A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7

Abstract

The invention discloses a method for detecting military tanks in remote sensing images based on MSG-YOLOv7, which belongs to the technical field of combining deep learning target detection with remote sensing military. The method comprises the following specific steps: obtaining a remote sensing military tank data set disclosed by Google Earth, dividing the data into a training set, a validation set and a test set in a ratio of 6:2:2; in a YOLOv7 network, using a lightweight MobileNetv3 to replace the original Backbone to reduce the size of the model; designing an SD-MP module that is more suitable for remote sensing targets, and reducing the feature loss of tank targets during downsampling; the invention replaces the traditional Backbone with MobileNetv3, and designs an SD-MP module and a GD-ELAN module, thereby achieving lightweight model, reducing the amount of calculation, and improving the representation ability of the model, solving the problem of excessive volume and calculation in military tank detection in remote sensing images, and providing a feasible solution for lightweight and efficient remote sensing tank target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of combining deep learning target detection with remote sensing military, and specifically to a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7. Background Art

[0002] In recent years, with the rapid development of satellite technology, our ability to obtain remote sensing images has made a qualitative leap. These images have not only increased significantly in quantity, but also greatly improved in quality. Remote sensing image target detection technology has shown its great potential in many fields, especially in the military field. Rapidly and accurately identifying military tank targets has extremely important strategic value for battlefield situation awareness and tactical decision-making.

[0003] In the past, we mainly relied on traditional detection algorithms to identify remote sensing military tank targets. For example, some methods use multi-scale fractal features or a combination of ant colony optimization and artificial neural networks. However, these methods usually require manual feature selection and continuous optimization of feature extraction and classification algorithms. Since these methods are not competent for complex scenes, their accuracy and robustness are relatively poor.

[0004] With the rapid advancement of computer technology, deep learning target detection technology has brought new developments to remote sensing image military tank target detection. At present, the main technical routes are divided into two categories: one is the two-stage detection method, such as RCNN and Faster-RCNN, which usually has high detection accuracy and can handle multi-scale and complex scenes, but its computational complexity is high and the speed is slow, which is not ideal in remote sensing military tank detection; the other is the single-stage detection method, such as SSD and YOLO, which realizes target detection through a single forward propagation, which is faster and performs well in real-time applications.

[0005] Although single-stage detection methods have achieved some results in the research of remote sensing military target detection, they also face problems such as different target directions and complex backgrounds. In order to solve these problems, researchers have proposed some improvement methods. For example, some scholars have proposed the YOLO-FSD algorithm, which reduces the prediction bias caused by small and dense targets by introducing the Swin-CSP structure and DWC head in each layer of the YOLOv7 network, and improves the performance of the algorithm in complex backgrounds. On the other hand, some scholars have proposed the DCN-YOLO one-stage target detection algorithm, which uses DCNv2 and residual structure to optimize the backbone network of YOLOv7 and enhance the feature extraction ability of the model. However, these algorithms still have the problem of excessive computation. In the field of military applications, the algorithm needs to run on resource-constrained devices, such as drones or single-soldier combat helmets. Therefore, a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 is invented to achieve high accuracy while reducing computational consumption, so as to ensure real-time and high efficiency and meet the needs of rapid decision-making. Summary of the invention

[0006] In view of the above and / or existing problems in a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7, the present invention is proposed.

[0007] Therefore, the purpose of the present invention is to provide a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7, which can solve the above-mentioned existing problems.

[0008] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0009] A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7, which includes the following specific steps:

[0010] S1: Obtain the remote sensing military tank dataset publicly available on Google Earth and divide it into training set, validation set, and test set in a ratio of 6:2:2;

[0011] S2: In the YOLOv7 network, lightweight MobileNetv3 is used to replace the original Backbone to reduce the size of the model;

[0012] S3: Design an SD-MP module that is more suitable for remote sensing targets and reduce the feature loss of tank targets during downsampling;

[0013] S4: GD-ELAN module is designed based on GCNet and depthwise separable convolution, which enhances the representation ability of the model while making it lightweight;

[0014] S5: Input the data set into the network for iterative training. After the training is completed, use the optimal model to detect the test set to obtain the final result.

