Single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicle
By building a dynamic channel pruning system, the adaptability problem of low-altitude UAVs to changes in resource status during flight is solved, dynamic adjustment of model accuracy and efficiency is achieved, and the computing power and energy consumption management of UAVs are improved.
Patent Information
- Application Number
- CN202510977675.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-10-17
AI Technical Summary
Existing technologies are unable to dynamically adjust pruning strategies based on the resource status during low-altitude UAV flight and lack resilience, resulting in degraded model performance and difficulty in adapting to resource changes.
A single-stage dynamic channel pruning system for low-altitude UAVs was constructed, including a channel pool construction module, a resource-aware initialization module, a channel scoring module, a mask optimization module, and a mask execution module. Reversible pruning was achieved through resource awareness and channel scoring predictor, and channel retention or pruning was dynamically adjusted.
Dynamic channel adjustment is achieved based on the real-time resource status and mission requirements of the drone, ensuring model accuracy, reducing inference latency and energy consumption, and improving the robustness and flexibility of the model.
Smart Images
Figure CN120806023A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of dynamic channel pruning, and particularly relates to a single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicles. BACKGROUND
[0002] With the rapid development of artificial intelligence technology, deep neural networks have shown excellent performance in tasks such as target detection and image recognition, and are widely used in low-altitude unmanned aerial vehicle inspection, monitoring, security and other scenarios. However, mainstream neural network models have large parameter quantities and high computational complexity, making it difficult to directly deploy them on unmanned aerial vehicle platforms with limited computing power and power consumption. To reduce computational overhead, related practitioners have proposed channel pruning technology to compress network structures and improve inference efficiency. Traditional pruning methods mostly use static strategies, which first prune redundant channels through importance evaluation, and then retrain to restore performance. This "offline pruning + static deployment" approach lacks flexibility and cannot adapt to changes in resource state during flight.
[0003] In addition, most current pruning methods are irreversible, and once important channels are mistakenly pruned, the model performance will decline and it will be difficult to recover. At the same time, existing methods often cannot dynamically adjust the pruning strength according to changes in power, communication bandwidth limitations or task complexity fluctuations during the flight of the unmanned aerial vehicle. SUMMARY
[0004] The application provides a single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicles, which solves the technical problem of being unable to dynamically adjust the pruning strategy according to the resource state during flight and lacking recoverability in related technologies.
[0005] The application provides a single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicles, which includes:
[0006] A channel pool construction module for extracting channel kernel weights of each convolutional layer in a target deep neural network, constructing a channel feature set according to the kernel weights, and training a channel score predictor;
[0007] A resource-aware initialization module for obtaining current resource state information of the unmanned aerial vehicle and calculating a pruning initial threshold value through a resource-threshold mapping function according to the resource state information;
[0008] A channel score module for obtaining activation feature maps and corresponding kernel weights output by each channel of the target deep neural network under the current input image, and calculating the importance score of each channel based on the activation feature maps and kernel weights through the channel score predictor, comparing the importance score with the pruning initial threshold value, and generating a channel mask label table for identifying whether each channel is retained;
[0009] The mask optimization module is configured to sample a plurality of mask combination schemes based on the channel mask label table, perform running cost calculation on each mask combination scheme, construct a cost function based on a multi-objective optimization strategy, and calculate a comprehensive cost function value of each mask combination scheme, and select a mask combination scheme with the minimum comprehensive cost function value as an optimal mask combination scheme.
[0010] The mask execution module is configured to perform reversible pruning operation on the target deep neural network according to the optimal mask combination scheme.
[0011] Further, the channel pool construction module further comprises:
[0012] The channel parameter extraction unit is configured to obtain channel weight parameters of each convolutional layer in the target deep neural network, and calculate L2 norms of the channels based on the channel weight parameters;
[0013] The activation feature statistics unit is configured to input a plurality of preset sample images to the target deep neural network, and extract activation feature maps output by each channel; and perform statistical processing on the activation feature maps to obtain activation mean, activation standard deviation and non-zero activation ratio, wherein the non-zero activation ratio represents a proportion of non-zero elements in the activation feature maps.
