Edge monitoring target detection method, system and equipment based on lightweight network

Abstract

The invention discloses an edge monitoring target detection method, system and device based on an EfcientNetV2 and GhostNet lightweight network improvement. The method comprises the steps that S1, a monitoring image data set is acquired, and the data set is preprocessed; s2, constructing an improved monitoring target detection network model based on an EffcientNetV2 and a GhostNet lightweight network; s3, a loss function BBox is changed based on IOU; s4, setting network parameters; s5: according to the neural network models reconstructed in the steps 2 and 3, training by using the training set in the step 1, and verifying an optimal weight file obtained by training on a test set; according to the method, the backbone network part of single-stage target detection is reconstructed in combination with the EffcientNetV2 and the GhostNet, so that the parameter quantity reaches 4140404, the floating point operand is 5.1 GFLOPS, and compared with an existing mainstream yolk algorithm, the parameter quantity is only increased by 31.25%, the floating point operand is reduced by 41.38%, the average precision is reduced by 1.2%, and the real-time requirement for deployment on edge equipment is met.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to an edge monitoring target detection method, system and device based on EffcientNetV2 and GhostNet lightweight network improvements. Background Art

[0002] Object detection is an important branch of computer vision. With the rapid development of security monitoring technology, the amount of monitoring image data has exploded. How to quickly and accurately extract useful information from these massive data has become an urgent problem to be solved. The emergence of object detection technology provides strong support for solving this problem, and it plays a key role in many fields.

[0003] Convolutional neural networks have played a huge role in addressing the problems that arise in security monitoring. From the perspective of implementation, there are currently two types of convolutional neural networks. One is two-stage target detection. In the first stage, the image is divided based on the features of the candidate area, and features are extracted manually. In the second stage, the features are used to determine the type of target. Representative algorithms include R-CNN, Faster-CNN, Mask R-CNN, etc. These algorithms achieve high-precision target detection, but due to the characteristics of the two stages, the algorithm is more complex, and in complex spaces, there are problems with feature extraction, which leads to problems such as decreased detection accuracy and poor generalization ability. The other is single-stage target detection, which can directly locate and classify the target after only one detection, thereby improving the detection speed of the algorithm and maintaining the required detection accuracy. Representative algorithms include SSD, yolo series, etc. However, the existing single-stage algorithms still have problems of excessive computational complexity and poor real-time performance in the deployment of edge devices. In the edge computing scenario of monitoring target detection, real-time and low latency become issues that need to be addressed while ensuring accuracy is met, which requires the model to respond quickly. In addition, reducing energy consumption, extending device working time, and improving user experience are also important issues. Neither single-stage nor dual-stage algorithms can take into account both accuracy and lightweight. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes an edge monitoring target detection method, system and device based on the improved EffcientNetV2 and GhostNet lightweight networks; by introducing EffcientNetV2 and GhostNet networks, reconstructing feature extraction and neck networks, reducing the amount of calculation, using the dual-element attention mechanism to balance the problem of decreased accuracy caused by model lightweighting, using EIOU instead of IOU to improve the space utilization of the post-processing part of the model, while greatly reducing the amount of model calculation, balancing the accuracy, and meeting the conditions for deployment on edge devices.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] An edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement includes the following steps:

[0007] S1: Obtain surveillance image dataset, preprocess the dataset, and obtain training set and validation set;

[0008] S2: Build an edge monitoring target detection network model based on the improved EffcientNetV2 and GhostNet lightweight networks, including the input network Input, the backbone network Backbone, the neck network Neck and the exploration head Head;

[0009] S3: Change the loss function BBox: Based on IOU, the EIOU loss function is proposed.

[0010] S4: Set network parameters.

[0011] S5: According to the neural network model reconstructed in steps 2 and 3, it is trained using the training set in step 1. The best weight file obtained through training is verified on the test set and compared with the traditional algorithm. The number of parameters is greatly reduced, and the accuracy is only reduced by 1.2%, which proves the effectiveness of the algorithm.

[0012] Furthermore, the input network Input adopts the Mosaic data enhancement algorithm to randomly splice and scale the training set image data to increase data diversity, and automatically calculates anchors according to the data set, and automatically adjusts the network input size and the original image to reduce redundant information, so as to achieve the purpose of improving the inference speed.

