Lightweight YOLO model target detection method based on group fast spatial pyramid pooling

Abstract

The present invention discloses a lightweight YOLO model based on group fast spatial pyramid pooling, comprising the following steps: step 1) using a cross-stage local network and group convolution to construct a group fast spatial pyramid pooling structure GFSPP; step 2) using a channel space intention module (CBAM) and a GFSPP structure to construct a backbone network module GFSPP-Backbone based on a group fast spatial pyramid pooling structure; step 3) using a phantom convolution structure to construct a lightweight neck network module Ghost-Slim-Neck; step 4) according to the order of the backbone network module GFSPP-Backbone, the neck network module Ghost-Slim-Neck, and the detection head network module, step 2), step 3) and the detection head network module are sequentially connected in series to form a lightweight target detection model GFSPP-YOLO. The method of the present invention can significantly improve the feature extraction capability of the backbone network module while significantly reducing the amount of calculation of the neck network module.

Description

Technical Field

[0001] The present invention relates to a lightweight YOLO model target detection method based on group fast spatial pyramid pooling, and belongs to the field of image target detection and recognition technology and computer vision technology. Background Art

[0002] As an important task in the field of computer vision, object detection is widely used in industrial scenarios, such as face detection, helmet detection, defect detection, etc. However, traditional object detection models have contradictions in terms of real-time performance, detection accuracy, and model size.

[0003] At present, in order to balance this contradiction, a common method is to use some lightweight network structures, such as MobileNet, ShuffleNet and GhostNet, to reduce the number of model parameters and calculations while maintaining detection accuracy. Among them, MobileNet uses DWConv instead of Conv to reduce the number of model parameters and calculations. ShuffleNet uses channel shuffling operations to achieve interactive fusion of information between channels. GhostNet takes into account the problem of feature map redundancy and obtains the same number of feature maps as ordinary convolution through simple linear operations, reducing the amount of calculation and improving the calculation speed. Some scholars integrate these lightweight networks into the target detection model, which can effectively reduce the computational complexity and number of parameters while maintaining a certain detection accuracy of the model, making real-time target detection on mobile devices possible.

[0004] For embedded environments, most of the above methods use DWConv to effectively reduce the number of parameters and the amount of calculation, but the detection accuracy of the model is too low and the detection accuracy is low, which cannot meet the detection accuracy requirements. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention provides a lightweight YOLO model target detection method based on group fast spatial pyramid pooling. Starting from the spatial pyramid pooling structure, while enhancing the feature representation ability of the backbone network, a lightweight design is made in the neck network, and a lightweight YOLO model based on group fast spatial pyramid pooling is proposed. The model balances the detection accuracy, calculation speed and model size to a certain extent, so that it can still maintain strong robustness and effective recognition ability for embedded device environments.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a lightweight YOLO model target detection method based on group fast spatial pyramid pooling, comprising the following steps:

[0007] Step 1: Use cross-stage local networks and group convolution to build a group fast spatial pyramid pooling structure GFSPP;

[0008] Step 2: Using the channel space intention module (CBAM) and GFSPP structure, construct the backbone network module GFSPP-Backbone based on the group fast spatial pyramid pooling structure;

[0009] Step 3: Use the ghost convolution structure to build a lightweight neck network module Ghost-Slim-Neck;

[0010] Step 4: In the order of the backbone network module GFSPP-Backbone, the neck network module Ghost-Slim-Neck, and the detection head network module, connect step 2), step 3), and the detection head network module in series to form the lightweight target detection model GFSPP-YOLO.

[0011] Furthermore, the specific contents of step 1 are:

[0012] Step 1-1: First, the original feature map is divided into two branches. In the first branch, it passes through a grouped convolution layer with convolution kernel sizes of 1×1, 3×3, and 1×1 respectively;

[0013] Step 1-2: Then, the output feature map is input into the SPPF structure, which consists of a group convolution layer with a convolution kernel size of 1×1 and three maximum pooling layers with a scale of 5×5; that is, the input feature map passes through three maximum pooling layers with a scale of 5×5, and each maximum pooling layer outputs a branch, and the original feature map and the output of the three branches are channel-joined, and then passed through a 1×1 group convolution layer for output;

[0014] Step 1-3: After passing through the SPPF structure, it is input into the grouped convolutional layers with convolution kernel sizes of 1×1 and 3×3 respectively, and the output feature map is channel-joined with the other branch of the original feature map (after passing through a grouped convolutional layer with a convolution kernel size of 1×1);

[0015] Step 1-4: Finally, the concatenated feature map is output through a grouped convolution layer with a convolution kernel size of 1×1 as the output of the GFSPP structure.