[0014] The channel feature fusion unit is configured to combine the L2 norm of each channel with the corresponding activation feature to form a channel feature vector, and to aggregate all channel feature vectors to form a channel pool.
[0015] The channel score predictor construction unit is configured to train a channel score predictor for scoring based on the channel feature vectors and corresponding training labels.
[0016] Further, the training labels are obtained by respectively pruning each channel and calculating the decline amplitude of the accuracy of the pruned target deep neural network, wherein the accuracy is calculated by the ratio between the number of correctly recognized samples of the target deep neural network on the preset sample images and the total number of sample images.
[0017] Further, the resource state information comprises maximum available computing power of the processor, current remaining power and communication bandwidth.
[0018] Further, the computing load ratio is obtained by performing ratio operation on the preset computing power requirement and the maximum available computing power of the processor, the power tension degree is obtained by performing ratio operation on the preset power safety threshold and the current remaining power, the communication resource occupancy ratio is obtained by performing ratio operation on the preset communication bandwidth requirement and the current communication bandwidth, and finally the pruning initial threshold is obtained by combining the computing load ratio, the power tension degree and the communication resource occupancy ratio.
[0019] Further, the importance score comprises a domain score and an activation domain score, wherein the domain score is represented by an L2 norm of the channel, and the activation domain score is calculated by computing a pixel value standard deviation of an activation feature map output by the channel for the input image and taking an average of the pixel value standard deviation; further, the importance score is obtained by weighted sum of the domain score and the activation domain score.
[0020] Further, the specific steps of the mask optimization module comprise:
[0021] S201, perturbing the current channel mask marking table and sampling to generate a plurality of mask combination schemes, each mask combination scheme representing a configuration vector of different channel reservation or clipping states;
[0022] S202, constructing a cost function based on a multi-objective optimization strategy;
[0023] S203, calculating a comprehensive cost function value of each mask combination scheme according to the cost function;
[0024] S204, selecting the scheme with the minimum comprehensive cost function value as the optimal mask combination scheme.
[0025] Further, the calculation formula of the cost function is: ;
[0026] Wherein, comprehensive cost function value of the mth mask combination scheme, m represents the mask combination scheme index, 、 and respectively represent a time weight coefficient, a power consumption weight coefficient and an accuracy weight coefficient, and the sum of the three is 1, representing the predicted inference time of the mth mask combination scheme, representing the power consumption prediction value of the mth mask combination scheme, representing the prediction accuracy of the mth mask combination scheme.
[0027] Further, the reversible pruning operation comprises performing a freezing process on the channels marked as clipped by the optimal mask combination scheme, and the freezing process comprises skipping the convolution calculation of the channel and retaining the convolution kernel parameters thereof.
[0028] Further, when the accuracy of the target deep neural network is lower than a preset accuracy threshold, the gradient L2 norm of the channel pruned by the reversible pruning operation is calculated, and when the gradient L2 norm is greater than a preset norm threshold, the corresponding channel is restored to a reserved state.
[0029] The beneficial effects of the present application are that the present application integrates resource-aware initialization, channel score prediction, mask perturbation optimization, reversible pruning execution and other modules, and constructs a closed-loop pruning mechanism with adjustable structure and controllable performance. The present application can dynamically adjust the channel according to the real-time resource state and task demand of the unmanned aerial vehicle, which not only guarantees the model accuracy, but also significantly reduces the inference delay and energy consumption. The present application introduces a gradient norm driven accuracy adaptive recovery mechanism, which effectively avoids the mis-pruning of key channels and improves the robustness. BRIEF DESCRIPTION OF DRAWINGS
[0030] Figure 1 is a module schematic diagram of the single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicles of the present application. DETAILED DESCRIPTION
[0031] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described with respect to some examples can be combined in other examples.