[0013] The backbone network Backbone uses the feature extraction layer of EffcientNetV2 as the basic network, including a series of stem structures, FusedMBConv modules, MBConv modules and CBAM self-attention mechanism modules. The stem structure includes a convolution layer Conv2d, BatchNorm2d batch normalization processing and an activation function SiLU; the FusedMBConv module includes a stem structure, dropout regularization processing, dilated convolution, projection convolution and a residual structure; the MBConv module includes a depthwise separable convolution, point-by-point convolution, dilated convolution, projection convolution, SE module and a residual structure; the CBAM module includes a CAM channel attention module and a SAM spatial attention module.

[0014] The neck network Neck adopts GhostConv module and C3Ghost module for lightweight design, and uses Concat operation to fuse features of different layers to extract effective information.

[0015] The detection head uses the P21, 24, and 27 layers of the exploration neural network, which correspond one-to-one to three sizes: large (feature map 20*20), medium (feature map 40*40), and small (feature map 80*80), to perform target detection, realize multi-scale detection, and improve detection accuracy.

[0016] Furthermore, the specific contents of step S1 are as follows:

[0017] Step 101: Obtain a camera dataset; use a camera to shoot and create a dataset D, which includes an original image dataset and an image label dataset.

[0018] Step 102: Divide the data set into training sets D according to the corresponding proportions. v , and the validation set D t .

[0019] Furthermore, the specific contents of step S2 are as follows:

[0020] Step 201: Data set enhancement: Mosaic algorithm is used to perform data enhancement processing to obtain an enhanced training set.

[0021] Step 202: Build a feature extraction network, use a stem module to preprocess the input data, and then connect 5 FusedMBconv modules, 3 MBconv modules and 3 Sequential modules in series to replace the traditional Conv module and C3 module to improve the smooth flow of gradients, and perform BN operation after each convolution to solve the problems of gradient disappearance and gradient explosion.

[0022] Step 203: Insert the CBAM self-attention mechanism module before the SPPF module. CBAM first uses the channel attention mechanism to determine the weights of different channels, uses global average pooling and global maximum pooling, and combines multi-layer perceptron to fuse information. Then, the spatial attention mechanism is used to focus on the high-weight image features after spatial weighting of the feature map to improve the accuracy of recognition and solve the problem of decreased accuracy caused by lightweight design.

[0023] Step 204: In the neck network Neck, GhostConv is built to divide a standard convolution operation into two convolution operations: a smaller standard convolution (1*1) and a larger low-rank convolution (3*3), thereby significantly reducing the number of model parameters. C3Ghost is built to divide the standard convolution into three convolution operations: point convolution, group convolution, and depth convolution. This reduces the number of model parameters while improving computational efficiency, thereby achieving the purpose of optimizing the algorithm.

[0024] Furthermore, the specific contents of step S3 are as follows:

[0025] Loss function BBOX: IOU is used to measure the overlap between the predicted box and the true bounding box. Its calculation expression is as follows:

[0026]

[0027] Where A represents the predicted box boundary, B represents the real box boundary, (A∩B) is their intersection area, and (A∪B) is their union area. In practical applications, the IOU loss function is usually defined as 1-IOU.

[0028] Based on IOU, the EIOU loss function is proposed, which uses additional geometric constraints to enhance the expression of IOU, which is divided into three parts:

[0029] a.IOU overlap loss.

[0030] b. Center distance loss: The center distance loss uses the comparison between the center position of the predicted box and the true bounding box to enhance the expression. It pushes the predicted box to move toward the center of the true box, solving the gradient vanishing problem caused by the non-overlapping predicted box and the true box.

[0031] c. Width and height loss: The width and height difference between the predicted box and the true box is used to improve the accuracy of the loss function regression.

[0032] The calculation expression of EIOU is as follows:

[0033]

[0034] IOU is the interaction ratio between the predicted box and the true box, D is the Euclidean distance between the center of the predicted box and the center of the true box, and C is the square root of the distance between the center of the predicted box and the center of the true box. p and h p are the width and height of the prediction box respectively. g and h g are the width and height of the ground-truth box, respectively. λ1, λ2, λ3 are hyperparameters used to balance the various loss components.

[0035] Furthermore, the specific contents of step S4 are as follows:

[0036] Set the initial learning rate Ir0, learning rate decay factor Irf, IOU threshold, training rounds epochs, batch size batch-size, and use SGD (stochastic gradient descent) as the optimizer.