[0016] Furthermore, in step 2, the backbone network module GFSPP-Backbone outputs image features of three scales: large, medium and small. Its network structure is mainly composed of five CBS modules, four C3 modules, the GFSPP structure in step 1) and the CBAM module;

[0017] The specific contents are:

[0018] Step 2-1: First, connect five CBS modules and four C3 modules into a basic feature extraction module in the following order: CBS module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module;

[0019] Step 2-2: The CBS module is mainly composed of a convolutional layer, a BN layer, and a SiLU activation function; the C3 module is an important part of the backbone network. Its main function is to increase the depth and feature extraction capability of the network. The module is mainly composed of three CBS modules and several BottleNeck modules. The BottleNeck module is a series connection of two CBS modules.

[0020] First, the original input is divided into two branches, and convolution operations are performed on each branch to halve the number of channels. Then, one branch is subjected to CBS operation, followed by BottleNeck×N operation, and the other branch is directly subjected to CBS operation. Finally, the two branches are concatenated and then subjected to a CBS operation. This ensures that the input and output of the feature map are the same size after entering the C3 module, and the repeated stacking of C3 modules can build a deeper feature extraction network.

[0021] Step 2-3: Connect the GFSPP structure described in step 1) after the above basic feature extraction module;

[0022] Step 2-4: Based on step 2-3), the CBAM modules are connected in series; CBAM consists of two parts: the channel attention module and the spatial attention module. The channel attention generates the channel attention weight through the global average pooling and the fully connected layer, and performs weighted fusion on the features of different channels; the spatial attention generates the spatial attention map through the convolution layer to highlight the important positions and suppress the unimportant positions; finally, the feature maps processed by the channel attention and the spatial attention are added to obtain the final representation to capture the key information and improve the representation ability of the network features.

[0023] Furthermore, the specific content of step 3 is:

[0024] The lightweight neck network module Ghost-Slim-Neck takes the output of three scales of the backbone network module GFSPP-Backbone as input, performs multi-scale fusion of features from top to bottom and bottom to top, and then outputs optimized features of three different scales. Its network structure mainly consists of four GhostConv modules, four C3Ghost modules, four channel concatenation (Concat) modules and two upsampling (Upsample) modules;

[0025] Step 3-1: The neck network consists of 14 modules, in the following order: GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module; the first Concat module is the channel splicing of the middle-scale output of the backbone network and the first Upsample module, the second Concat module is the channel splicing of the small-scale output of the backbone network and the second Upsample module, the third Concat module is the channel splicing of the fifth GhostConv module and the ninth GhostConv module in the neck network, and the fourth Concat module is the channel splicing of the first GhostConv module and the twelfth GhostConv module in the neck network; the eighth module, the eleventh module, and the fourteenth module output small-scale, middle-scale, and large-scale image features, respectively;

[0026] Step 3-2: The GhostConv module consists of a CBS module and a DWConv module. The original feature map is first passed through a CBS module and then divided into two branches. One branch performs a DWConv operation, and the other branch does not perform any operation. Then the two branches are channel-joined as the output of the GhostConv module.

[0027] Step 3-3: The C3Ghost module is mainly composed of three CBS modules and one GhostBottleneck module. Its structure is the same as that of the C3 module, except that the GhostBottleneck module is replaced. The GhostBottleneck module is composed of two series-connected GhostConv modules and a parallel DWConv module to perform element-by-element addition and output.

[0028] Furthermore, the specific content of step 4 is:

[0029] Step 4-1: The feature maps output by the GFSPP-Backbone backbone network are used as the input of the large-scale, medium-scale, and small-scale features of the Ghost-Slim-Neck network respectively;

[0030] Step 4-2: The output feature maps of the large-scale, medium-scale, and small-scale features of the Ghost-Slim-Neck network are connected to the large-scale, medium-scale, and small-scale detection head networks respectively;

[0031] Step 4-3: Finally, the three detection head networks output large-scale, medium-scale, and small-scale target object positioning boxes Box, detection confidence Conf, and output category Class, respectively, to achieve lightweight target positioning and recognition functions.

