a fourth micro-downsampling layer, and a fifth micro-downsampling layer;

[0024] The macroscopic branch includes the first macroscopic downsampling layer, the first macroscopic dual-domain rectification and aggregation layer, the second macroscopic downsampling layer, the second macroscopic dual-domain rectification and aggregation layer, the third macroscopic downsampling layer, the third macroscopic dual-domain rectification and aggregation layer, the fourth macroscopic downsampling layer, and the fifth macroscopic downsampling layer;

[0025] The input of the dual-domain rectification and aggregation module is processed sequentially through a first microscopic downsampling layer and a first microscopic dual-domain rectification and aggregation layer to obtain a first microscopic feature. The first microscopic feature is then processed sequentially through a second microscopic downsampling layer and a second microscopic dual-domain rectification and aggregation layer to obtain a second microscopic feature. The second microscopic feature and the second macroscopic feature are cascaded and then processed sequentially through a third microscopic downsampling layer and a third microscopic dual-domain rectification and aggregation layer to obtain a third microscopic feature. The first microscopic feature is then processed through a fourth microscopic downsampling layer to obtain a first downsampled microscopic feature. The second microscopic feature is then processed through a fifth microscopic downsampling layer to obtain a second downsampled microscopic feature. The first downsampled microscopic feature, the second downsampled microscopic feature, and the third microscopic feature are summed element-wise to obtain the microscopic branch output feature.

[0026] The input of the dual-domain rectification and aggregation module is processed sequentially through a first macroscopic downsampling layer and a first macroscopic dual-domain rectification and aggregation layer to obtain a first macroscopic feature. The first macroscopic feature is then processed sequentially through a second macroscopic downsampling layer and a second macroscopic dual-domain rectification and aggregation layer to obtain a second macroscopic feature. The second macroscopic feature and the second microscopic feature are cascaded and then processed sequentially through a third macroscopic downsampling layer and a third macroscopic dual-domain rectification and aggregation layer to obtain a third macroscopic feature. The first macroscopic feature is then processed through a fourth macroscopic downsampling layer to obtain a first downsampled macroscopic feature. The second macroscopic feature is then processed through a fifth macroscopic downsampling layer to obtain a second downsampled macroscopic feature. The first downsampled macroscopic feature, the second downsampled macroscopic feature, and the third macroscopic feature are summed element-wise to obtain the macroscopic branch output feature.

[0027] The micro-branch output characteristics and the macro-branch output characteristics are used together as the output of the dual-domain rectification aggregation module;

[0028] The structures of the first microscopic downsampling layer to the fifth microscopic downsampling layer and the first macroscopic downsampling layer to the fifth macroscopic downsampling layer are all the same; the downsampling layers are convolutional layers.

[0029] The structures of the first microscopic dual-domain rectification aggregation layer to the third microscopic dual-domain rectification aggregation layer and the first macroscopic dual-domain rectification aggregation layer to the third macroscopic dual-domain rectification aggregation layer are all identical; the processing procedure of the dual-domain rectification aggregation layer includes the following steps:

[0030] Dual-domain rectifier aggregation layer includes Convolutional layers Hollow convolutional layer, signal rectification layer, and dual-domain feature aggregation layer;

[0031] The input data is divided into two paths: one path passes through... The convolutional layer processes the data to obtain dual-domain rectified aggregated micro-features; another path is passed through... The dilated convolutional layer is used to process the macroscopic features of the dual-domain rectification aggregation. The microscopic features and macroscopic features of the dual-domain rectification aggregation are input together into the signal rectification layer for statistical description and cross-rectification. Then, the dual-domain feature aggregation layer performs detail extraction, semantic capture, geometric alignment and recombination to obtain the output of the dual-domain rectification aggregation layer.

[0032] The signal rectifier layer and the signal rectifier module have the same structure; the processing includes the following steps:

[0033] The input data includes microscopic features and macro characteristics ;

[0034] Microscopic features Passing in sequence The micro-rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ;

[0035] Macro characteristics Passing in sequence The macroscopic rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ;

[0036] By cascading micro and macro features along the channel dimension, composite features are obtained. ;

[0037] Composite features The micro-modulation coefficients are obtained through processing by the global descriptor generation layer. and macro modulation coefficient ;

[0038] The following formula is used for parallel parameter processing to obtain micro-cross features. and macroscopic cross-features :

[0039] In the formula This is element-wise multiplication;

[0040] Micro-intersection features with microscopic features After performing element-wise multiplication, then combining with microscopic features By adding elements one by one, the microscopic composite rectification characteristics are obtained;

[0041] Macro-intersection characteristics Macro characteristics After performing element-wise multiplication, then combine with macroscopic features. By adding elements one by one, the macroscopic composite rectification characteristics are obtained;

[0042] After cascading the micro-composite rectification features and macro-composite rectification features along the channel dimension, and then through... The convolutional layer processes the data to obtain the output data.

[0043] The processing steps of the global descriptor generation layer include the following:

[0044] The input data is divided into three paths: the first path is processed by a global average pooling layer to obtain the first sub-feature; the second path is processed by a global max pooling layer to obtain the second sub-feature; the third path is processed by a standard deviation statistics layer to obtain the third sub-feature; the first, second, and third sub-features are concatenated along the channel dimension to obtain the comprehensive sub-feature.

[0045] The synthesized sub-features are processed sequentially through a first fully connected layer, a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer, and then channel splitting is performed to obtain the micro-modulation coefficients. and macro modulation coefficient .

[0046] The processing steps of the dual-domain feature aggregation layer include the following:

[0047] The input features are divided into two paths;

[0048] First path of input features After processing by the convolutional layer, the features are divided into two paths: the first path of aggregate features is processed by... The deep convolutional layer processes the first aggregated sub-features; the second aggregated sub-features are then processed by... The deep convolutional layer processes the second aggregated sub-feature; the first aggregated sub-feature and the second aggregated sub-feature are added together to obtain the first dual-domain aggregated feature;

[0049] The second path of the input features is further divided into three sub-features. The first path of features uses features with a dilation rate of 1. Dilated convolution is used to process the data, resulting in the first dilated sub-feature; the second feature is obtained through a dilation rate of 2. Dilated convolution is used to obtain the second type of dilated sub-features; the third type of feature is obtained through a dilation rate of 3. The dilated convolution is processed to obtain the third dilated sub-feature; Softmax normalized weights are calculated for the first, second and third dilated sub-features respectively, and the three dilated sub-features are weighted and summed according to the obtained weights to obtain the weighted fused features; the weighted fused features are then input into the deformable convolutional layer for processing to obtain the dilated aggregate features.