[0037] An edge monitoring target detection system based on EffcientNetV2 and GhostNet lightweight network improvement, including:

[0038] Data processing module: uses a camera to collect images, divides the data set and the verification set according to corresponding proportions, labels them, and obtains the processed data.

[0039] Image preprocessing module: The labeled data set is processed using the Mosaic algorithm to perform data enhancement to obtain an enhanced training set.

[0040] Network building module: used to build an edge monitoring target detection neural network based on the improved EffcientNetV2 and GhostNet lightweight networks, which includes feature extraction units, image recognition units and detection units.

[0041] Network training module: used to train on the training set according to the set network parameters, output the training weight file, verify the weight file on the verification set, select the training weight file with the highest accuracy, and obtain the required detection inference model.

[0042] Target detection module: Deploy the obtained detection inference model on the edge monitoring device, build a post-processing unit, and complete real-time detection of the edge monitoring device.

[0043] An edge monitoring target detection device based on EffcientNetV2 and GhostNet lightweight network improvement, comprising:

[0044] Machine executable instructions: a computer program storing the improved edge monitoring target detection method based on EffcientNet V2 and GhostNet lightweight network, which is a computer-readable executable instruction;

[0045] Processor: A processor used to execute the computer program to implement the improved edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network.

[0046] A computer-readable storage medium: The computer-readable storage medium stores a computer program, and when the computer program is processed by a processor, the storage medium can store the edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement. The storage medium can be any physical storage medium, which can include or store information, such as RAM, EMMC, ROM, SD, DVD, etc. The beneficial effects of the present invention are mainly manifested in:

[0047] The present invention makes lightweight improvements to the network based on EffcientNetV2 and GhostNet. By improving the feature extraction layer and replacing the Neck layer, the amount of calculation is greatly reduced while maintaining a high detection accuracy, meeting the requirements for deployment on edge devices, and being able to increase the service life of edge devices and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a network training flow chart of the present invention;

[0049] Figure 2 It is a diagram of the improved network structure of the present invention;

[0050] Figure 3 is a target detection graph of the present invention;

[0051] Figure 4 is a system block diagram of the present invention;

[0052] Figure 5 It is a hardware structure diagram of the electronic equipment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings.

[0054] Reference Figure 1 to Figure 5 , an edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement, comprising the following steps:

[0055] S1: Obtain image dataset, preprocess dataset, obtain training set and validation set;

[0056] The specific process in step S1 is as follows:

[0057] Step 101: Obtain a camera dataset; use a camera to shoot and create a dataset D, which includes an original image dataset and an image label dataset.

[0058] Step 102: Divide the data set. Divide the data set D obtained in step 101 into training set D according to the ratio of 8:2. v , and the validation set D t.

[0059] S2: Build an edge monitoring target detection network model based on the improved EffcientNetV2 and GhostNet lightweight networks, including the input network Input, the backbone network Backbone, the neck network Neck and the exploration head Head;

[0060] Input network Input, using the Mosaic data enhancement algorithm, randomly splicing and scaling four pictures to increase data diversity, and automatically calculating anchors based on the data set, using the network input size and the original picture to automatically adjust and reduce redundant information, in order to achieve the purpose of improving the inference speed.

[0061] The backbone network Backbone uses the feature extraction layer of EffcientNetV2 as the basic network, including a series of stem structures, FusedMBConv modules, MBConv modules and CBAM self-attention mechanism modules. The stem structure includes convolutional layer Conv2d, BatchNorm2d batch normalization processing and activation function SiLU; the FusedMBConv module includes stem structure, dropout regularization processing, dilated convolution, projection convolution and residual structure. The MBConv module includes depthwise separable convolution, point-by-point convolution, dilated convolution, projection convolution, SE module and residual structure. The CBAM module includes CAM channel attention module and SAM spatial attention module.

[0062] The neck network Neck adopts GhostConv module and C3Ghost module for lightweight design, and uses Concat operation to fuse features of different layers to extract effective information.

[0063] The detection head uses the P21, 24, and 27 layers of the exploration neural network, which correspond to the three sizes of large (feature map 20*20), medium (feature map 40*40), and small (feature map 80*80) to detect targets, realize multi-scale detection, and improve detection accuracy.

