Unmanned aerial vehicle face and signal recognition safety system based on military artificial intelligence

Abstract

The invention relates to the technical field of military safety, and discloses an unmanned aerial vehicle face and signal recognition safety system based on military artificial intelligence, an unmanned aerial vehicle is used for collecting images and signal data in real time, and a face recognition module is used for receiving a face initial image; the initial face image is preprocessed and input into a face recognition model for face recognition, and a face recognition result is obtained; the signal identification module is used for receiving the radio signal, performing preprocessing and feature extraction on the radio signal, and inputting the radio signal into a neural network model for identification to obtain a signal identification result; the block chain storage module is used for encrypting and recording each recognition process, a face recognition result and a signal recognition result into a block chain; and the decision control module is used for adjusting a task strategy in real time, and ensuring that the unmanned aerial vehicle does not accidentally injure friends when executing the task. According to the invention, the identification accuracy and safety of the unmanned aerial vehicle in military actions are improved, and accidental injury to own people is effectively prevented.

Description

Technical Field

[0001] The present invention relates to the field of military security technology, and in particular to a military artificial intelligence-based unmanned aerial vehicle face and signal recognition security system. Background Art

[0002] In modern military operations, drones are increasingly used. However, drone security issues are becoming increasingly prominent. Traditional drone security systems mainly rely on passwords and encrypted communications, but these methods are easily cracked or interfered with. In addition, in a complex military environment, how to quickly and accurately identify target people and signals is crucial to ensuring the safety of military operations. Summary of the invention

[0003] The purpose of the present invention is to solve the above problems and to design a UAV face and signal recognition security system based on military artificial intelligence.

[0004] The present invention provides a military artificial intelligence-based unmanned aerial vehicle face and signal recognition security system, the military artificial intelligence-based unmanned aerial vehicle face and signal recognition security system comprises an unmanned aerial vehicle, a face recognition module, a signal recognition module, a blockchain storage module and a decision control module, wherein:

[0005] Unmanned aerial vehicles (UAVs) for real-time image and signal data collection, equipped with at least a high-definition camera and a signal receiver;

[0006] A face recognition module is used to receive an initial facial image of a target person collected by a drone, preprocess the initial facial image, obtain a preprocessed initial facial image through image grayscale processing, sharpening processing and block processing, build a face recognition model by combining IResNet50, MobileFaceNet and Efficient Transformer networks, input the preprocessed initial facial image into the face recognition model for face recognition, obtain a face recognition result, distinguish between friendly and enemy personnel based on the face recognition result, and mark the target person as a friendly person or an enemy person;

[0007] A signal recognition module is used to receive the radio signal collected by the signal receiver of the drone, pre-process the radio signal, extract features of the pre-processed radio signal using a variational mode decomposition algorithm, and then input the feature into a neural network model optimized based on a krill swarm algorithm for recognition, thereby obtaining a signal recognition result;

[0008] A blockchain storage module, used to encrypt and record each recognition process of the face recognition module and the signal recognition module, the face recognition result and the signal recognition result into the blockchain;

[0009] The decision control module is used to combine the face recognition result and the signal recognition result with the flight control system of the UAV to adjust the mission strategy in real time to ensure that the UAV will not accidentally injure friendly personnel when performing the mission.

[0010] Optionally, in a first implementation of the present invention, the face recognition module includes a grayscale processing submodule, a sharpening processing submodule and a block processing submodule, wherein:

[0011] A grayscale processing submodule is used to receive an initial facial image of a target person collected by a drone, perform image grayscale processing on the initial facial image, and obtain a grayscale image of the initial facial image by using a weighted average algorithm;

[0012] The sharpening submodule is used to sharpen the grayscale image of the initial facial image using the Sobel operator to enhance the edge and contour of the face, thereby obtaining:

[0013]

[0014] In the formula, G represents the gradient value of any pixel in the image, θ represents the gradient direction, and G x Represents the lateral gradient value, G y Indicates the longitudinal gradient value;

[0015] The block processing submodule is used to perform block processing on the initial facial image after sharpening processing, divide the initial facial image after sharpening processing into 9 sub-blocks of 3×3 in total, and convolve the grayscale values ​​of eight peripheral pixels with eight Kitch operators with any pixel point in each sub-block as the center to obtain eight edge gradient values, and calculate the average value of the eight edge gradient values. The preprocessed initial facial image is obtained based on the average value of the eight edge gradient values, wherein the preprocessed initial facial image is a feature histogram of each sub-block.