[0050] The void aggregation feature is divided into two paths: the first path void aggregation feature is processed through a global max pooling layer to obtain the first pooled void aggregation feature; the second path void aggregation feature is processed through a global average pooling layer to obtain the second pooled void aggregation feature; the first and second pooled void aggregation features are then summed element-wise, and then sequentially processed... Spatial attention features are obtained by processing convolutional layers and sigmoid activation function layers.

[0051] After multiplying the spatial attention features and the first-path dual-domain aggregation features element-wise, the spatial aggregation features are obtained.

[0052] After summing the spatial aggregation features, the hole aggregation features, and the input features element-wise, then... The convolutional layer is subjected to projection processing to obtain the output of the dual-domain feature aggregation layer.

[0053] The processing procedure of the target projection module includes the following steps:

[0054] The input features are processed through the spatial projection branch to obtain spatial projection features; the input features are also processed through the channel projection branch to obtain channel projection features.

[0055] The spatial projection branch processing includes: the input features are divided into two paths, one of which passes through... The deep convolutional layer processes the data to obtain the first spatial projection sub-features; the other path passes through... The second spatial projection sub-feature is obtained by deep convolutional layer processing; the first and second spatial projection sub-features are concatenated along the channel dimension, and then processed... The convolutional layer is used to process the spatial projection features. ;

[0056] The channel projection branch processing includes: the input features are processed by a normalization layer and then divided into three paths. The first path is... The convolutional layer processes the data to obtain the first channel projection sub-feature, and the second path passes through... The convolutional layer processes the data to obtain the second channel projection sub-features, and the third path passes through... The convolutional layer processes the data to obtain the third channel projection feature; the first, second, and third channel projection features are then added element-wise to obtain the channel projection feature. ;

[0057] After cascading the spatial projection features and channel projection features along the channel dimension, they are then sequentially passed through... The convolutional layer and sigmoid activation function layer are used to process the data to obtain the gated fusion matrix. ;

[0058] Adopting a gated fusion matrix The spatial projection features and channel projection features are fused to obtain the aggregated projection features. , represented as ;

[0059] After summing the aggregated projection features and spatial projection features element-wise, they are then sequentially passed through the first... Convolutional layer and second The convolutional layer processes the data to obtain the bounding box detection results of the target.

[0060] After summing the aggregated projection features and channel projection features element-wise, they are then sequentially passed through the third... Convolutional layer, fourth The target classification result is obtained by processing the data through convolutional layers and softmax activation function layers.

[0061] The training described in step S4 specifically includes the following steps:

[0062] The following formula is used as the loss function for the training process:

[0063] In the formula The value of the loss function; This is the first weight value set. For classification loss; This is the set second weight value; This represents the bounding box overlap loss. This is the set third weight value; For distributed regression loss; This is the fourth weight value that is set; For orthogonal constraint loss, and , This represents a vectorization mapping, used to flatten the input feature tensor into a one-dimensional vector. Represents the inner product between vectors. Let L2 be the norm of the vector; This is a set minimum number to prevent the denominator from being 0.

[0064] This invention also provides a system for implementing the feature aggregation-based target detection method, comprising a data acquisition module, a data processing module, a network construction module, a network training module, and a target detection module; the data acquisition module, data processing module, network construction module, network training module, and target detection module are connected in series; the data acquisition module is used to acquire image data information corresponding to the target image to be detected and upload the data information to the data processing module; the data processing module is used to preprocess the acquired image data information according to the received data information to construct a training dataset and upload the data information to the network construction module; the network construction module is used to construct an initial target detection network based on feature aggregation according to the received data information, based on convolution schemes, pooling schemes, dual-channel schemes, spatial projection schemes, and channel projection schemes, and upload the data information to the network training module; the constructed initial target detection network based on feature aggregation includes an input processing module, a dual-domain rectification aggregation module, a signal rectification module, and a target projection module connected in series; the input processing module is constructed based on the convolution scheme; the input processing module is used to process the input... The image undergoes downsampling processing, and the processing result is uploaded to the dual-domain rectification and aggregation module. The dual-domain rectification and aggregation module is constructed based on convolution, pooling, and dual-channel schemes. This module extracts local features from the input data through micro-branching and global features through macro-branching, achieving feature extraction by aggregating local and global features. A signal rectification module is constructed based on convolution and pooling schemes to implement boundary compensation and background suppression of the input data. A target projection module is constructed based on convolution, spatial projection, and channel projection schemes. This module classifies targets and detects their bounding boxes through spatial and channel projection. A network training module trains the initial target detection network based on feature aggregation using the constructed training dataset, obtaining a feature aggregation-based target detection network, and uploads the data to the target detection module. The target detection module then uses the obtained feature aggregation-based target detection network to perform target detection on the image based on the received data.

[0065] The target detection method and system based on feature aggregation provided by this invention constructs a dual-domain rectification aggregation module based on convolution, pooling, and dual-channel schemes to improve the complexity and expressive power of feature modeling; constructs a signal rectification module based on convolution and pooling schemes to improve the precision of cross-level feature redistribution; and constructs a target projection module based on convolution, spatial projection, and channel projection schemes to reduce gradient interference caused by task conflicts. Therefore, the method of this invention can not only achieve target detection based on feature aggregation, but also has higher reliability and better accuracy. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0067] Figure 2 This is a schematic diagram of the target detection network based on feature aggregation according to the method of the present invention.

[0068] Figure 3 This is a schematic diagram of the structure of the dual-domain rectification and aggregation module of the present invention.

[0069] Figure 4 This is a schematic diagram of the structure of the dual-domain rectifier polymerization layer of the present invention.