[0064] The specific construction process in step S2 is as follows:

[0065] Step 201: Data set enhancement: Mosaic algorithm is used to perform data enhancement processing to obtain an enhanced training set.

[0066] Step 202: Build a feature extraction network, use a stem module to preprocess the input data, and then connect 5 FusedMBconv modules and 6 MBconv modules in series to replace the traditional Conv module and C3 module to improve the smooth flow of gradients, and perform BN operation after each convolution to solve the problems of gradient disappearance and gradient explosion.

[0067] Step 203: Insert the CBAM self-attention mechanism module before the SPPF module. CBAM first uses the channel attention mechanism to determine the weights of different channels, uses global average pooling and global maximum pooling, and combines multi-layer perceptron to fuse information to generate a weight map. Then, the spatial attention mechanism is used to focus on the high-weight image features after the feature map space is weighted, so as to improve the accuracy of recognition and solve the problem of decreased accuracy caused by lightweight design.

[0068] Step 204: In the neck network Neck, GhostConv is built to divide a standard convolution operation into two convolution operations: a 1*1 standard convolution and a 3*3 low-rank convolution, thereby significantly reducing the number of model parameters. C3Ghost is built to divide the standard convolution into three convolution operations: point convolution, group convolution, and depth convolution. This reduces the number of model parameters while improving computational efficiency, thereby achieving the purpose of optimizing the algorithm.

[0069] S3: Change the loss function BBox:IOU to EIOU.

[0070] The specific process in step S3 is as follows:

[0071] Loss function BBOX: IOU is used to measure the overlap between the predicted box and the true bounding box. Its calculation expression is as follows:

[0072]

[0073] Where A represents the predicted box boundary, B represents the real box boundary, (A∩B) is their intersection area, and (A∪B) is their union area. In practical applications, the IOU loss function is usually defined as 1-IOU.

[0074] Based on IOU, the EIOU loss function is proposed, which uses additional geometric constraints to enhance the expression of IOU, which is divided into three parts:

[0075] a.IOU overlap loss.

[0076] b. Center distance loss: The center distance loss uses the comparison between the center position of the predicted box and the true bounding box to enhance the expression. It pushes the predicted box to move toward the center of the true box, solving the gradient vanishing problem caused by the non-overlapping predicted box and the true box.

[0077] c. Width and height loss: The width and height difference between the predicted box and the true box is used to improve the accuracy of the loss function regression.

[0078] The calculation expression of EIOU is as follows:

[0079]

[0080] IOU is the interaction ratio between the predicted box and the true box, D is the Euclidean distance between the center of the predicted box and the center of the true box, and C is the square root of the distance between the center of the predicted box and the center of the true box. p and h p are the width and height of the prediction box respectively. g and h g are the width and height of the ground-truth box respectively. λ1, λ2, λ3 are hyperparameters used to balance the various loss components.

[0081] S4: Adjust network parameters.

[0082] The specific contents of step S4 are as follows:

[0083] Set the initial learning rate Ir0 = 0.01, the learning rate decay factor Irf = 0.01, the IOU threshold = 0.2, the training round epochs = 300, the batch size batch-size = 16, and the optimizer uses SGD (stochastic gradient descent).

[0084] S5: Based on the neural network model reconstructed in steps 2 and 3, the training set in step 1 is used for training. The optimal weight file obtained through training is verified on the test set and compared with the traditional algorithm. The amount of calculation is greatly reduced, and the accuracy is only reduced by 1.2%, which proves the effectiveness of the algorithm.

[0085] An edge monitoring target detection system based on EffcientNetV2 and GhostNet lightweight network improvement, including:

[0086] Data processing module: Use a camera to take images, divide the data set and the verification set according to the ratio of 8:2, annotate them, and obtain the processed data.

[0087] Image preprocessing module: The labeled image data is spliced ​​using the Mosaic data enhancement algorithm to expand the training set and improve the accuracy of the model.

[0088] Network building module: used to build an edge monitoring target detection neural network based on the improved EffcientNetV2 and GhostNet lightweight networks, which includes feature extraction units, image recognition units and detection units.

[0089] Network training module: used to train on the training set according to the set network parameters, output the training weight file, verify the weight file on the verification set, select the training weight file with the highest accuracy, and obtain the required detection inference model.