[0015] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the remote sensing image military tank data set in S1 is specifically: obtaining Google Earth data from open source, which contains 4000 pictures of different heights, orientations and times, and dividing them into 2400 training sets, 800 verification sets and 800 test sets for experiments.

[0016] As a preferred solution of a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, wherein: the lightweight MobileNetv3 is used in S2, which is specifically: using MobileNetv3 as the backbone network of MSG-YOLOv7, significantly reducing the number of parameters and the amount of calculation of the model, wherein MobileNetv3 is a lightweight neural network structure designed specifically for image processing tasks, which is mainly composed of a Bneck structure, which is mainly composed of an expanded convolution, a depth-separable convolution, an SE module and a residual connection. In this structure, the expanded convolution helps to increase the receptive field, the depth-separable convolution reduces the computational cost and promotes the flow of information, and the SE module can better model the relationship between features, and the residual connection can fuse information between different feature layers. The integration of these modules can effectively improve the efficiency of feature learning and information transmission, help to keep the model lightweight while providing good performance, and make the model more suitable for resource-constrained scenarios.

[0017] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the depth-separable convolution is a commonly used operation in convolutional neural networks, which is designed to reduce the number of parameters and the amount of calculation while maintaining model performance, and is mainly divided into two steps:

[0018] Step 1: Depth convolution, performing a separate convolution operation on each input channel to capture local features;

[0019] Step 2: Point-by-point convolution, using a 1×1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel;

[0020] Among them, the formulas for the standard convolution parameter quantity P and the calculation quantity F are as follows:

[0021] P=D K ×D K ×M×N

[0022] F=D K ×D K ×M×N×D W ×D H

[0023] Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D H Represents high output characteristics;

[0024] The parameter amount DP and computation amount DF formula of depth-separable convolution are as follows:

[0025] DP=D K ×D K ×M+M×N

[0026] DF=D K ×D K ×M×D W ×D H +M×N×D W ×D H

[0027] The comparison of the parameters and calculation amount between the two is as follows:

[0028]

[0029] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the SD-MP module in S3 is mainly used to solve the feature loss problem that tank targets that may appear in remote sensing images may face during the downsampling process. The SD-MP module uses the separation and splicing technology to first divide the input image into small blocks, and recombine them by channel, and then perform convolution processing to capture richer feature information. This operation mode effectively alleviates the information loss that may be caused by downsampling. In addition, the deep separable convolution is introduced to replace the original standard convolution to improve the efficiency and performance of the module.

[0030] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the GD-ELAN module in S4 is an innovatively designed module to overcome the challenges faced in target detection under complex backgrounds. Its core is built on two main components: global context GCNet and deep separable convolution. Its main function is to enhance the network's perception of the overall background and its ability to express features, and solve the problem of difficulty in detecting military targets in complex backgrounds.

[0031] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the global context GCNet is a global context module that can obtain global information in image processing. This global information is very important for correctly distinguishing the target from the background and understanding the relationship between the target and the surrounding environment. It is very effective for improving the accuracy and reliability of tank target detection. The calculation process is as follows:

[0032]

[0033] Y=X×σ(W v2 Relu(LN(W v1 Y1)))

[0034] In the formula, X i Represents any position information in the feature map, N p represents the number of all positions, W k represents the weight parameter, Softmax represents the normalization operation, Y1 is the obtained feature matrix, Relu is the activation function, LN is the layer normalization, W v1 , W v2 are the weight parameters of the two convolutional layers v1 and v2, σ is the sigmoid function, and Y is the final feature map.