[0032] As shown in Figure 1 The single-stage dynamic channel pruning system for low-altitude unmanned aerial vehicles includes:
[0033] The channel pool construction module 101 is configured to extract channel convolution kernel weights of each convolution layer in a target deep neural network, construct a channel feature set according to the convolution kernel weights, and train a channel score predictor;
[0034] The resource-aware initialization module 102 is configured to obtain current resource state information of the unmanned aerial vehicle, and calculate a pruning initial threshold value through a resource-threshold mapping function according to the resource state information;
[0035] The channel score module 103 is configured to obtain activation feature maps output by each channel and corresponding convolution kernel weights of the target deep neural network under a current input image, calculate an importance score of each channel based on the activation feature maps and the convolution kernel weights through the channel score predictor, compare the importance score with the pruning initial threshold value, and generate a channel mask marking table for identifying whether each channel is retained;
[0036] The mask optimization module 104 is configured to sample a plurality of mask combination schemes based on the channel mask marking table, calculate a running cost of each mask combination scheme, construct a cost function based on a multi-objective optimization strategy and calculate a comprehensive cost function value of each mask combination scheme, and select a mask combination scheme with the minimum comprehensive cost function value as the optimal mask combination scheme.
[0037] The mask execution module 105 is configured to perform reversible pruning operation on the target deep neural network according to the optimal mask combination scheme.
[0038] In an embodiment of the present application, the channel pool construction module is configured to extract and organize the channel hierarchy and its feature information for the target deep neural network model before system deployment, to establish a structured channel information set, i.e., a channel pool, and to construct a channel score predictor based on the set to support rapid pruning decision-making during flight; the target deep neural network model refers to a trained convolutional neural network.
[0039] In an embodiment of the present application, the channel pool construction module further comprises:
[0040] The channel parameter extraction unit is configured to obtain the channel weight parameters of each convolutional layer in the target deep neural network, and calculate the L2 norm of the channel based on the channel weight parameters, to represent the structural expression ability of the channel.
[0041] Specifically, for the output channel of each convolutional layer, the corresponding convolution kernel weight is a three-dimensional tensor, and the L2 norm can be calculated by flattening the tensor into a one-dimensional vector, and the calculation formula of the L2 norm is:
[0042] wherein, represents the L2 norm of the one-dimensional vector, J represents the number of parameters of the one-dimensional vector, j represents the parameter index, and i represents the channel index, represents the jth parameter of the ith channel; the calculation of the L2 norm of the channel provides an intuitive reflection of the structural complexity and expression ability of the channel, and provides basic structural information for subsequent channel importance evaluation.
[0043] The activation feature statistics unit is configured to input a plurality of preset sample images to the target deep neural network, and extract the activation feature maps output by each channel; and perform statistical processing on the activation feature maps to obtain the activation mean, the activation standard deviation, and the non-zero activation ratio, wherein the non-zero activation ratio represents the proportion of non-zero elements in the activation feature map, and is used to represent the response degree of the channel under input data.
[0044] wherein, the activation feature map refers to the output result of the convolutional layer channel after receiving the sample image during the forward inference process of the target deep neural network, and the non-zero activation ratio represents the proportion of non-zero elements in the activation feature map of the convolutional layer channel during the process of the target deep neural network processing the sample image, and is used to measure whether the channel is in an active state on the entire input sample, and the non-zero activation ratio is obtained by calculating the ratio of the number of non-zero elements to the total number of elements.
[0045] a channel feature fusion unit configured to combine the L2 norm of each channel with the corresponding activation feature to form a channel feature vector, and to aggregate all the channel feature vectors to form a channel pool for subsequent channel scoring and channel selection in the pruning control process, wherein the channel feature vector comprises the L2 norm of the channel, the activation mean, the activation standard deviation and the non-zero activation ratio;
[0046] a channel score predictor construction unit configured to train a channel score predictor for scoring based on the channel feature vectors and their corresponding training labels, the channel score predictor being configured to quickly evaluate the importance of the channels during flight.
[0047] Specifically, the training process comprises: inputting a plurality of preset sample images into the target deep neural network, and extracting channel feature vectors of all channels as training inputs of the score predictor;
[0048] For each channel, the training label is obtained by respectively pruning each channel and calculating the decline amplitude of the accuracy of the target deep neural network after pruning, wherein the accuracy is calculated by the ratio between the number of correctly identified samples of the target deep neural network on the preset sample images and the total number of samples of the preset sample images, and the correctly identified sample refers to a sample whose identification result of the neural network is consistent with its corresponding true label;
[0049] The least mean square error loss function is used for regression model training, and the training target is to predict the influence degree of the channel on the model accuracy;
[0050] The score predictor is implemented by using a multilayer perceptron.