[0070] Figure 5 This is a schematic diagram of the signal rectification module of the method of the present invention.

[0071] Figure 6 This is a schematic diagram of the structure of the global descriptor generation layer in the method of the present invention.

[0072] Figure 7 This is a schematic diagram of the structure of the dual-domain feature aggregation layer of the method of the present invention.

[0073] Figure 8 This is a schematic diagram of the target projection module of the method of the present invention.

[0074] Figure 9 This is a schematic diagram illustrating the implementation effect of the method of the present invention on the PASCAL VOC dataset; wherein, Figure 9 (a) is a schematic diagram of target detection in a scene with a dog and a bicycle. Figure 9 (b) is a schematic diagram of bird target detection. Figure 9 (c) is a schematic diagram of target detection for multiple horses.

[0075] Figure 10 This is a schematic diagram illustrating the implementation effect of the method of the present invention on the MS COCO dataset; wherein, Figure 10 (a) is a schematic diagram of target detection in a scene with a giraffe and a zebra. Figure 10 (b) is a schematic diagram of target detection in a bus and pedestrian scene. Figure 10 (c) is a schematic diagram of target detection in a scene of kites and people on the beach.

[0076] Figure 11 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0077] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The target detection method based on feature aggregation disclosed in this invention includes the following steps:

[0078] S1. Obtain image data information corresponding to the target image to be detected; for example, if target detection is required for various scenes, animals and people, the MS COCO standard dataset can be used, or the PASCAL VOC dataset can be used.

[0079] S2. Preprocess the image data information obtained in step S1 to construct a training dataset; the preprocessing process includes operations such as image cleaning and image enhancement;

[0080] S3. Construct an initial network for target detection based on feature aggregation using convolution, pooling, dual-channel, spatial projection, and channel projection schemes;

[0081] The constructed initial network for target detection based on feature aggregation includes an input processing module, a dual-domain rectification and aggregation module, a signal rectification module, and a target projection module connected in sequence, as shown in the structure below. Figure 2 As shown;

[0082] An input processing module is constructed based on a convolution scheme; the input processing module is used to downsample the input image and upload the processing result to the dual-domain rectification and aggregation module;

[0083] A dual-domain rectification and aggregation module is constructed based on convolution, pooling, and dual-channel schemes. The dual-domain rectification and aggregation module extracts local features of the input data through micro-branch and global features of the input data through macro-branch. The feature extraction of the input data is achieved by aggregating local and global features.

[0084] A signal rectification module is constructed based on convolution and pooling schemes; the signal rectification module is used to implement boundary compensation and background suppression of input data;

[0085] A target projection module is constructed based on convolution, spatial projection, and channel projection schemes. The target projection module is used to classify targets and detect their bounding boxes through spatial and channel projection.

[0086] In practice, the input processing module's processing procedure includes the following steps:

[0087] The input processing module includes a first input downsampling layer and a second output downsampling layer connected in series.

[0088] The first input downsampling layer and the second output downsampling layer have the same structure; the input downsampling layer is a convolutional layer.

[0089] The input downsampling layer is used to downsample the input image;

[0090] In practice, the processing steps of the dual-domain rectification and aggregation module include the following:

[0091] The structure of the dual-domain rectifier aggregation module is as follows: Figure 3 As shown;

[0092] The dual-domain rectification aggregation module includes micro-branch and macro-branch;

[0093] The micro-branch includes a first micro-downsampling layer, a first micro-dual-domain rectification and aggregation layer, a second micro-downsampling layer, a second micro-dual-domain rectification and aggregation layer, a third micro-downsampling layer, a third micro-dual-domain rectification and aggregation layer, a fourth micro-downsampling layer, and a fifth micro-downsampling layer;

[0094] The macroscopic branch includes the first macroscopic downsampling layer, the first macroscopic dual-domain rectification and aggregation layer, the second macroscopic downsampling layer, the second macroscopic dual-domain rectification and aggregation layer, the third macroscopic downsampling layer, the third macroscopic dual-domain rectification and aggregation layer, the fourth macroscopic downsampling layer, and the fifth macroscopic downsampling layer;

[0095] The input of the dual-domain rectification and aggregation module is processed sequentially through a first microscopic downsampling layer and a first microscopic dual-domain rectification and aggregation layer to obtain a first microscopic feature. The first microscopic feature is then processed sequentially through a second microscopic downsampling layer and a second microscopic dual-domain rectification and aggregation layer to obtain a second microscopic feature. The second microscopic feature and the second macroscopic feature are cascaded and then processed sequentially through a third microscopic downsampling layer and a third microscopic dual-domain rectification and aggregation layer to obtain a third microscopic feature. The first microscopic feature is then processed through a fourth microscopic downsampling layer to obtain a first downsampled microscopic feature. The second microscopic feature is then processed through a fifth microscopic downsampling layer to obtain a second downsampled microscopic feature. The first downsampled microscopic feature, the second downsampled microscopic feature, and the third microscopic feature are summed element-wise to obtain the microscopic branch output feature.

[0096] The input of the dual-domain rectification and aggregation module is processed sequentially through a first macroscopic downsampling layer and a first macroscopic dual-domain rectification and aggregation layer to obtain a first macroscopic feature. The first macroscopic feature is then processed sequentially through a second macroscopic downsampling layer and a second macroscopic dual-domain rectification and aggregation layer to obtain a second macroscopic feature. The second macroscopic feature and the second microscopic feature are cascaded and then processed sequentially through a third macroscopic downsampling layer and a third macroscopic dual-domain rectification and aggregation layer to obtain a third macroscopic feature. The first macroscopic feature is then processed through a fourth macroscopic downsampling layer to obtain a first downsampled macroscopic feature. The second macroscopic feature is then processed through a fifth macroscopic downsampling layer to obtain a second downsampled macroscopic feature. The first downsampled macroscopic feature, the second downsampled macroscopic feature, and the third macroscopic feature are summed element-wise to obtain the macroscopic branch output feature.

[0097] The micro-branch output characteristics and the macro-branch output characteristics are used together as the output of the dual-domain rectification aggregation module;

[0098] The structures of the first microscopic downsampling layer to the fifth microscopic downsampling layer and the first macroscopic downsampling layer to the fifth macroscopic downsampling layer are all the same; the downsampling layers are convolutional layers.