[0016] Optionally, in a second implementation of the present invention, the face recognition module further includes a channel attention mechanism processing submodule, a cross fusion submodule and a capture information submodule, wherein:

[0017] The channel attention mechanism processing submodule is used to extract multi-scale features from the image branch and the key point branch of the input image through the IResNet50 and MobileFaceNet networks, and use the channel attention mechanism to process the output features of each stage in the two branches to extract key point features with local information and image features with global information;

[0018] The cross-fusion submodule is used to feed the features extracted by the IResNet50 and MobileFaceNet networks into the Efficient Transformer network respectively, and fuse the image and key point features of the same scale through the cross-fusion mechanism;

[0019] The information capture submodule is used to add a receptive field enhancement module after the cross-fusion of the Efficient Transformer network. It constructs convolution kernels with different receptive fields through different void rates, aggregates multi-scale context information, and uses upsampling and downsampling operations of the feature pyramid network structure to capture global semantic information and local detail information, thereby realizing face recognition and outputting face recognition results.

[0020] Optionally, in a third implementation of the present invention, the channel attention mechanism processing submodule passes the input image through two pooling layers, combines the obtained features, and then uses the Sigmoid activation function to determine the weight value of each channel, multiplies the weight value of each channel by the feature map of the corresponding scale, and generates output features;

[0021] The cross-fusion submodule concatenates the image features and key point features according to the patch dimension, uses 1×1 convolution to adjust the number of channels of each scale feature map to 256, cross-fuses the concatenated features through a multi-head self-attention mechanism, normalizes the attention heads at each position so that the attention heads at different positions learn the corresponding weights, concatenates the output values ​​of each head and linearly projects them, and outputs the cross-fused features;

[0022] The information capture submodule is composed of five stacked branches. The second branch is a 1×1 convolution. The second to fourth branches are depth-separable convolutions. The fifth branch is a global average pooling, and then a 1×1 convolution is performed to adjust the number of channels and then upsampled to obtain global features. The features of the five branches are stacked in the channel dimension.

[0023] Optionally, in a fourth implementation of the present invention, the signal identification module includes a signal decomposition submodule, a model building submodule and a signal identification submodule, wherein:

[0024] The signal decomposition submodule is used to receive the radio signal collected by the signal receiver of the drone, pre-process the radio signal, decompose the pre-processed radio signal into several IMF components using the variational mode decomposition algorithm and obtain the permutation entropy value of the IMF component, arrange the permutation entropy value after the decomposition of a single signal in order to construct a feature vector, and construct all the permutation entropy values ​​to obtain a feature vector set, wherein the pre-processing process includes segmenting the signal stream in a window manner;

[0025] A model building submodule, used for inputting the feature vector set into a pre-built neural network model, wherein the pre-built neural network model is optimized based on a krill swarm algorithm, and assigning an optimal hyperparameter set obtained by using the optimal fitness value in the krill swarm algorithm population and its corresponding parameters to the neural network model;

[0026] The signal recognition submodule is used to recognize the radio signal through the neural network model, capture the information in the radio signal, adopt self-attention, residual connection feedforward neural network module and residual connection layer feature processing and signal recognition in the neural network model, and the last layer adopts normalization layer and full connection classification for processing to output the signal recognition result.

[0027] Optionally, in a fifth implementation of the present invention, the signal decomposition submodule includes a coarse-grained processing unit and a sorting unit, wherein:

[0028] The coarse-grained processing unit is used to obtain the original sequence of the preprocessed radio signal, perform coarse-grained processing on the original sequence, divide the original sequence into a plurality of sub-intervals, add all data points in each sub-interval and take the average value as a new data point:

[0029]

[0030] In the formula, y pd represents the multi-scale time series after coarse-grained processing, p s represents the scale parameter, x pd =x pd (1), X pd (2), …, x pd (N pd ) represents the original sequence, N pd represents the length of the original sequence, and l represents the number of the subinterval;

[0031] The sorting unit is used to spatially reconstruct the original sequence after coarse-graining processing to obtain a matrix, reconstruct the IMF components of each row element of the matrix, and sort the elements of each column in ascending order to obtain the permutation entropy value of the IMF component based on the information entropy:

[0032]

[0033] Where PE represents the permutation entropy, p j represents the probability of the jth arrangement, d w represents the embedding dimension and k represents the length of the sequence.