[0035] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, the model training in S5 is specifically as follows: the experiment is carried out in the Ubuntu 16.04 environment, using Python 3.8 and PyTorch 1.9.0 as the main development tools, and in terms of hardware configuration, GeForce RTX3090 GPU and Core i9-10900 CPU are used, equipped with CUDA 11.1, and the system memory capacity is 64GB. The image size is set to 640×640 in the model training, and the SGD optimizer is used, the batch size is 16, and the number of training iterations is 200.

[0036] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, wherein: the model detection result is obtained in S5, which is specifically: the improved algorithm is compared with different algorithms, and the weight size, number of frames per second, parameter amount, calculation amount and average precision AP index are selected to evaluate the model performance, and the AP calculation formula is as follows:

[0037]

[0038]

[0039]

[0040] In the formula, P is precision, R is recall, TP is the number of correctly detected samples, FP is the number of samples that are falsely detected as positive samples, and FN is the number of samples that are falsely detected as negative samples.

[0041] As a preferred solution of the lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 described in the present invention, wherein: the improved algorithm is compared with different algorithms in S5, specifically: in order to accurately verify the performance of the MSG-YOLOv7 algorithm in the remote sensing tank detection task, it is compared with SSD, YOLOv3, YOLOv4, YOLOv5, YOLOv7, YOLO-FSD and DCN-YOLO algorithms, and the above algorithms are trained and tested under the same conditions.

[0042] Compared with existing technologies:

[0043] The present invention replaces the traditional Backbone with MobileNetv3, and designs the SD-MP module and the GD-ELAN module, thereby achieving model lightweight, reducing the amount of computation, and improving the model's representation ability, solving the problems of excessive volume and computation in military tank detection in remote sensing images, and providing a feasible solution for lightweight and efficient remote sensing tank target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the method of the present invention;

[0045] Figure 2 It is the overall network structure diagram of the present invention;

[0046] Figure 3 This is a partial data set picture of the present invention;

[0047] Figure 4 Bneck and MobileNetv3 of the present invention;

[0048] Figure 5 The depth-separable convolutional DSC of the present invention;

[0049] Figure 6 The SD-MP module of the present invention;

[0050] Figure 7 The GD-ELAN module of the present invention;

[0051] Figure 8 It is the GCNet module of the present invention. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0053] This paper provides a lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7, please refer to Figure 1-Figure 8 , including the following specific steps:

[0054] S1: Obtain the remote sensing military tank dataset publicly available on Google Earth and divide it into training set, validation set, and test set in a ratio of 6:2:2;

[0055] S2: In the YOLOv7 network, lightweight MobileNetv3 is used to replace the original Backbone to reduce the size of the model;

[0056] S3: Design an SD-MP module that is more suitable for remote sensing targets and reduce the feature loss of tank targets during downsampling;

[0057] S4: GD-ELAN module is designed based on GCNet and depthwise separable convolution, which enhances the representation ability of the model while making it lightweight;

[0058] S5: Input the data set into the network for iterative training. After the training is completed, use the optimal model to detect the test set to obtain the final result.

[0059] In this embodiment, after the above steps, the overall network structure of MSG-YOLOv7 constructed by the present invention is as follows: Figure 2 shown.

[0060] The remote sensing image military tank dataset in step S1 comes from the open source Google Earth data. The data samples are as follows: Figure 3 As shown in the figure, it contains 4000 pictures of different heights, orientations, and times, which are divided into 2400 training sets, 800 validation sets, and 800 test sets for experiments.

[0061] The lightweight MobileNetv3 in step S2 is specifically:

[0062] Using MobileNetv3 as the backbone network of MSG-YOLOv7 significantly reduces the number of model parameters and computational complexity. MobileNetv3 is a lightweight neural network structure designed specifically for image processing tasks. It is mainly composed of the Bneck structure, such as Figure 4As shown. This structure is mainly composed of dilated convolution, depthwise separable convolution (DSC), SE module and residual connection. In this structure, dilated convolution helps to increase the receptive field, depthwise separable convolution reduces the computational cost and promotes information flow, SE module can better model the relationship between features, and residual connection can fuse information between different feature layers. The integration of these modules can effectively improve the efficiency of feature learning and information transmission, help keep the model lightweight while providing good performance, and make the model more suitable for resource-constrained scenarios, such as edge devices or environments with limited computing power.