[0099] The structures of the first microscopic dual-domain rectification and aggregation layer to the third microscopic dual-domain rectification and aggregation layer, and the first macroscopic dual-domain rectification and aggregation layer to the third macroscopic dual-domain rectification and aggregation layer are all the same, such as... Figure 4 As shown, the processing procedure for the dual-domain rectifier aggregation layer includes the following steps:

[0100] Dual-domain rectifier aggregation layer includes Convolutional layers Hollow convolutional layer, signal rectification layer, and dual-domain feature aggregation layer;

[0101] The input data is divided into two paths: one path passes through... The convolutional layer processes the data to obtain dual-domain rectified aggregated micro-features; another path is passed through... The dilated convolutional layer is used to process the macroscopic features of the dual-domain rectified aggregation. The microscopic features and macroscopic features of the dual-domain rectified aggregation are input together into the signal rectified layer for statistical description and cross-rectification. Then, the dual-domain feature aggregation layer is used for detail extraction, semantic capture, geometric alignment and recombination to obtain the output of the dual-domain rectified aggregation layer.

[0102] In practice, the signal rectifier layer and the signal rectifier module have the same structure; the processing includes the following steps:

[0103] The input data includes microscopic features and macro characteristics ;

[0104] Microscopic features Passing in sequence The micro-rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ;

[0105] Macro characteristics Passing in sequence The macroscopic rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ;

[0106] By cascading micro and macro features along the channel dimension, composite features are obtained. ;

[0107] Composite features The micro-modulation coefficients are obtained through processing by the global descriptor generation layer. and macro modulation coefficient ;

[0108] The following formula is used for parallel parameter processing to obtain micro-cross features. and macroscopic cross-features :

[0109] In the formula This is element-wise multiplication;

[0110] Micro-intersection features with microscopic features After performing element-wise multiplication, then combining with microscopic features By adding elements one by one, the microscopic composite rectification characteristics are obtained;

[0111] Macro-intersection characteristics Macro characteristics After performing element-wise multiplication, then combine with macroscopic features. By adding elements one by one, the macroscopic composite rectification characteristics are obtained;

[0112] After cascading the micro-composite rectification features and macro-composite rectification features along the channel dimension, and then through... The convolutional layer processes the data to obtain the output.

[0113] In practice, the global descriptor generation layer's processing steps include the following:

[0114] The structure of the global descriptor generation layer is as follows: Figure 6 As shown;

[0115] The input data is divided into three paths: the first path is processed by a global average pooling layer to obtain the first sub-feature; the second path is processed by a global max pooling layer to obtain the second sub-feature; the third path is processed by a standard deviation statistics layer to obtain the third sub-feature; the first, second, and third sub-features are concatenated along the channel dimension to obtain the comprehensive sub-feature.

[0116] The synthesized sub-features are processed sequentially through a first fully connected layer, a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer, and then channel splitting is performed to obtain the micro-modulation coefficients. and macro modulation coefficient ;

[0117] This invention proposes a signal rectification layer or signal rectification module that includes a global descriptor generation layer and introduces a composite modulation mechanism of "multi-statistical description - coefficient decoupling - cross-channel rectification - spatially assisted rectification". First, it extracts the average response, extreme response, and fluctuation response from the input dual-stream features to form a joint global descriptor containing first-order and second-order statistical information. Then, through a bottleneck mapping network, the global descriptor is split into two sets of modulation coefficients for detail paths and semantic paths, and the dual-stream features are cross-weighted by combining the spatial rectification map generated from the opposite side features. This mechanism not only realizes the boundary compensation of shallow textures for deep semantics, but also realizes the background suppression of shallow details by deep semantics, thereby improving the precision of cross-level feature redistribution.

[0118] In practice, the processing steps of the dual-domain feature aggregation layer include the following:

[0119] The structure of the dual-domain feature aggregation layer is as follows: Figure 7 As shown;

[0120] The input features are divided into two paths;

[0121] First path of input features After processing by the convolutional layer, the features are divided into two paths: the first path of aggregate features is processed by... The deep convolutional layer processes the first aggregated sub-features; the second aggregated sub-features are then processed by... The deep convolutional layer processes the second aggregated sub-feature; the first aggregated sub-feature and the second aggregated sub-feature are added together to obtain the first dual-domain aggregated feature;

[0122] The second path of the input features is further divided into three sub-features. The first path of features uses features with a dilation rate of 1. Dilated convolution is used to process the data, resulting in the first dilated sub-feature; the second feature is obtained through a dilation rate of 2. Dilated convolution is used to obtain the second type of dilated sub-features; the third type of feature is obtained through a dilation rate of 3. The dilated convolution is processed to obtain the third dilated sub-feature; Softmax normalized weights are calculated for the first, second and third dilated sub-features respectively, and the three dilated sub-features are weighted and summed according to the obtained weights to obtain the weighted fused features; the weighted fused features are then input into the deformable convolutional layer for processing to obtain the dilated aggregate features.

[0123] The void aggregation feature is divided into two paths: the first path void aggregation feature is processed through a global max pooling layer to obtain the first pooled void aggregation feature; the second path void aggregation feature is processed through a global average pooling layer to obtain the second pooled void aggregation feature; the first and second pooled void aggregation features are then summed element-wise, and then sequentially processed... Spatial attention features are obtained by processing convolutional layers and sigmoid activation function layers.

[0124] After multiplying the spatial attention features and the first-path dual-domain aggregation features element-wise, the spatial aggregation features are obtained.

[0125] After summing the spatial aggregation features, the hole aggregation features, and the input features element-wise, then... The convolutional layer is subjected to projection processing to obtain the output of the dual-domain feature aggregation layer;

[0126] The dual-domain feature aggregation layer proposed in this invention is based on the construction of two differentiated paths: one is the semantic alignment flow, which establishes a hierarchical receptive field through multi-dilation convolution groups and then uses deformable convolution (DCN) to perform geometric deformation correction on macroscopic semantic features to ensure the semantic localization accuracy of large-scale targets and non-rigid targets; the other is the microscopic detail flow, which uses a combination of pointwise convolution compression and multi-scale depth convolution decomposition to preserve high-frequency components such as edges, corners, and textures. Subsequently, a semantic gating mechanism is introduced, which uses the spatial weight mask generated by the macroscopic branch to selectively modulate the microscopic branch and achieves cross-domain recombination output through residual projection, thereby significantly improving the feature modeling complexity and expressive power without changing the module name.