[0090] Target detection module: Based on the above-mentioned detection inference model, it is deployed in the edge monitoring device, and an image post-processing unit is built to complete real-time detection.

[0091] Reference Figure 5 , an edge monitoring target detection device based on EffcientNetV2 and GhostNet lightweight network improvement, including:

[0092] Machine executable instructions: A computer program storing the improved edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight networks, which is a computer-readable executable instruction.

[0093] Processor: used to execute the computer program to implement the improved edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network.

[0094] A computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer program, and when the computer program is processed by a processor, it can implement the edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement. It can be any physical storage medium that can include or store information, such as: RAM, EMMC, ROM, SD, DVD, etc.

[0095] The comparison of the present invention with the existing mainstream algorithms in terms of parameter quantity, floating point operation quantity and precision mAP is shown in Table 1 below:

[0096] Table 1

[0097]

[0098] In summary, 1) the present invention combines EffcientNetV2 and GhostNet to reconstruct the backbone network part of single-stage target detection, so that the number of parameters reaches 4140404, and the floating-point operation amount is 5.1GFLOPS. Compared with the existing mainstream yolo algorithm, the number of parameters only increases by 31.25%, and the floating-point operation amount decreases by 41.38%, meeting the real-time requirements of deployment on edge devices.

[0099] 2) In order to address the problem of decreased accuracy caused by reconstructing the backbone network, the CBAM attention mechanism was introduced to solve the problem of decreased accuracy after the structure was lightweight, so that the network's mAP accuracy only dropped by 1.2% while the amount of computation was reduced.

[0100] 3) Based on the post-processing loss function, IOU is reconstructed into EIOU, so that in training, not only the loss between bounding boxes but also the loss of center point, width and height are utilized, so that its accuracy is further improved.

[0101] In this specification, the schematic description of the present invention does not necessarily refer to the same embodiment or example, and those skilled in the art may combine and combine different embodiments or examples described in this specification. In addition, the contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept, and the protection scope of the present invention should not be considered as limited to the specific forms described in the implementation cases. The protection scope of the present invention also includes equivalent technical means that those skilled in the art can think of based on the inventive concept.

Claims

1. An edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement, characterized in that: The following steps are involved: S1: Obtain surveillance image dataset, preprocess the dataset, and obtain training set and validation set; S2: Build a surveillance target detection network model based on the improved EffcientNetV2 and GhostNet lightweight networks, including the input network Input, the backbone network Backbone, the neck network Neck and the exploration head Head; The input network Input adopts the Mosaic data enhancement algorithm to randomly splice and scale the dataset images to increase data diversity, and automatically calculates anchors based on the dataset, and automatically adjusts the network input size and the original image to reduce redundant information; The backbone network Backbone uses the feature extraction layer of EffcientNetV2 as the basic network, including a series of stem structures, a FusedMBConv module, an MBConv module and a CBAM self-attention mechanism module; the stem structure includes a convolution layer Conv2d, a BatchNorm2d batch normalization process and an activation function SiLU; the FusedMBConv module includes a stem structure, a dropout regularization process, an extended convolution, a projection convolution and a residual structure, the MBConv module includes a depthwise separable convolution, a point-by-point convolution, an extended convolution, a projection convolution, a SE module and a residual structure, and the CBAM module includes a CAM channel attention module and a SAM spatial attention module; The neck network Neck uses the built GhostConv module and C3Ghost module, and uses the Concat operation to fuse the features of different layers to extract information; The detection head uses the P21, 24, and 27 layers of the exploration neural network to correspond to large, medium, and small targets one by one to perform detection and realize multi-scale detection; S3: Change the loss function BBox: Based on IOU, the EIOU loss function is proposed; S4: Set network parameters; S5: Based on the neural network model reconstructed by S2 and S3, the training set in S1 is used for training, and the best weight file obtained from the training is verified on the test set.

2. The edge monitoring target detection method based on the EffcientNetV2 and GhostNet lightweight network improvement as described in claim 1 is characterized in that: The specific steps of S1 are as follows: Step 101: Obtain a camera data set; use a camera to shoot and create a data set D, which includes an original image data set and an image label data set; Step 102: Divide the data set; divide the data set D obtained in step 101 into training sets D according to the proportion v , and the validation set D t .