[0034] Optionally, in a sixth implementation of the present invention, the model building submodule includes, wherein:

[0035] Logistic chaotic mapping is used to generate the initial population of the krill swarm algorithm, and the number of krill individuals, the maximum number of iterations, the maximum guiding speed, the maximum foraging speed, the maximum diffusion speed, and the upper and lower boundaries are determined;

[0036] In the initial iteration, a group of krill populations under the initialization conditions are searched. After a single iteration, the average fitness of all IKs in the krill population is calculated, and random perturbation processing is performed on the IK at the local optimal position:

[0037]

[0038]

[0039] In the formula, r g represents a random value belonging to [0, 1], represents the optimal IK positioning of the tth iteration, G(0,1) represents a random value with a mean of 0 and a variance of 1. express The random disturbance of is, Step is the iteration step length of the krill population;

[0040] After the operation is completed, if the fitness of IK exceeds the original value, then this IK will replace the original IK and become the representative individual of the new generation. It will be updated iteratively to determine whether the predetermined number of iterations or the upper and lower boundaries are reached. If not, continue to iterate. If so, the optimal hyperparameter set obtained by the optimal fitness value in the krill swarm algorithm population and its corresponding parameters is assigned to the neural network model.

[0041] Optionally, in a seventh implementation of the present invention, the method for implementing the military artificial intelligence-based drone face and signal recognition security system comprises the following steps:

[0042] Receive the initial facial image of the target person collected by the drone, pre-process the initial facial image, obtain the pre-processed initial facial image through image grayscale processing, sharpening processing and block processing, build a face recognition model by combining IResNet50, MobileFaceNet and Efficient Transformer network, input the pre-processed initial facial image into the face recognition model for face recognition, obtain face recognition results, distinguish between friendly and enemy personnel based on the face recognition results, and mark the target person as a friendly person or an enemy person;

[0043] Receiving the radio signal collected by the signal receiver of the UAV, preprocessing the radio signal, extracting features of the preprocessed radio signal using a variational mode decomposition algorithm, and then inputting the features into a neural network model optimized based on a krill swarm algorithm for recognition, thereby obtaining a signal recognition result;

[0044] Encrypting each recognition process, the face recognition result and the signal recognition result and recording them in the blockchain;

[0045] The face recognition result and the signal recognition result are linked with the flight control system of the UAV to adjust the mission strategy in real time to ensure that the UAV will not accidentally injure friendly personnel when performing the mission.

[0046] Optionally, in an eighth implementation of the present invention, the receiving drone collects an initial facial image of the target person, preprocesses the initial facial image, and obtains the preprocessed initial facial image by grayscale processing, sharpening processing, and block processing, including:

[0047] The grayscale image of the initial facial image is sharpened using the Sobel operator to enhance the edge and contour of the face, and the following is obtained:

[0048]

[0049] In the formula, G represents the gradient value of any pixel in the image, θ represents the gradient direction, and G x Represents the lateral gradient value, G y Indicates the longitudinal gradient value;

[0050] The sharpened initial facial image is divided into 9 sub-blocks (3×3) after block processing. Taking any pixel in each sub-block as the center, the grayscale values ​​of eight peripheral pixels are convolved with eight Kitch operators to obtain eight edge gradient values, and the average of the eight edge gradient values ​​is calculated. The preprocessed initial facial image is obtained based on the average of the eight edge gradient values, where the preprocessed initial facial image is the feature histogram of each sub-block.

[0051] Optionally, in a ninth implementation of the present invention, the face recognition model is constructed by combining IResNet50, MobileFaceNet and Efficient Transformer networks, the preprocessed initial facial image is input into the face recognition model for face recognition, and a face recognition result is obtained. Based on the face recognition result, the enemy and friendly personnel are distinguished, and the target personnel are marked as friendly personnel or enemy personnel, including:

[0052] The input image is extracted from the image branch and the key point branch through the IResNet50 and MobileFaceNet networks, and the channel attention mechanism is used to process the output features of each stage in the two branches to extract the key point features with local information and the image features with global information;

[0053] The features extracted by IResNet50 and MobileFaceNet networks are fed into the EfficientTransformer network respectively, and the image and key point features of the same scale are fused through the cross fusion mechanism;

[0054] After the cross-fusion of the Efficient Transformer network, a receptive field enhancement module is added. Convolution kernels with different receptive fields are constructed through different void rates, multi-scale context information is aggregated, and upsampling and downsampling operations of the feature pyramid network structure are used to capture global semantic information and local detail information, thereby realizing face recognition and outputting face recognition results.