[0127] In practice, the target projection module's processing steps include the following:

[0128] The structure of the target projection module is as follows: Figure 8 As shown;

[0129] The input features are processed through the spatial projection branch to obtain spatial projection features; the input features are also processed through the channel projection branch to obtain channel projection features.

[0130] The spatial projection branch processing includes: the input features are divided into two paths, one of which passes through... The deep convolutional layer processes the data to obtain the first spatial projection sub-features; the other path passes through... The second spatial projection sub-feature is obtained by deep convolutional layer processing; the first and second spatial projection sub-features are concatenated along the channel dimension, and then processed... The convolutional layer is used to process the spatial projection features. ;

[0131] The channel projection branch processing includes: the input features are processed by a normalization layer and then divided into three paths. The first path is... The convolutional layer processes the data to obtain the first channel projection sub-feature, and the second path passes through... The convolutional layer processes the data to obtain the second channel projection sub-features, and the third path passes through... The convolutional layers process the data to obtain the third channel projection feature. The first, second, and third paths use the same 1×1 convolutional layer structure, but their kernel parameters are independent, each used to perform different channel mappings on the normalized input features to generate three sets of complementary channel projection features. This structure simultaneously considers different types of inter-channel dependencies. The first, second, and third channel projection features are then element-wise summed to obtain the channel projection feature. ;

[0132] After cascading the spatial projection features and channel projection features along the channel dimension, they are then sequentially passed through... The convolutional layer and sigmoid activation function layer are used to process the data to obtain the gated fusion matrix. ;

[0133] Adopting a gated fusion matrix The spatial projection features and channel projection features are fused to obtain the aggregated projection features. , represented as ;

[0134] After summing the aggregated projection features and spatial projection features element-wise, they are then sequentially passed through the first... Convolutional layer and second The convolutional layer processes the data to obtain the bounding box detection results of the target.

[0135] After summing the aggregated projection features and channel projection features element-wise, they are then sequentially passed through the third... Convolutional layer, fourth The convolutional layer and the softmax activation function layer are used to process the data to obtain the classification result of the target.

[0136] The target projection module proposed in this invention refines feature processing into two orthogonal directions: "spatial projection branch" and "channel projection branch," and introduces a gated fusion matrix between the two to complete adaptive decoupling mapping. The spatial branch enhances local geometric response through multi-scale deep convolution, while the channel branch strengthens inter-class separability through point-by-point mapping. Ultimately, they serve the classification head and the bounding box regression head respectively, thereby reducing gradient interference caused by task conflicts.

[0137] S4. Using the training dataset constructed in step S2, train the initial target detection network based on feature aggregation constructed in step S3 to obtain the target detection network based on feature aggregation.

[0138] The training process includes the following steps:

[0139] The following formula is used as the loss function for the training process:

[0140] In the formula The value of the loss function; This is the first weight value set. For classification loss; This is the set second weight value; This represents the bounding box overlap loss. This is the set third weight value; For distributed regression loss; This is the fourth weight value that is set; For orthogonal constraint loss, and , This represents a vectorization mapping used to flatten an input feature tensor into a one-dimensional vector; Represents the inner product between two vectors. The L2 norm of a vector; The minimum value is set to prevent the denominator from being 0; through this joint loss function, the feature coupling redundancy between the classification branch and the regression branch can be reduced, and the stability and generalization ability of the decoupled projection terminal can be improved.

[0141] Mosaic enhancement was enabled in the early stage of training to enrich the background semantics, and the enhancement was turned off in the last 10 epochs to avoid sample distribution bias. At the same time, conventional enhancement methods such as random flipping and scale perturbation were introduced. The optimizer used was SGD, combined with automatic mixed precision (AMP) technology to improve training throughput and reduce memory usage.

[0142] S5. Use the feature aggregation-based target detection network obtained in step S4 to perform target detection on the image to be detected.

[0143] The effects of the method of the present invention will be further illustrated below with reference to an embodiment:

[0144] The method of the present invention and existing solutions were tested on the PASCAL VOC dataset and the MS COCO dataset to verify the performance of the method of the present invention.

[0145] The experimental results on the PASCAL VOC dataset are shown in Table 1:

[0146] The experimental results on the MS COCO dataset are shown in Table 2:

[0147] Among them, the YOLOv5-N and YOLOv5-S schemes are lightweight variants of the YOLOv5 series released by the Ultralytics team in 2020; the YOLOX-Tiny scheme is proposed by Zheng Ge et al. in the paper "YOLOX: Exceeding YOLO Series in 2021" in 2021; the YOLOv6-N scheme is proposed by Chuyi Li et al. in the paper "YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications" in 2022; the YOLOv7-Tiny scheme is proposed by Chien-Yao Wang et al. in the paper "YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors" in 2022; the YOLOv8-N scheme is a lightweight variant of the YOLOv8 series released by the Ultralytics team in 2023; and the YOLOv10-N scheme is proposed by Ao The YOLOv10 scheme was proposed by Wang et al. in the 2024 paper "YOLOv10: Real-Time End-to-End Object Detection"; the Faster R-CNN scheme was proposed by Shaoqing Ren et al. in the 2015 paper "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks"; and the Gold-YOLO-N scheme was proposed by Chengcheng Wang et al. in the 2023 paper "Gold-YOLO: Efficient Object Detector via Gather-and-DistributeMechanism".

[0148] As shown in Table 1, the proposed solution achieves an accuracy (P) of 81.9% and a mean average accuracy (mAP) of 78.5%. Compared with the YOLOv7-Tiny model of the same magnitude, the proposed solution improves mAP by 2.2 percentage points while reducing the number of parameters by about 30% (4.27M vs 6.22M), and is superior to the YOLOv10-N model.