[0063] The depthwise separable convolution (DSC) in step S2 is specifically:

[0064] DSC is a commonly used operation in convolutional neural networks, which is designed to reduce the number of parameters and computation while maintaining model performance. Figure 5 As shown, it is mainly divided into two steps:

[0065] (1) Depthwise convolution, which performs a separate convolution operation on each input channel to capture local features;

[0066] (2) Point-by-point convolution, using a 1x1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel.

[0067] The formulas for the standard convolution parameter quantity P and the calculation quantity F are as follows:

[0068] P=D K ×D K ×M×N

[0069] F=D K ×D K ×M×N×D W ×D H

[0070] Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D H Represents high output characteristics.

[0071] The formulas for the parameter DP and calculation DF of DSC are as follows:

[0072] DP=D K ×D K ×M+M×N

[0073] DF=D K ×D K ×M×D W ×D H +M×N×DW ×D H

[0074] The comparison of the parameters and calculation amount between the two is as follows:

[0075]

[0076] The SD-MP module in step S3 is specifically:

[0077] The SD-MP module is mainly used to solve the problem of feature loss that may occur in the downsampling process of tank targets in remote sensing images, such as Figure 6 As shown in the figure. SD-MP uses the separation and splicing technology to first divide the input image into small blocks and reassemble them by channel, and then process them through convolution to capture richer feature information. This operation mode effectively alleviates the information loss that may be caused by downsampling. In addition, depthwise separable convolution is introduced to replace standard convolution to improve the efficiency and performance of the module.

[0078] The GD-ELAN module in step S4 is specifically:

[0079] GD-ELAN is an innovative module designed to overcome the challenges of object detection in complex backgrounds. Figure 7 Its core construction is based on GCNet and DSC, and its main function is to improve the network's perception of the overall background and its ability to express features, solving the problem of difficult detection of military targets in complex backgrounds.

[0080] The global context GCNet in step S4 is specifically:

[0081] GCNet is a global context module that can obtain global information in image processing, such as Figure 8 This global information is very important for correctly distinguishing the target from the background and understanding the relationship between the target and the surrounding environment. It is very effective in improving the accuracy and reliability of tank target detection. The calculation process is as follows:

[0082]

[0083] Y=X×σ(W v2 Relu(LN(W v1 Y1)))

[0084] In the formula, X i Represents any position information in the feature map, N p represents the number of all positions, W k represents the weight parameter, Softmax represents the normalization operation, and Y1 is the obtained feature matrix. Relu is the activation function, LN is the layer normalization, and Wv1 , W v2 are the weight parameters of the two convolutional layers v1 and v2, σ is the sigmoid function, and Y is the final feature map.

[0085] The model training in step S5 is specifically as follows:

[0086] The experiment was conducted in Ubuntu 16.04, using Python 3.8 and PyTorch 1.9.0 as the main development tools. In terms of hardware configuration, GeForce RTX 3090 GPU and Core i9-10900 CPU were used, equipped with CUDA11.1, and the system memory capacity was 64GB. The image size was set to 640×640 in model training, and the SGD optimizer was used, the batch size was 16, and the number of training iterations was 200.

[0087] In step S5, the model detection result is obtained, specifically:

[0088] The improved algorithm is compared with different algorithms, and the weight size (Weight), number of frames per second (FPS), number of parameters (Params), amount of calculation (GFlops) and average precision AP (Average Precision) are selected to evaluate the model performance. The AP calculation formula is as follows:

[0089]

[0090]

[0091]

[0092] In the formula, P is precision, R is recall, TP is the number of correctly detected samples, FP is the number of samples that are falsely detected as positive samples, and FN is the number of samples that are falsely detected as negative samples.