[0055] The embodiments of the present invention include the following advantages: by collecting image and signal data in real time through the drone, and utilizing the high-definition camera and signal receiver equipped therewith, it is possible to quickly obtain multi-dimensional information in a complex military environment, providing a rich data basis for subsequent identification and decision-making; it is possible to accurately perform face recognition, distinguish between enemy and friendly personnel and mark them, greatly improving the accuracy and efficiency of personnel identification in military operations, effectively avoiding accidental injury to friendly personnel, and ensuring the safety of friendly personnel; it is possible to accurately identify various radio signals, provide important intelligence support for military operations, help to timely grasp the enemy's dynamics, and formulate more scientific and reasonable combat strategies; the data is encrypted and recorded in the blockchain, ensuring the security and non-tamperability of the data, and providing a reliable basis for subsequent data analysis and review; it is linked with the flight control system of the drone to adjust the mission strategy in real time, ensuring that the drone can act more intelligently and efficiently when performing tasks, and improving the adaptability and safety of the drone in military missions. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0057] Figure 1 A schematic diagram of the structure of a UAV face and signal recognition security system based on military artificial intelligence provided by an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of the structure of a face recognition module provided by an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of the structure of a signal recognition module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0061] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 The structural diagram of the UAV face and signal recognition security system based on military artificial intelligence provided by an embodiment of the present invention includes a UAV, a face recognition module, a signal recognition module, a blockchain storage module and a decision control module, wherein:

[0062] Unmanned aerial vehicles (UAVs) for real-time image and signal data collection, equipped with at least a high-definition camera and a signal receiver;

[0063] The face recognition module is used to receive the initial facial image of the target person collected by the drone, preprocess the initial facial image, obtain the preprocessed initial facial image through image grayscale processing, sharpening processing and block processing, build a face recognition model by combining IResNet50, MobileFaceNet and Efficient Transformer networks, input the preprocessed initial facial image into the face recognition model for face recognition, obtain the face recognition result, distinguish between friendly and enemy personnel based on the face recognition result, and mark the target person as a friendly person or an enemy person;

[0064] The signal recognition module is used to receive the radio signal collected by the signal receiver of the UAV, pre-process the radio signal, use the variational mode decomposition algorithm to extract the features of the pre-processed radio signal, and then input it into the neural network model optimized by the krill swarm algorithm for recognition to obtain the signal recognition result;

[0065] A blockchain storage module, used to encrypt and record each recognition process, face recognition result and signal recognition result of the face recognition module and the signal recognition module into the blockchain;

[0066] The decision-making control module is used to combine the face recognition results and signal recognition results with the UAV's flight control system to adjust the mission strategy in real time to ensure that the UAV will not accidentally injure friendly personnel when performing missions.

[0067] See also Figure 2 , a schematic diagram of the structure of a face recognition module provided by an embodiment of the present invention, wherein the face recognition module includes a grayscale processing submodule, a sharpening processing submodule and a block processing submodule, wherein,

[0068] The grayscale processing submodule is used to receive the initial facial image of the target person collected by the drone, perform grayscale processing on the initial facial image, and obtain a grayscale image of the initial facial image using a weighted average algorithm;

[0069] The sharpening submodule is used to sharpen the grayscale image of the initial facial image using the Sobel operator to enhance the edge and contour of the face, and obtain:

[0070]

[0071] In the formula, G represents the gradient value of any pixel in the image, θ represents the gradient direction, and G x Represents the lateral gradient value, G y Indicates the longitudinal gradient value;

[0072] The block processing submodule is used to perform block processing on the initial facial image after sharpening processing, divide the initial facial image after sharpening processing into 9 sub-blocks of 3×3 in total, and convolve the grayscale values ​​of eight peripheral pixels with eight Kitch operators with any pixel point in each sub-block as the center to obtain eight edge gradient values, and calculate the average value of the eight edge gradient values. The preprocessed initial facial image is obtained based on the average value of the eight edge gradient values, wherein the preprocessed initial facial image is a feature histogram of each sub-block.