[0051] After the training is completed, the channel score predictor can quickly predict the importance score of each channel in real time according to the channel feature vector during flight, for guiding the channel pruning decision.
[0052] In an embodiment of the present application, the resource state information comprises: maximum available computing power of the processor, current remaining power and communication bandwidth. The maximum available computing power of the processor is obtained by querying the task scheduling system, the current remaining power is obtained by the power management unit on the flight platform, and the communication bandwidth is obtained by the network interface.
[0053] In an embodiment of the present application, the calculation formula of the resource-threshold mapping function is: ;
[0054] wherein, represents the initial pruning threshold, and represents the threshold for preliminarily judging whether to retain a certain channel in the pruning process of the target deep neural network. 、 and respectively represent a first weight coefficient, a second weight coefficient and a third weight coefficient, represents a preset computing power requirement, represents a maximum available computing power of a processor, represents a preset power safety threshold, represents a current remaining power, represents a preset communication bandwidth requirement, represents a communication bandwidth.
[0055] In an embodiment of the application, the importance score comprises a domain score and an activation domain score, wherein the domain score is represented by an L2 norm of the channel, and the activation domain score is obtained by calculating a pixel value standard deviation of an activation feature map output by the channel for the input image and averaging the standard deviation; further, the importance score is obtained by weighted summation of the domain score and the activation domain score.
[0056] Specifically, the activation domain score is used to reflect the change in response intensity of the channel under different input images, for each channel, the activation feature map output by the channel under consecutive N input images is obtained, and the pixel standard deviation of each frame of the activation map is calculated to obtain the activity degree of the channel response of the frame. The standard deviation values of all frames are averaged to obtain the activation domain score of the channel. The calculation formula of the activation domain score is: ;
[0057] wherein, represents the activation feature map of the i th channel under the t th image, N represents the number of image frames, i represents the channel index, t represents the frame index, H and W respectively represent the height and width of the activation feature map, j represents the row index of the activation feature map, and k represents the column index of the activation feature map, represents the value of the i th channel at the j th row and the k th column in the t th frame, represents the pixel mean value of the i th channel in the t th frame.
[0058] The calculation formula of the importance score is: ;
[0059] wherein, represents the importance score of the i th channel, and respectively represent a third weight coefficient and a fourth weight coefficient, represents the domain score of the i th channel, i.e. the L2 norm.
[0060] In an embodiment of the present application, the channel score module determines whether to retain each channel according to the relationship between the importance score of the channel and the pruning initial threshold; specifically, when the importance score of the channel is greater than the pruning initial threshold, the channel mask of the channel is marked as 1, indicating that the channel is retained; when the importance score of the channel does not exceed the pruning initial threshold, the channel mask of the channel is marked as 0, indicating that the channel is pruned; the channel mask marks of all channels constitute a channel mask mark table, which is used as the basis for the subsequent reversible pruning operation of the mask execution module. By comparing the channel importance score with the pruning initial threshold, a structure optimization decision based on task resource perception and comprehensive trade-off of channel characteristics can be realized. The mask marking mechanism has the characteristics of small computational overhead and high execution efficiency, and is suitable for real-time pruning control in dynamic scenarios.
[0061] In an embodiment of the present application, the specific steps of the mask optimization module include:
[0062] S201, perturbing the current channel mask mark table and sampling to generate a plurality of mask combination schemes, each mask combination scheme representing a configuration vector of different channel retention or pruning states;
[0063] Specifically, a new mask combination scheme is generated by combining a Gaussian distribution with thresholding, and the specific steps include:
[0064] S301, regarding the current channel mask mark table as a set of initial channel scores, and adding a normally distributed noise term based thereon to obtain a perturbed configuration vector;
[0065] S302, performing thresholding processing on the perturbed configuration vector to convert it into a new mask combination scheme, the thresholding processing including: when the parameter of the perturbed configuration vector exceeds a first preset threshold, replacing it with 1, and when the parameter of the perturbed configuration vector is lower than the first preset threshold, replacing it with 0, and combining them into a new mask combination scheme;
[0066] In this way, a plurality of mask combinations with differences can be generated in the local structure space near the original mask for subsequent running cost evaluation and multi-objective optimization selection;
[0067] S202, constructing a cost function based on a multi-objective optimization strategy;
[0068] S203, calculating the comprehensive cost function value of each mask combination scheme according to the cost function;
[0069] S204, selecting the scheme with the minimum comprehensive cost function value as the optimal mask combination scheme.