[0149] As shown in Table 2, the proposed solution achieved an AP value of 40.8%, which is significantly better than two-stage networks such as Faster R-CNN and advanced single-stage networks such as YOLOv8-N in terms of detection accuracy. In particular, compared with the benchmark model YOLOv7-Tiny, the proposed solution improved the AP by 2.0%, while the computational load (FLOPs) decreased significantly from 13.8G to 8.2G. This fully demonstrates the efficiency of the proposed solution in feature extraction and its advantages in multi-scale fusion, and also shows that the proposed solution can maintain extremely high detection performance under the condition of limited computing resources.

[0150] The following ablation experiments illustrate the performance and function of the dual-domain rectification aggregation layer and signal rectification module (signal rectification layer) of the present invention.

[0151] The results of the ablation experiment are shown in Table 3:

[0152] As shown in Table 3, when only the dual-domain rectification aggregation layer is embedded, the detection accuracy AP increases from 36.8% to 40.5%, which is a performance gain of 3.7%. When the signal rectification module (signal rectification layer) is embedded at the same time, the AP is further improved to 40.8%. This shows that the feature map generated by the present invention can more accurately cover the complete geometric contour of the target object, rather than just focusing on the most discriminative local area.

[0153] Figure 9 This is a schematic diagram illustrating the implementation effect of the method of the present invention on the PASCAL VOC dataset. Figure 9 In (a), the foreground includes a dog and a bicycle, while the background area also includes vehicle targets. The method of this invention can accurately detect and locate different types of targets in transitional indoor / outdoor scenes and in complex backgrounds, with good matching between the detection bounding box and the target's main body area. Figure 9 In (b), the target body is relatively complete and its outline is clear; the method of the present invention can accurately cover the entire area of ​​the bird target and better reflect the edge range and morphological information of the target. Figure 9In (c), multiple similar targets appear simultaneously, and there are adjacent, overlapping and partially occluded situations between them; the method of the present invention can still distinguish and locate different horse targets relatively stably, indicating that it has good adaptability to densely distributed targets and partially occluded scenes.

[0154] Figure 10 This is a schematic diagram illustrating the implementation effect of the method of the present invention on the MS COCO dataset. Figure 10 In (a), different types of animals appear simultaneously and there is a certain degree of mutual occlusion; the method of this invention can effectively detect different types of targets under conditions with complex vegetation backgrounds and maintain good bounding box localization results. Figure 10 (b) includes a large bus and multiple pedestrian targets, representing a typical complex urban road scenario. The method of this invention can accurately detect buses and pedestrians even when multiple targets coexist and have significant scale differences, demonstrating good multi-target detection capabilities. Figure 10 (c) contains multiple targets with significant differences in scale, including kites in the distance that are small and scattered, and a large number of human targets. The method of the present invention can achieve relatively stable detection output for small targets and distant targets, while also taking into account the detection of nearby human targets.

[0155] pass Figure 9 and Figure 10 As can be seen, the method of the present invention has good target detection capability and can achieve relatively accurate and reliable target detection in various complex scenarios.

[0156] like Figure 11The diagram shows the functional modules of the system of the present invention: The system for implementing the target detection method based on feature aggregation disclosed in this invention includes a data acquisition module, a data processing module, a network construction module, a network training module, and a target detection module; the data acquisition module, data processing module, network construction module, network training module, and target detection module are connected in series; the data acquisition module is used to acquire image data information corresponding to the target image to be detected and upload the data information to the data processing module; the data processing module is used to preprocess the acquired image data information according to the received data information to construct a training dataset and upload the data information to the network construction module; the network construction module is used to construct an initial target detection network based on feature aggregation according to the received data information, based on convolution schemes, pooling schemes, dual-channel schemes, spatial projection schemes, and channel projection schemes, and upload the data information to the network training module; the constructed initial target detection network based on feature aggregation includes an input processing module, a dual-domain rectification aggregation module, a signal rectification module, and a target projection module connected in series; the input processing module is constructed based on a convolution scheme; input processing... The module performs downsampling on the input image and uploads the processing result to the dual-domain rectification and aggregation module. The dual-domain rectification and aggregation module is constructed based on convolution, pooling, and dual-channel schemes. This module extracts local features from the input data through micro-branching and global features through macro-branching, and achieves feature extraction by aggregating local and global features. A signal rectification module is constructed based on convolution and pooling schemes. This module performs boundary compensation and background suppression on the input data. A target projection module is constructed based on convolution, spatial projection, and channel projection schemes. This module performs target classification and bounding box detection through spatial and channel projection. A network training module trains the initial target detection network based on feature aggregation using the constructed training dataset, obtaining the target detection network based on feature aggregation, and uploads the data to the target detection module. The target detection module uses the obtained target detection network based on feature aggregation to perform target detection on the image to be detected, based on the received data.

Claims

1. A target detection method based on feature aggregation, characterized in that... Includes the following steps: S1. Obtain image data information corresponding to the target image to be detected; S2. Preprocess the image data information obtained in step S1 to construct a training dataset; S3. Construct an initial network for target detection based on feature aggregation using convolution, pooling, dual-channel, spatial projection, and channel projection schemes; The constructed target detection initial network based on feature aggregation includes an input processing module, a dual-domain rectification and aggregation module, a signal rectification module, and a target projection module connected in sequence. An input processing module is constructed based on a convolution scheme; The input processing module is used to downsample the input image and upload the processing result to the dual-domain rectification and aggregation module; A dual-domain rectification and aggregation module is constructed based on convolution, pooling, and dual-channel schemes. The dual-domain rectification and aggregation module extracts local features of the input data through micro-branch and global features of the input data through macro-branch. The feature extraction of the input data is achieved by aggregating local and global features. A signal rectification module is constructed based on convolution and pooling schemes; the signal rectification module is used to implement boundary compensation and background suppression of input data; A target projection module is constructed based on convolution, spatial projection, and channel projection schemes. The target projection module is used to classify targets and detect their bounding boxes through spatial and channel projection. S4. Using the training dataset constructed in step S2, train the initial target detection network based on feature aggregation constructed in step S3 to obtain the target detection network based on feature aggregation. S5. Use the feature aggregation-based target detection network obtained in step S4 to perform target detection on the image to be detected.