[0073] The face recognition module also includes a channel attention mechanism processing submodule, a cross fusion submodule and a capture information submodule, among which,

[0074] The channel attention mechanism processing submodule is used to extract multi-scale features from the image branch and the key point branch of the input image through the IResNet50 and MobileFaceNet networks, and use the channel attention mechanism to process the output features of each stage in the two branches to extract key point features with local information and image features with global information;

[0075] The cross-fusion submodule is used to feed the features extracted by the IResNet50 and MobileFaceNet networks into the Efficient Transformer network respectively, and fuse the image and key point features of the same scale through the cross-fusion mechanism;

[0076] The information capture submodule is used to add a receptive field enhancement module after the cross-fusion of the Efficient Transformer network. It constructs convolution kernels with different receptive fields through different void rates, aggregates multi-scale context information, and uses upsampling and downsampling operations of the feature pyramid network structure to capture global semantic information and local detail information, thereby realizing face recognition and outputting face recognition results.

[0077] In this embodiment, the channel attention mechanism processing submodule passes the input image through two pooling layers, combines the obtained features, and then uses the Sigmoid activation function to determine the weight value of each channel, multiplies the weight value of each channel by the feature map of the corresponding scale, and generates output features;

[0078] The cross-fusion submodule concatenates image features and key point features according to the patch dimension, uses 1×1 convolution to adjust the number of channels of each scale feature map to 256, and cross-fuses the concatenated features through a multi-head self-attention mechanism. The attention heads at each position are normalized so that the attention heads at different positions learn the corresponding weights. The output values ​​of each head are connected in series and linearly projected to output the cross-fused features.

[0079] The information capture submodule is composed of 5 stacked branches. The second branch is a 1×1 convolution. The second to fourth branches are depth-separable convolutions. The fifth branch is a global average pooling, and then a 1×1 convolution is performed to adjust the number of channels and then upsampled to obtain global features. The features of the 5 branches are stacked in the channel dimension.

[0080] See also Figure 3 , a schematic diagram of the structure of a signal recognition module provided by an embodiment of the present invention, wherein the signal recognition module comprises a signal decomposition submodule, a model building submodule and a signal recognition submodule, wherein,

[0081] The signal decomposition submodule is used to receive the radio signal collected by the signal receiver of the UAV, pre-process the radio signal, decompose the pre-processed radio signal into several IMF components using the variational mode decomposition algorithm and obtain the permutation entropy value of the IMF component, arrange the permutation entropy value after the decomposition of a single signal in order to construct a feature vector, and construct all the permutation entropy values ​​to obtain a feature vector set, wherein the pre-processing process includes segmenting the signal stream in a window manner;

[0082] A model building submodule is used to input the feature vector set into a pre-built neural network model, wherein the pre-built neural network model is optimized based on the krill swarm algorithm, and the optimal hyperparameter set obtained by the optimal fitness value in the krill swarm algorithm population and its corresponding parameters is assigned to the neural network model;

[0083] The signal recognition submodule is used to identify radio signals through a neural network model and capture the information in the radio signals. Self-attention, residual connection feedforward neural network modules and residual connection layer feature processing and signal recognition are used in the neural network model. The last layer uses a normalization layer and a fully connected classification for processing to output the signal recognition result.

[0084] In this embodiment, the signal decomposition submodule includes a coarse-grained processing unit and a sorting unit, wherein:

[0085] The coarse-grained processing unit is used to obtain the original sequence of the preprocessed radio signal, perform coarse-grained processing on the original sequence, divide the original sequence into several sub-intervals, add all the data points in each sub-interval and take the average value as the new data point:

[0086]

[0087] In the formula, y pd represents the multi-scale time series after coarse-graining processing, p s represents the scale parameter, x pd =x pd (1), x pd (2), …, x pd (N pd ) represents the original sequence, N pd represents the length of the original sequence, and l represents the number of the subinterval;

[0088] The sorting unit is used to spatially reconstruct the original sequence after coarse-graining processing to obtain a matrix, reconstruct the IMF components of each row element of the matrix, and sort the elements of each column in ascending order to obtain the permutation entropy value of the IMF component based on the information entropy:

[0089]

[0090] Where PE represents the permutation entropy, p j represents the probability of the jth arrangement, d w represents the embedding dimension and k represents the length of the sequence.