3. The edge monitoring target detection method based on the improved EffcientNetV2 and GhostNet lightweight networks as described in claim 1 is characterized in that: The specific steps of S2 are as follows: Step 201: Data set enhancement: Mosaic algorithm is used to perform data enhancement processing to obtain an enhanced training set; Step 202: construct a feature extraction network; use a stem module to preprocess the input data, then connect 5 FusedMBconv modules, 3 MBconv modules and 3 Sequential modules in series to replace the traditional Conv module and C3 module, and perform a BN operation after each convolution; Step 203: insert the CBAM self-attention mechanism module before the SPPF module. CBAM first uses the channel attention mechanism to determine the weights of different channels, uses global average pooling and global maximum pooling, combines multi-layer perceptron fusion information, generates a weight map, and then uses the spatial attention mechanism to focus on the high-weight image features after the feature map space is weighted; Step 204: In the neck network Neck, build GhostConv to divide a standard convolution operation into two convolution operations, including: one standard convolution and one low-rank convolution; build C3Ghost to divide the standard convolution into three convolution operations: point convolution, group convolution and depth convolution, which reduces the number of model parameters and improves computational efficiency.

4. The edge monitoring target detection method based on the EffcientNetV2 and GhostNet lightweight network improvement as claimed in claim 1 is characterized in that: The specific steps of S3 are as follows: Loss function BBOX: IOU is used to measure the overlap between the predicted box and the true bounding box. Its calculation expression is as follows: Where A represents the predicted box boundary, B represents the real box boundary, (A∩B) is their intersection area, and (A∪B) is their union area; Based on IOU, the EIOU loss function is proposed, and additional geometric constraints are used to enhance the expression of IOU, which is divided into three parts: 1) IOU overlap loss; 2) Center distance loss: The center distance loss uses the comparison between the center position of the predicted box and the true bounding box to enhance the expression and push the predicted box to move toward the center of the true box. 3) Width and height loss: Use the width and height difference between the predicted box and the real box to improve the accuracy of loss function regression; The calculation expression of EIOU is as follows: IOU is the interaction ratio between the predicted box and the true box, D is the Euclidean distance between the center of the predicted box and the center of the true box, and C is the square root of the distance between the center of the predicted box and the center of the true box; w p and h p are the width and height of the prediction box respectively; w g and h g are the width and height of the true box respectively; λ1, λ2, λ3 are hyperparameters used to balance the various loss components.

5. The edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight network improvement as claimed in claim 1 is characterized in that: The specific steps of S4 are as follows: Set the initial learning rate Ir0, learning rate decay factor Irf, IOU threshold, training rounds epochs, batch size batch-size, and select the optimizer.

6. A low-power monitoring target detection system based on EffcientNetV2 and GhostNet lightweight network improvement, characterized in that: include: Data processing module: using the camera to collect images, and dividing the data set according to the ratio, annotating it, and obtaining the processed data; Image preprocessing module: Use the Mosaic data enhancement algorithm to splice the annotated image data, expand the training set, and improve the accuracy of the model; Network building module: used to build a low-power monitoring target detection network model based on the EffcientNetV2 and GhostNet lightweight network improvements; These include feature extraction unit, image recognition unit and detection unit; Network training module: used to train on the training set according to the set network parameters, output the training weight file, verify the weight file on the verification set, select the training weight file with the highest accuracy, and obtain the required detection inference model; Target detection module: Based on the above-mentioned detection inference model, it is deployed in the edge monitoring device, and an image post-processing unit is built to complete real-time detection.

7. A low-power target detection device based on EffcientNetV2 and GhostNet lightweight network improvement, characterized in that: include: Machine executable instructions: a computer program storing an improved edge monitoring target detection method based on EffcientNet V2 and GhostNet lightweight network as described in any one of claims 1 to 5, which is a computer-readable executable instruction; Processor: a processor for executing the computer program to implement the edge monitoring target detection method based on the EffcientNetV2 and GhostNet lightweight network improvement described in any one of claims 1-5; A computer-readable storage medium: the computer-readable storage medium stores a computer program, and when the computer program is processed by a processor, the computer-readable storage medium can store a readable storage medium of an improved edge monitoring target detection method based on EffcientNetV2 and GhostNet lightweight networks as described in any one of claims 1-5.

Images