[0032] The beneficial effects of the present invention are as follows: the present invention proposes a lightweight YOLO model based on group fast spatial pyramid pooling. Different from the existing YOLOv5 model, firstly, at the end of the backbone network, the idea of ​​cross-stage local network and group convolution is used to expand the receptive field, and a group fast spatial pyramid pooling structure GFSPP is constructed to avoid false detection and missed detection caused by image distortion; and the CBAM attention mechanism is introduced into the backbone network to enhance the representation ability of network features. Secondly, the Slim-Neck paradigm is used in the neck network to combine the GhostConv lightweight convolution module to compress the network structure. Experimental results show that the GFSPP-YOLO model proposed in the present invention reduces the amount of calculation and the amount of parameters by 10% and 3.5% respectively compared with the traditional YOLOv5 model on the PASCAL VOC2007+12 dataset, and at the same time, the mAP0.5 is increased by 2%, making the model of the present invention more suitable for application in mobile terminal embedded environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of a process for constructing a lightweight target detection network according to the present invention;

[0034] Figure 2 This is a schematic diagram of the GFSPP network structure proposed by the present invention;

[0035] Figure 3 This is a schematic diagram of the GFSPP-Backbone network structure proposed by the present invention;

[0036] Figure 4 This is a schematic diagram of the Ghost-Slim-Neck neck network structure proposed by the present invention;

[0037] Figure 5 The lightweight YOLO model based on grouped fast spatial pyramid pooling proposed by the present invention;

[0038] Figure 6 Schematic diagram of the comparison between the lightweight target detection network of the present invention and other lightweight detection models in terms of target detection accuracy and model parameter volume. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below through the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0041] like Figure 1 As shown, a lightweight YOLO model target detection method based on group fast spatial pyramid pooling includes the following steps:

[0042] Step 101: Use the cross-stage local network and group convolution to build a group fast spatial pyramid pooling structure GFSPP; the specific content is:

[0043] like Figure 2 As shown in the figure, the group fast spatial pyramid pooling structure GFSPP (Group Fast Spatial PyramidPooling) combines the ideas of cross-stage local network and group convolution on the basis of SPPF, aiming to improve the receptive field and feature expression ability of the model. This structure introduces several group convolution layers at the front end and back end of the SPPF structure, where the convolution kernel size is 1×1 and 3×3. The specific structure is as follows:

[0044] Step 1-1 First, the original feature map is divided into two branches. In the first branch, it passes through the grouped convolution layer with convolution kernel sizes of 1×1, 3×3, and 1×1 respectively;

[0045] Step 1-2 Then, the output feature map is input into the SPPF structure. The SPPF structure consists of a group convolution layer with a convolution kernel size of 1×1 and three maximum pooling layers with a scale of 5×5. That is, the input feature map passes through three maximum pooling layers with a scale of 5×5, and each maximum pooling layer outputs a branch, and the original feature map and the output of the three branches are channel-joined, and then output through a 1×1 group convolution layer.

[0046] After passing through the SPPF structure, steps 1-3 are input into the grouped convolution layers with convolution kernel sizes of 1×1 and 3×3 respectively, and the output feature map is channel-joined with another branch of the original feature map (after passing through a grouped convolution layer with a convolution kernel size of 1×1).

[0047] Step 1-4 Finally, the concatenated feature map is output through a grouped convolution layer with a convolution kernel size of 1×1 as the output of the GFSPP structure.

[0048] Step 102: Using the channel space intention module (CBAM) and the GFSPP structure, a backbone network module GFSPP-Backbone based on the group fast spatial pyramid pooling structure is constructed; the specific content is:

[0049] like Figure 3 As shown,

[0050] Step 2-1 First, connect five CBS modules and four C3 modules into a basic feature extraction module in the following order: CBS module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module.