[0091] In this embodiment, the model building submodule includes, among which,

[0092] Logistic chaotic mapping is used to generate the initial population of the krill swarm algorithm, and the number of krill individuals, the maximum number of iterations, the maximum guiding speed, the maximum foraging speed, the maximum diffusion speed, and the upper and lower boundaries are determined;

[0093] In the initial iteration, a group of krill populations under the initialization conditions are searched. After a single iteration, the average fitness of all IKs in the krill population is calculated, and random perturbation processing is performed on the IK at the local optimal position:

[0094]

[0095]

[0096] In the formula, r g represents a random value belonging to [0, 1], represents the optimal IK positioning of the tth iteration, G(0,1) represents a random value with a mean of 0 and a variance of 1. express The random disturbance of is, Step is the iteration step length of the krill population;

[0097] After the operation is completed, if the fitness of IK exceeds the original value, then this IK will replace the original IK and become the representative individual of the new generation. It will be updated iteratively to determine whether the predetermined number of iterations or the upper and lower boundaries are reached. If not, continue to iterate. If so, the optimal hyperparameter set obtained by the optimal fitness value in the krill swarm algorithm population and its corresponding parameters is assigned to the neural network model.

[0098] Through the implementation of the above solution, the problems of low drone recognition accuracy and easy data tampering in the existing technology have been solved, the recognition accuracy and safety of drones in military operations have been improved, and accidental injuries to one's own personnel have been effectively prevented.

[0099] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A drone face and signal recognition security system based on military artificial intelligence, characterized in that: The military artificial intelligence-based drone face and signal recognition security system includes a drone, a face recognition module, a signal recognition module, a blockchain storage module and a decision control module, wherein: Unmanned aerial vehicles (UAVs) for real-time image and signal data collection, equipped with at least a high-definition camera and a signal receiver; A face recognition module is used to receive an initial facial image of a target person collected by a drone, preprocess the initial facial image, obtain a preprocessed initial facial image through image grayscale processing, sharpening processing and block processing, build a face recognition model by combining IResNet50, MobileFaceNet and Efficient Transformer networks, input the preprocessed initial facial image into the face recognition model for face recognition, obtain a face recognition result, distinguish between friendly and enemy personnel based on the face recognition result, and mark the target person as a friendly person or an enemy person; A signal recognition module is used to receive the radio signal collected by the signal receiver of the drone, pre-process the radio signal, extract features of the pre-processed radio signal using a variational mode decomposition algorithm, and then input the feature into a neural network model optimized based on a krill swarm algorithm for recognition, thereby obtaining a signal recognition result; A blockchain storage module, used to encrypt and record each recognition process of the face recognition module and the signal recognition module, the face recognition result and the signal recognition result into the blockchain; The decision control module is used to combine the face recognition result and the signal recognition result with the flight control system of the UAV to adjust the mission strategy in real time to ensure that the UAV will not accidentally injure friendly personnel when performing the mission.

2. A military artificial intelligence-based drone face and signal recognition security system as claimed in claim 1, characterized in that: The face recognition module includes a grayscale processing submodule, a sharpening processing submodule and a block processing submodule, wherein: A grayscale processing submodule is used to receive an initial facial image of a target person collected by a drone, perform image grayscale processing on the initial facial image, and obtain a grayscale image of the initial facial image by using a weighted average algorithm; The sharpening submodule is used to sharpen the grayscale image of the initial facial image using the Sobel operator to enhance the edge and contour of the face, thereby obtaining: In the formula, G represents the gradient value of any pixel in the image, θ represents the gradient direction, and G x Represents the lateral gradient value, G y Indicates the longitudinal gradient value; The block processing submodule is used to perform block processing on the initial facial image after sharpening processing, divide the initial facial image after sharpening processing into 9 sub-blocks of 3×3 in total, and convolve the grayscale values ​​of eight peripheral pixels with eight Kitch operators with any pixel point in each sub-block as the center to obtain eight edge gradient values, and calculate the average value of the eight edge gradient values. The preprocessed initial facial image is obtained based on the average value of the eight edge gradient values, wherein the preprocessed initial facial image is a feature histogram of each sub-block.

3. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 1, characterized in that: The face recognition module also includes a channel attention mechanism processing submodule, a cross fusion submodule and a capture information submodule, wherein: The channel attention mechanism processing submodule is used to extract multi-scale features from the image branch and the key point branch of the input image through the IResNet50 and MobileFaceNet networks, and use the channel attention mechanism to process the output features of each stage in the two branches to extract key point features with local information and image features with global information; The cross-fusion submodule is used to feed the features extracted by the IResNet50 and MobileFaceNet networks into the Efficient Transformer network respectively, and fuse the image and key point features of the same scale through the cross-fusion mechanism; The information capture submodule is used to add a receptive field enhancement module after the cross-fusion of the Efficient Transformer network. It constructs convolution kernels with different receptive fields through different void rates, aggregates multi-scale context information, and uses upsampling and downsampling operations of the feature pyramid network structure to capture global semantic information and local detail information, thereby realizing face recognition and outputting face recognition results.

4. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 3, characterized in that: The channel attention mechanism processing submodule passes the input image through two pooling layers, combines the obtained features, and then uses the Sigmoid activation function to determine the weight value of each channel, multiplies the weight value of each channel by the feature map of the corresponding scale, and generates output features; The cross-fusion submodule concatenates the image features and key point features according to the patch dimension, uses 1×1 convolution to adjust the number of channels of each scale feature map to 256, cross-fuses the concatenated features through a multi-head self-attention mechanism, normalizes the attention heads at each position so that the attention heads at different positions learn the corresponding weights, concatenates the output values ​​of each head and linearly projects them, and outputs the cross-fused features; The information capture submodule is composed of five stacked branches. The second branch is a 1×1 convolution. The second to fourth branches are depth-separable convolutions. The fifth branch is a global average pooling, and then a 1×1 convolution is performed to adjust the number of channels and then upsampled to obtain global features. The features of the five branches are stacked in the channel dimension.

5. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 1, characterized in that: The signal identification module includes a signal decomposition submodule, a model building submodule and a signal identification submodule, wherein: The signal decomposition submodule is used to receive the radio signal collected by the signal receiver of the drone, pre-process the radio signal, decompose the pre-processed radio signal into several IMF components using the variational mode decomposition algorithm and obtain the permutation entropy value of the IMF component, arrange the permutation entropy value after the decomposition of a single signal in order to construct a feature vector, and construct all the permutation entropy values ​​to obtain a feature vector set, wherein the pre-processing process includes segmenting the signal stream in a window manner; A model building submodule, used for inputting the feature vector set into a pre-built neural network model, wherein the pre-built neural network model is optimized based on a krill swarm algorithm, and assigning an optimal hyperparameter set obtained by using the optimal fitness value in the krill swarm algorithm population and its corresponding parameters to the neural network model; The signal recognition submodule is used to recognize the radio signal through the neural network model, capture the information in the radio signal, adopt self-attention, residual connection feedforward neural network module and residual connection layer feature processing and signal recognition in the neural network model, and the last layer adopts normalization layer and full connection classification for processing to output the signal recognition result.

6. A military artificial intelligence-based drone face and signal recognition security system as claimed in claim 5, characterized in that: The signal decomposition submodule includes a coarse-grained processing unit and a sorting unit, wherein: The coarse-grained processing unit is used to obtain the original sequence of the preprocessed radio signal, perform coarse-grained processing on the original sequence, divide the original sequence into a plurality of sub-intervals, add all data points in each sub-interval and take the average value as a new data point: In the formula, y pd represents the multi-scale time series after coarse-graining processing, p s represents the scale parameter, xpd = Xpd(1), xpd(2), ..., xpd(Npd) represents the original sequence, Npd represents the length of the original sequence, and l represents the number of the subinterval; The sorting unit is used to spatially reconstruct the original sequence after coarse-graining processing to obtain a matrix, reconstruct the IMF components of each row element of the matrix, and sort the elements of each column in ascending order to obtain the permutation entropy value of the IMF component based on the information entropy: Where PE represents the permutation entropy, p j represents the probability of the jth arrangement, d w represents the embedding dimension and k represents the length of the sequence.

7. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 5, characterized in that: The model building submodule includes, among which, Logistic chaotic mapping is used to generate the initial population of the krill swarm algorithm, and the number of krill individuals, the maximum number of iterations, the maximum guiding speed, the maximum foraging speed, the maximum diffusion speed, and the upper and lower boundaries are determined; In the initial iteration, a group of krill populations under the initialization conditions are searched. After a single iteration, the average fitness of all IKs in the krill population is calculated, and random perturbation processing is performed on the IK at the local optimal position: In the formula, r g represents a random value belonging to [0, 1], represents the optimal IK positioning of the tth iteration, G(0,1) represents a random value with a mean of 0 and a variance of 1. express The random disturbance of is, Step is the iteration step length of the krill population; After the operation is completed, if the fitness of IK exceeds the original value, then this IK will replace the original IK and become the representative individual of the new generation. It will be updated iteratively to determine whether the predetermined number of iterations or the upper and lower boundaries are reached. If not, continue to iterate. If so, the optimal hyperparameter set obtained by the optimal fitness value in the krill swarm algorithm population and its corresponding parameters is assigned to the neural network model.

8. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 1, characterized in that: The method for implementing the military artificial intelligence-based drone face and signal recognition security system includes the following steps: Receive the initial facial image of the target person collected by the drone, pre-process the initial facial image, obtain the pre-processed initial facial image through image grayscale processing, sharpening processing and block processing, build a face recognition model by combining IResNet50, MobileFaceNet and Efficient Transformer network, input the pre-processed initial facial image into the face recognition model for face recognition, obtain face recognition results, distinguish between friendly and enemy personnel based on the face recognition results, and mark the target person as a friendly person or an enemy person; Receiving the radio signal collected by the signal receiver of the UAV, preprocessing the radio signal, extracting features of the preprocessed radio signal using a variational mode decomposition algorithm, and then inputting the features into a neural network model optimized based on a krill swarm algorithm for recognition, thereby obtaining a signal recognition result; Encrypting each recognition process, the face recognition result and the signal recognition result and recording them in the blockchain; The face recognition result and the signal recognition result are linked with the flight control system of the UAV to adjust the mission strategy in real time to ensure that the UAV will not accidentally injure friendly personnel when performing the mission.

9. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 8, characterized in that: The receiving drone collects an initial facial image of the target person, preprocesses the initial facial image, and obtains the preprocessed initial facial image by grayscale processing, sharpening processing, and block processing, including: The grayscale image of the initial facial image is sharpened using the Sobel operator to enhance the edge and contour of the face, and the following is obtained: In the formula, G represents the gradient value of any pixel in the image, θ represents the gradient direction, and G x Represents the lateral gradient value, G y Indicates the longitudinal gradient value; The sharpened initial facial image is divided into 9 sub-blocks (3×3) after block processing. Taking any pixel in each sub-block as the center, the grayscale values ​​of eight peripheral pixels are convolved with eight Kitch operators to obtain eight edge gradient values, and the average of the eight edge gradient values ​​is calculated. The preprocessed initial facial image is obtained based on the average of the eight edge gradient values, where the preprocessed initial facial image is the feature histogram of each sub-block.

10. The military artificial intelligence-based drone face and signal recognition security system as claimed in claim 9, characterized in that: The face recognition model is constructed by combining IResNet50, MobileFaceNet and Efficient Transformer network, the preprocessed initial face image is input into the face recognition model for face recognition, and the face recognition result is obtained. Based on the face recognition result, the enemy and friendly personnel are distinguished, and the target personnel are marked as friendly personnel or enemy personnel, including: The input image is extracted from the image branch and the key point branch through the IResNet50 and MobileFaceNet networks, and the channel attention mechanism is used to process the output features of each stage in the two branches to extract the key point features with local information and the image features with global information; The features extracted by IResNet50 and MobileFaceNet networks are fed into the Efficient Transformer network respectively, and the image and key point features of the same scale are fused through the cross fusion mechanism; After the cross-fusion of the Efficient Transformer network, a receptive field enhancement module is added. Convolution kernels with different receptive fields are constructed through different void rates, multi-scale context information is aggregated, and upsampling and downsampling operations of the feature pyramid network structure are used to capture global semantic information and local detail information, thereby realizing face recognition and outputting face recognition results.

Images