2. The target detection method based on feature aggregation according to claim 1, characterized in that... The input processing module's processing procedure includes the following steps: The input processing module includes a first input downsampling layer and a second output downsampling layer connected in series. The first input downsampling layer and the second output downsampling layer have the same structure; the input downsampling layer is a convolutional layer. The input downsampling layer is used to downsample the input image.

3. The target detection method based on feature aggregation according to claim 2, characterized in that... The processing steps of the dual-domain rectification and aggregation module include the following: The dual-domain rectification aggregation module includes micro-branch and macro-branch; The micro-branch includes a first micro-downsampling layer, a first micro-dual-domain rectification and aggregation layer, a second micro-downsampling layer, a second micro-dual-domain rectification and aggregation layer, a third micro-downsampling layer, a third micro-dual-domain rectification and aggregation layer, a fourth micro-downsampling layer, and a fifth micro-downsampling layer; The macroscopic branch includes the first macroscopic downsampling layer, the first macroscopic dual-domain rectification and aggregation layer, the second macroscopic downsampling layer, the second macroscopic dual-domain rectification and aggregation layer, the third macroscopic downsampling layer, the third macroscopic dual-domain rectification and aggregation layer, the fourth macroscopic downsampling layer, and the fifth macroscopic downsampling layer; The input of the dual-domain rectification and aggregation module is processed sequentially through a first microscopic downsampling layer and a first microscopic dual-domain rectification and aggregation layer to obtain a first microscopic feature. The first microscopic feature is then processed sequentially through a second microscopic downsampling layer and a second microscopic dual-domain rectification and aggregation layer to obtain a second microscopic feature. The second microscopic feature and the second macroscopic feature are cascaded and then processed sequentially through a third microscopic downsampling layer and a third microscopic dual-domain rectification and aggregation layer to obtain a third microscopic feature. The first microscopic feature is then processed through a fourth microscopic downsampling layer to obtain a first downsampled microscopic feature. The second microscopic feature is then processed through a fifth microscopic downsampling layer to obtain a second downsampled microscopic feature. The first downsampled microscopic feature, the second downsampled microscopic feature, and the third microscopic feature are summed element-wise to obtain the microscopic branch output feature. The input of the dual-domain rectification and aggregation module is processed sequentially through a first macroscopic downsampling layer and a first macroscopic dual-domain rectification and aggregation layer to obtain a first macroscopic feature. The first macroscopic feature is then processed sequentially through a second macroscopic downsampling layer and a second macroscopic dual-domain rectification and aggregation layer to obtain a second macroscopic feature. The second macroscopic feature and the second microscopic feature are cascaded and then processed sequentially through a third macroscopic downsampling layer and a third macroscopic dual-domain rectification and aggregation layer to obtain a third macroscopic feature. The first macroscopic feature is then processed through a fourth macroscopic downsampling layer to obtain a first downsampled macroscopic feature. The second macroscopic feature is then processed through a fifth macroscopic downsampling layer to obtain a second downsampled macroscopic feature. The first downsampled macroscopic feature, the second downsampled macroscopic feature, and the third macroscopic feature are summed element-wise to obtain the macroscopic branch output feature. The micro-branch output characteristics and the macro-branch output characteristics are used together as the output of the dual-domain rectification aggregation module; The structures of the first microscopic downsampling layer to the fifth microscopic downsampling layer and the first macroscopic downsampling layer to the fifth macroscopic downsampling layer are all the same; the downsampling layers are convolutional layers.

4. The target detection method based on feature aggregation according to claim 3, characterized in that... The structures of the first microscopic dual-domain rectification aggregation layer to the third microscopic dual-domain rectification aggregation layer and the first macroscopic dual-domain rectification aggregation layer to the third macroscopic dual-domain rectification aggregation layer are all identical; the processing procedure of the dual-domain rectification aggregation layer includes the following steps: Dual-domain rectifier aggregation layer includes Convolutional layers Hollow convolutional layer, signal rectification layer, and dual-domain feature aggregation layer; The input data is divided into two paths: one path passes through... The convolutional layer processes the data to obtain dual-domain rectified aggregated micro-features; the other path passes through... The dilated convolutional layer is used to process the macroscopic features of the dual-domain rectification aggregation. The microscopic features and macroscopic features of the dual-domain rectification aggregation are input together into the signal rectification layer for statistical description and cross-rectification. Then, the dual-domain feature aggregation layer performs detail extraction, semantic capture, geometric alignment and recombination to obtain the output of the dual-domain rectification aggregation layer.

5. The target detection method based on feature aggregation according to claim 4, characterized in that... The signal rectifier layer and the signal rectifier module have the same structure; the processing includes the following steps: The input data includes microscopic features and macro characteristics ; Microscopic features Passing in sequence The micro-rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ; Macro characteristics Passing in sequence The macroscopic rectified features are obtained by processing convolutional layers and the Sigmoid activation function. ; By cascading micro and macro features along the channel dimension, composite features are obtained. ; Composite features The micro-modulation coefficients are obtained through processing by the global descriptor generation layer. and macro modulation coefficient ; The following formula is used for parallel parameter processing to obtain micro-cross features. and macroscopic cross-features : In the formula This is element-wise multiplication; Micro-intersection features with microscopic features After performing element-wise multiplication, then combining with microscopic features By adding elements one by one, the microscopic composite rectification characteristics are obtained; Macro-intersection characteristics Macro characteristics After performing element-wise multiplication, then combine with macroscopic features. By adding elements one by one, the macroscopic composite rectification characteristics are obtained; After cascading the micro-composite rectification features and macro-composite rectification features along the channel dimension, and then through... The convolutional layer processes the data to obtain the output data.

6. The target detection method based on feature aggregation according to claim 5, characterized in that... The processing steps of the global descriptor generation layer include the following: The input data is divided into three paths: the first path is processed by a global average pooling layer to obtain the first sub-feature; the second path is processed by a global max pooling layer to obtain the second sub-feature; the third path is processed by a standard deviation statistics layer to obtain the third sub-feature; the first, second, and third sub-features are concatenated along the channel dimension to obtain the comprehensive sub-feature. The synthesized sub-features are processed sequentially through a first fully connected layer, a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer, and then channel splitting is performed to obtain the micro-modulation coefficients. and macro modulation coefficient .