[0051] Step 2-2 Among them, the CBS module is mainly composed of convolutional layers, BN layers and SiLU activation functions; the C3 module is an important part of the backbone network. Its main function is to increase the depth and feature extraction capabilities of the network. This module is mainly composed of three CBS modules and several BottleNeck modules. The BottleNeck module is a series connection of two CBS modules. First, the original input is divided into two branches, and convolution operations are performed on each branch to halve the number of channels. Then one branch performs a CBS operation, and then undergoes a BottleNeck×N operation. The other branch directly performs a CBS operation. Finally, the two branches are spliced ​​and then a CBS operation is performed. This ensures that the input and output of the feature map are the same size after entering the C3 module, and the repeated stacking of the C3 module can build a deeper feature extraction network.

[0052] Step 2-3: Connect the GFSPP structure described in step 1) after the above basic feature extraction module;

[0053] Step 2-4 Based on step 2-3), the CBAM module is connected in series. CBAM consists of two parts: channel attention module and spatial attention module. Channel attention generates channel attention weights through global average pooling and fully connected layers, and performs weighted fusion of different channel features. Spatial attention generates spatial attention maps through convolutional layers to highlight important positions and suppress unimportant positions. Finally, the feature maps processed by channel attention and spatial attention are added to obtain the final representation to capture key information and improve the representation ability of network features.

[0054] Step 103: Use the ghost convolution structure to build a lightweight neck network module Ghost-Slim-Neck; the specific content is:

[0055] like Figure 4 As shown in the figure, the light neck network module Ghost-Slim-Neck takes the output of the three scales of the backbone network module GFSPP-Backbone as input, performs multi-scale fusion of features from top to bottom and bottom to top, and then outputs optimized features of three different scales. Its network structure mainly consists of four GhostConv modules, four C3Ghost modules, four channel concatenation (Concat) modules and two upsampling (Upsample) modules. The specific structure is as follows:

[0056] Step 3-1 The neck network consists of 14 modules, in the following order: GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module. The first Concat module is the channel splicing of the middle-scale output of the backbone network and the first Upsample module, the second Concat module is the channel splicing of the small-scale output of the backbone network and the second Upsample module, the third Concat module is the channel splicing of the fifth GhostConv module and the ninth GhostConv module in the neck network, and the fourth Concat module is the channel splicing of the first GhostConv module and the twelfth GhostConv module in the neck network. The eighth module, the eleventh module, and the fourteenth module output small-scale, middle-scale, and large-scale image features, respectively.

[0057] Step 3-2: The GhostConv module consists of a CBS module and a DWConv module. The original feature map is first passed through a CBS module and then divided into two branches. One branch performs a DWConv operation, and the other branch does not perform any operation. The two branches are then channel-joined as the output of the GhostConv module.

[0058] Step 3-3 The C3Ghost module is mainly composed of three CBS modules and one GhostBottleneck module. Its structure is the same as that of the C3 module, except that the GhostBottleneck module is replaced. The GhostBottleneck module is composed of two series-connected GhostConv modules and a parallel DWConv module to perform element-wise addition and output.

[0059] Step 104: In the order of the backbone network module GFSPP-Backbone, the neck network module Ghost-Slim-Neck, and the detection head network module, step 2), step 3), and the detection head network module are connected in series to form a lightweight target detection model GFSPP-YOLO. The specific contents are:

[0060] like Figure 5 As shown in the figure, the output feature map of the GFSPP-Backbone backbone network is used as the input of the large-scale, medium-scale and small-scale features of the Ghost-Slim-Neck neck network; the output feature maps of the large-scale, medium-scale and small-scale features of the Ghost-Slim-Neck neck network are connected to the large-scale, medium-scale and small-scale detection head networks respectively; finally, the three detection head networks output three target detection result parameters, namely the large-scale, medium-scale and small-scale target object positioning box Box, the detection confidence Conf and the output category Class, respectively, to realize lightweight target positioning and recognition functions.