[0093] In step S5, the improved algorithm is compared with different algorithms, specifically:

[0094] In order to accurately verify the performance of the MSG-YOLOv7 algorithm in the remote sensing tank detection task, it is compared with the SSD, YOLOv3, YOLOv4, YOLOv5, YOLOv7, YOLO-FSD and DCN-YOLO algorithms, and the above algorithms are trained and tested under the same conditions. The final experimental results are shown in Table 1.

[0095] MSG-YOLOv7 achieved 99.02% in AP (average precision) indicator, showing that MSG-YOLOv7 performed very well in target detection accuracy. In terms of model size, MSG-YOLOv7 is only 28.6MB, the smallest of all models, which shows that the model requires the least space for storage and deployment. In terms of GFlops (computational power), MSG-YOLOv7 is 18.59GFlops, the lowest of all models. This shows that MSG-YOLOv7 has a lighter computational load during inference. In terms of FPS (frames per second), MSG-YOLOv7 reaches 41 frames, the highest frame rate among all models, showing its advantage in speed, which can perform remote sensing tank target detection more quickly.

[0096] Table 1 Performance comparison with different target detection algorithms

[0097]

[0098]

[0099] Note: Bold indicates the optimal value.

[0100] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7, characterized in that: The specific steps are as follows: S1: Obtain the remote sensing military tank dataset publicly available on Google Earth and divide it into training set, validation set, and test set in a ratio of 6:2:2; S2: In the YOLOv7 network, lightweight MobileNetv3 is used to replace the original Backbone to reduce the size of the model; S3: Design an SD-MP module that is more suitable for remote sensing targets and reduce the feature loss of tank targets during downsampling; S4: GD-ELAN module is designed based on GCNet and depthwise separable convolution, which enhances the representation ability of the model while making it lightweight; S5: Input the data set into the network for iterative training. After the training is completed, use the optimal model to test the test set to obtain the final result; In the SD-MP module, S represents separation and splicing, and D represents a DBS module, including a combination of depth-separable convolution, batch normalization, and SILU activation function, which is used to further process and enrich the reorganized feature information. In the SD-MP module, the input image is subjected to feature extraction through two branches A respectively, and then the features processed by the two branches A are spliced ​​in the channel dimension, and finally the merged feature map is output. At this point, the feature loss problem occurring in the downsampling process is effectively alleviated, so that the module can capture richer detail information; wherein, the two branches A are specifically as follows: The first branch A adopts a separation and splicing method, which first divides the input feature map into 4 smaller sub-blocks and recombines them along the channel dimension to form a richer feature expression. Then, the recombined features are further integrated through a convolution, and the recombined feature maps are sequentially passed through a DBS module for feature extraction. The second branch A processes the input feature map using two consecutive DBSs to extract fine-grained local features layer by layer; In the GD-ELAN module, G represents GCNet, D represents depthwise separable convolution, DBS_GC represents the combination of depthwise separable convolution, batch normalization, SILU activation function and GCNet, and in the GD-ELAN module structure, its input features will pass through two branches B, and then all the feature maps obtained will be spliced ​​to form a comprehensive feature map containing local and global information, and then further processed by the DBS module to obtain the final feature map; wherein, the two branches B are as follows: The first branch B processes the input feature map through a DBS_GC module, first extracts fine-grained local features using the DBS module, and then captures the global context information in the image through the GCNet module to achieve comprehensive integration of feature information; The second branch B uses two DBS modules in sequence to extract rich local features, which are then further processed in the DBS_GC module to achieve deep fusion of local and global features. The features of each step in the extraction process are passed to the splicing module to ensure that the multi-scale features can fully represent the target details.

2. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 1, characterized in that: The remote sensing image military tank dataset in S1 is specifically: the open source Google Earth data is obtained, which contains 4,000 pictures of different heights, orientations, and times, and is divided into a training set of 2,400 pictures, a validation set of 800 pictures, and a test set of 800 pictures for experiments.

3. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 1, characterized in that: The lightweight MobileNetv3 is used in S2, specifically: MobileNetv3 is used as the backbone network of MSG-YOLOv7, which significantly reduces the number of parameters and computational complexity of the model. MobileNetv3 is a lightweight neural network structure designed specifically for image processing tasks, which consists of a Bneck structure, which consists of dilated convolution, depthwise separable convolution, SE module and residual connection. In this structure, dilated convolution helps to increase the receptive field, depthwise separable convolution reduces computational cost and promotes information flow, while SE module can better model the relationship between features, and residual connection can fuse information between different feature layers. The integration of these modules can effectively improve the efficiency of feature learning and information transmission, help keep the model lightweight while providing good performance, and make the model more suitable for resource-constrained scenarios.

4. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 3, characterized in that: The depthwise separable convolution is a commonly used operation in convolutional neural networks. It is designed to reduce the number of parameters and computation while maintaining model performance. It is divided into two steps: Step 1: Depth convolution, performing a separate convolution operation on each input channel to capture local features; Step 2: Point-by-point convolution, using a 1×1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel; Among them, the formulas for the standard convolution parameter quantity P and the calculation quantity F are as follows: P=D K ×D K ×M×N F=D K ×D K ×M×N×D W ×D H Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D H Represents high output characteristics; The parameter amount DP and computation amount DF formula of depth-separable convolution are as follows: DP=D K ×D K ×M+M×N DF=D K ×D K ×M×D W ×D H +M×N×D W ×D H The comparison of the parameters and calculation amount between the two is as follows:

5. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 1, characterized in that: The GCNet is a global context module that can obtain global information in image processing. This global information is very important for correctly distinguishing the target from the background and understanding the relationship between the target and the surrounding environment. It is very effective in improving the accuracy and reliability of tank target detection. The calculation process is as follows: Y=X×σ(W v2 Relu(LN(W v1 Y1))) In the formula, X i Represents any position information in the feature map, N p represents the number of all positions, W k represents the weight parameter, Softmax represents the normalization operation, Y1 is the obtained feature matrix, Relu is the activation function, LN is the layer normalization, W v1 , W v2 are the weight parameters of the two convolutional layers v1 and v2, σ is the sigmoid function, and Y is the final feature map.

6. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 1, characterized in that: The model training in S5 is specifically as follows: the experiment was carried out in the Ubuntu 16.04 environment, using Python 3.8 and PyTorch 1.9.0 as development tools. In terms of hardware configuration, GeForce RTX3090 GPU and Core i9-10900 CPU were used, equipped with CUDA 11.1, and the system memory capacity was 64GB. The image size was set to 640×640 in the model training, and the SGD optimizer was used, the batch size was 16, and the number of training iterations was 200.

7. A lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 1, characterized in that: The model detection result is obtained in S5 as follows: the improved algorithm is compared with different algorithms, and the weight, number of frames per second, number of parameters, amount of calculation and average precision AP indicators are selected to evaluate the model performance. The AP calculation formula is as follows: In the formula, P is precision, R is recall, TP is the number of correctly detected samples, FP is the number of samples that are falsely detected as positive samples, and FN is the number of samples that are falsely detected as negative samples.

8. The lightweight remote sensing image military tank target detection algorithm based on MSG-YOLOv7 according to claim 7 is characterized in that: In S5, the improved algorithm is compared with different algorithms. Specifically, in order to accurately verify the performance of the MSG-YOLOv7 algorithm in the remote sensing tank detection task, it is compared with the SSD, YOLOv3, YOLOv4, YOLOv5, YOLOv7, YOLO-FSD and DCN-YOLO algorithms, and the above algorithms are trained and tested under the same conditions.

Images