7. The target detection method based on feature aggregation according to claim 6, characterized in that... The processing steps of the dual-domain feature aggregation layer include the following: The input features are divided into two paths; First path of input features After processing by the convolutional layer, the features are divided into two paths: the first path of aggregate features is processed by... The first aggregated sub-features are obtained by processing the deep convolutional layer. The second path aggregate feature is passed through The deep convolutional layer process yields the second aggregated sub-features; The first-path aggregation sub-feature and the second-path aggregation sub-feature are added together to obtain the first-path dual-domain aggregation feature; The second path of the input features is further divided into three sub-features. The first path of features uses features with a dilation rate of 1. Dilated convolution is used to process the data, resulting in the first dilated sub-feature; the second feature is obtained through a dilation rate of 2. Dilated convolution is used to obtain the second type of dilated sub-features; the third type of feature is obtained through a dilation rate of 3. The dilated convolution is processed to obtain the third dilated sub-feature; Softmax normalized weights are calculated for the first, second and third dilated sub-features respectively, and the three dilated sub-features are weighted and summed according to the obtained weights to obtain the weighted fused features; the weighted fused features are then input into the deformable convolutional layer for processing to obtain the dilated aggregate features. The void aggregation feature is divided into two paths: the first path void aggregation feature is processed through a global max pooling layer to obtain the first pooled void aggregation feature; the second path void aggregation feature is processed through a global average pooling layer to obtain the second pooled void aggregation feature; the first and second pooled void aggregation features are then summed element-wise, and then sequentially processed... Spatial attention features are obtained by processing convolutional layers and sigmoid activation function layers. After multiplying the spatial attention features and the first-path dual-domain aggregation features element-wise, the spatial aggregation features are obtained. After summing the spatial aggregation features, the hole aggregation features, and the input features element-wise, then... The convolutional layer is subjected to projection processing to obtain the output of the dual-domain feature aggregation layer.

8. The target detection method based on feature aggregation according to claim 7, characterized in that... The processing procedure of the target projection module includes the following steps: The input features are processed through the spatial projection branch to obtain spatial projection features; the input features are also processed through the channel projection branch to obtain channel projection features. The spatial projection branch processing includes: the input features are divided into two paths, one of which passes through... The deep convolutional layer processes the data to obtain the first spatial projection sub-features; the other path passes through... The second spatial projection sub-feature is obtained by deep convolutional layer processing; the first and second spatial projection sub-features are concatenated along the channel dimension, and then processed... The convolutional layer is used to process the spatial projection features. ; The channel projection branch processing includes: the input features are processed by a normalization layer and then divided into three paths. The first path is... The convolutional layer processes the data to obtain the first channel projection sub-feature, and the second path passes through... The convolutional layer processes the data to obtain the second channel projection sub-features, and the third path passes through... The convolutional layer processes the data to obtain the third channel projection feature; the first, second, and third channel projection features are then added element-wise to obtain the channel projection feature. ; After cascading the spatial projection features and channel projection features along the channel dimension, they are then sequentially passed through... The convolutional layer and sigmoid activation function layer are used to process the data to obtain the gated fusion matrix. ; Adopting a gated fusion matrix The spatial projection features and channel projection features are fused to obtain the aggregated projection features. , represented as ; After summing the aggregated projection features and spatial projection features element-wise, they are then sequentially passed through the first... Convolutional layer and second The convolutional layer processes the data to obtain the bounding box detection results of the target. After summing the aggregated projection features and channel projection features element-wise, they are then sequentially passed through the third... Convolutional layer, fourth The target classification result is obtained by processing the data through convolutional layers and softmax activation function layers.

9. The target detection method based on feature aggregation according to claim 8, characterized in that... The training described in step S4 specifically includes the following steps: The following formula is used as the loss function for the training process: In the formula The value of the loss function; This is the first weight value set. For classification loss; This is the set second weight value; This represents the bounding box overlap loss. This is the set third weight value; For distributed regression loss; This is the fourth weight value that is set; For orthogonal constraint loss, and , Represents a vectorized mapping. This represents the inner product of two vectors. The L2 norm of a vector; This is a set minimum number to prevent the denominator from being 0.

10. A system for implementing the target detection method based on feature aggregation as described in any one of claims 1 to 9, characterized in that... It includes a data acquisition module, a data processing module, a network construction module, a network training module, and a target detection module; the data acquisition module, data processing module, network construction module, network training module, and target detection module are connected in series; the data acquisition module is used to acquire image data information corresponding to the image of the target to be detected and upload the data information to the data processing module; The data processing module is used to preprocess the acquired image data based on the received data information to construct a training dataset and upload the data information to the network construction module; The network construction module is used to construct an initial target detection network based on feature aggregation based on received data, using convolution, pooling, dual-channel, spatial projection, and channel projection schemes, and then uploads the data to the network training module. The constructed initial target detection network based on feature aggregation includes an input processing module, a dual-domain rectification and aggregation module, a signal rectification module, and a target projection module connected in sequence. The input processing module is constructed based on the convolution scheme. It downsamples the input image and uploads the processing result to the dual-domain rectification and aggregation module. The dual-domain rectification and aggregation module is constructed based on the convolution, pooling, and dual-channel schemes. This module extracts local features from the input data through micro-branches and global features through macro-branches, achieving feature extraction by aggregating local and global features. The signal rectification module is constructed based on the convolution and pooling schemes. It performs boundary compensation and background suppression on the input data. The target projection module is constructed based on the convolution, spatial projection, and channel projection schemes. It classifies targets and detects their bounding boxes through spatial and channel projection. The network training module is used to train the initial target detection network based on feature aggregation using the constructed training dataset based on the received data information, thereby obtaining the target detection network based on feature aggregation, and uploading the data information to the target detection module. The target detection module is used to perform target detection on the image to be detected based on the received data information and the obtained feature aggregation-based target detection network.

Images