[0061] Figure 6The lightweight model of the present invention is compared with other lightweight models, including YOLOv5s, YOLOv5-GSConv-Slim-Neck, YOLOv5-ShuffleNetv2, YOLOv5-MobileNetv3, and YOLOv5-GhostNet on the VOC2007+12 public data set. The experimental environment and settings of the present invention are as follows: the experiments of the present invention are all carried out in Intel(R) Xeon® Silver 4214 CPU 2.20GH, the system of the server is Ubuntu18.04.5 LTS, equipped with NVIDIA GeForce RTX 3090 graphics card, and run under Pytorch1.8+CUDA11.1 deep learning framework. The hyperparameters of training are as follows: the optimizer is stochastic gradient descent SGD, a linear decay learning rate adjustment strategy is adopted, the initial learning rate is 0.01, the momentum and weight decay are 0.937 and 0.0005 respectively, the batch size is 64, and the epochs are 300 rounds. The experimental data set is PASCAL VOC07+12, of which the training set is 16551 pictures, the test set is 4952 pictures, and there are 20 categories in total. For the evaluation indicators of the model, the present invention uses parameter size (Params), floating point operations (FLOPs), and frames per second (FPS) to respectively indicate how many parameters the model contains, the complexity, and the reasoning speed. mAP0.5 and mAP0.5:0.95 are used to measure the accuracy of the model, where mAP0.5 represents the average accuracy of all categories when the IoU threshold is 0.5. mAP0.5:0.95 represents the average accuracy of all categories with an IoU threshold from 0.5 to 0.95 with a step size of 0.05 and weighted average. Table 1 shows the comparative experiment of the lightweight model proposed in the present invention with other lightweight models.

[0062] As can be seen from Table 1, compared with the baseline (original YOLOv5s), our model reduces the amount of computation by 10%, the amount of parameters by 3.5%, and the mAP0.5 and mAP0.5:0.95 are increased by 2% and 4.7% respectively. Compared with other improved lightweight models, although the amount of parameters and computation is slightly higher, the detection accuracy is much lower than the model in this paper, which is not suitable for use in mobile devices. The model of the present invention has achieved a balance in terms of the amount of parameters, computation and detection accuracy of the model, so that the model can be deployed on edge devices and still maintain strong robustness and effective recognition capabilities while performing real-time processing (FPS>35).

[0063] Table 1 Comparative test results of lightweight target detection models

[0064] Model GFLOPS Params(M) mAP0.5 mAP0.5:0.95 FPS YOLOv5s 16.0 7.06 0.789 0.527 128 YOLOv5s+MobileNetv3 6.4 3.59 0.703 0.429 103 YOLOv5s+ShuffleNetv2 8.1 3.84 0.733 0.460 105 YOLOv5s+GhostNet 8.2 3.73 0.788 0.538 113 YOLOv5s+GSConv+Slim-Neck 12.8 5.89 0.791 0.551 120 Ours 14.4 6.81 0.809 0.571 106

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A lightweight YOLO model target detection method based on group fast spatial pyramid pooling, characterized in that: The steps include: Step 1: Use cross-stage local networks and group convolution to build a group fast spatial pyramid pooling structure GFSPP; Wherein, the specific content of step 1 is: Step 1-1: First, the original feature map is divided into two branches. In the first branch, it passes through a grouped convolution layer with convolution kernel sizes of 1×1, 3×3, and 1×1 respectively; Step 1-2: Then, the output feature map is input into the SPPF structure, which consists of a group convolution layer with a convolution kernel size of 1×1 and three maximum pooling layers with a scale of 5×5; that is, the input feature map passes through three maximum pooling layers with a scale of 5×5, and each maximum pooling layer outputs a branch, and the original feature map and the output of the three branches are channel-joined, and then passed through a 1×1 group convolution layer for output; Step 1-3: After passing through the SPPF structure, it is input into the grouped convolutional layers with convolution kernel sizes of 1×1 and 3×3 respectively, and the output feature map is channel-joined with the other branch of the original feature map, that is, a grouped convolutional layer with a convolution kernel size of 1×1; Step 1-4: Finally, the concatenated feature map is output through a grouped convolution layer with a convolution kernel size of 1×1 as the output of the GFSPP structure; Step 2: Using the channel space attention module CBAM and GFSPP structure, construct the backbone network module GFSPP-Backbone based on the group fast spatial pyramid pooling structure; Wherein, in step 2, the backbone network module GFSPP-Backbone outputs image features of three scales: large, medium and small, and its network structure consists of five CBS modules, four C3 modules, the GFSPP structure in step 1 and the CBAM module; The specific contents are: First, five CBS modules and four C3 modules are connected into a basic feature extraction module in the following order: CBS module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module, CBS module, C3 module; Among them, the CBS module is composed of a convolutional layer, a BN layer and a SiLU activation function; the C3 module is an important part of the backbone network. Its function is to increase the depth and feature extraction capability of the network. This module is composed of three CBS modules and several BottleNeck modules. The BottleNeck module is a series connection of two CBS modules. The specific operation of the C3 module is to first divide the original input into two branches, perform convolution operations on each branch to halve the number of channels, then perform CBS operation on one branch, and then perform BottleNeck×N operation, and directly perform CBS operation on the other branch. Finally, the two branches are spliced ​​and then a CBS operation is performed, so that the feature map ensures that the input and output are the same size after entering the C3 module, and the repeated stacking of C3 modules can build a deeper feature extraction network; Connect the GFSPP structure described in step 1 after the above basic feature extraction module; On the basis of the above, the CBAM module is connected in series; CBAM consists of two parts: channel attention module and spatial attention module. Channel attention generates channel attention weights through global average pooling and fully connected layers, and performs weighted fusion of different channel features; spatial attention generates spatial attention maps through convolutional layers to highlight important positions and suppress unimportant positions; finally, the feature maps processed by channel attention and spatial attention are added to obtain the final representation to capture key information and improve the representation ability of network features; Step 3: Use the ghost convolution structure to build a lightweight neck network module Ghost-Slim-Neck; Step 4: In the order of the backbone network module GFSPP-Backbone, the neck network module Ghost-Slim-Neck, and the detection head network module, step 2, step 3, and the detection head network module are connected in series to form a lightweight target detection model GFSPP-YOLO to achieve lightweight target positioning and recognition functions. The experimental data set is PASCAL VOC07+12, with a training set of 16,551 images and a test set of 4,952 images, with a total of 20 categories.

2. A lightweight YOLO model target detection method based on grouped fast spatial pyramid pooling according to claim 1, characterized in that: The specific content of step 3 is: The lightweight neck network module Ghost-Slim-Neck takes the output of three scales of the backbone network module GFSPP-Backbone as input, performs multi-scale fusion of features from top to bottom and bottom to top, and then outputs optimized features of three different scales. Its network structure consists of four GhostConv modules, four C3Ghost modules, four channel concatenation Concat modules and two upsample Upsample modules; The neck network consists of 14 modules in total, in the following order: GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Upsample module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module, GhostConv module, Concat module, C3Ghost module; the first Concat module is the channel splicing of the middle-scale output of the backbone network and the first Upsample module, the second Concat module is the channel splicing of the large-scale output of the backbone network and the second Upsample module, the third Concat module is the channel splicing of the second GhostConv module and the third GhostConv module in the neck network, and the fourth Concat module is the channel splicing of the first GhostConv module and the fourth GhostConv module in the neck network; the eighth module, the eleventh module, and the fourteenth module output large-scale, middle-scale, and small-scale image features, respectively; The GhostConv module consists of a CBS module and a DWConv module. The original feature map is first passed through a CBS module and then divided into two branches. One branch performs a DWConv operation, and the other branch does not perform any operation. Then the two branches are channel-joined as the output of the GhostConv module. The C3Ghost module consists of three CBS modules and one GhostBottleneck module. Its structure is the same as that of the C3 module, except that the GhostBottleneck module is replaced. The GhostBottleneck module is composed of two series-connected GhostConv modules and a parallel DWConv module, which perform element-wise addition operations and output.

3. A lightweight YOLO model target detection method based on grouped fast spatial pyramid pooling according to claim 1, characterized in that: The specific content of step 4 is: Step 4-1: The feature maps output by the GFSPP-Backbone backbone network are used as the input of the large-scale, medium-scale, and small-scale features of the Ghost-Slim-Neck network respectively; Step 4-2: The output feature maps of the large-scale, medium-scale, and small-scale features of the Ghost-Slim-Neck network are connected to the large-scale, medium-scale, and small-scale detection head networks respectively; Step 4-3: Finally, the three detection head networks output large-scale, medium-scale, and small-scale target object positioning boxes Box, detection confidence Conf, and output category Class, respectively, to achieve lightweight target positioning and recognition functions.

Images