[0070] In an embodiment of the present application, the calculation formula of the cost function is: ;
[0071] wherein, represents the comprehensive cost function value of the mth mask combination scheme, m represents a mask combination scheme index, 、 and respectively represent a time weight coefficient, a power consumption weight coefficient and an accuracy weight coefficient, and the sum of the three is 1, represents the predicted inference time of the mth mask combination scheme, represents the power consumption prediction value of the mth mask combination scheme, represents the prediction accuracy of the mth mask combination scheme;
[0072] In the mask optimization process, the running cost parameters of each group of mask combination schemes are predicted and estimated, the inference time and power consumption can be calculated based on the channel reservation ratio and the preset delay / power consumption model; the recognition accuracy can be predicted by the channel importance score. The above estimates can be obtained without relying on the real deployment model, and have engineering feasibility.
[0073] In an embodiment of the present application, the reversible pruning operation includes performing a freezing process on the channels marked for pruning by the optimal mask combination scheme, the freezing process including skipping the convolution calculation of the channel and retaining its convolution kernel parameters; by performing a freezing process on the pruned channels, the convolution calculation is skipped but the convolution kernel parameters are retained, so that the channels can be quickly recovered according to the task requirements or the change in recognition accuracy. This method does not require retraining, reduces the cost of mistaken pruning, improves the flexibility and robustness of the pruning strategy, and is suitable for resource-constrained and dynamically changing flight platform deployment.
[0074] When the accuracy of the target deep neural network is lower than the preset accuracy threshold, the gradient L2 norm of the channel pruned by the reversible pruning operation is calculated, and when the gradient L2 norm is greater than the preset norm threshold, the corresponding channel is restored to the reserved state; wherein the gradient L2 norm refers to the Euclidean norm of the gradient tensor of the loss function of the target deep neural network with respect to the convolution kernel parameters of a certain channel under the current input sample, which is used to measure the sensitivity of the channel to the model output error. By introducing the gradient L2 norm as a channel importance indicator when the recognition accuracy decreases, it is dynamically determined whether the pruned channel needs to be restored, and the adaptive structure adjustment according to the accuracy is realized. Compared with static pruning, key channels can be recovered more accurately, the stability and intelligence of the model in complex environments are improved, and the dual requirements of real-time performance and accuracy for flight tasks are met.
[0075] It should be noted that the interval and the threshold size are set for the convenience of comparison, wherein the size of the threshold depends on how much sample data and the base number set by the person skilled in the art for each group of sample data, as long as it does not affect the proportional relationship between the parameters and the quantized values. And the above formula is the calculation of the dimensionless value, the formula is obtained by collecting a large amount of data to simulate the formula of the nearest real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.
[0076] The above describes the embodiments of the present application, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative but not limiting, and the person skilled in the art can make many forms under the inspiration of the present embodiment, which all belong to the protection of the present embodiment.
Claims
1. A single-stage dynamic channel pruning system for low-altitude UAVs, characterized by: include: A channel pool construction module is used to extract the channel convolution kernel weights of each convolution layer in the target deep neural network, construct a channel feature set based on the convolution kernel weights, and train a channel score predictor; The resource perception initialization module is used to obtain the current resource status information of the UAV and calculate the initial pruning threshold through the resource-threshold mapping function based on the resource status information; The channel scoring module is used to obtain the activation feature map and corresponding convolution kernel weight of each channel output of the target deep neural network under the current input image, and calculate the importance score of each channel based on the activation feature map and convolution kernel weight through the channel score predictor. The importance score is compared with the initial pruning threshold to generate a channel mask marking table to identify whether each channel is retained; The mask optimization module is used to sample multiple mask combination schemes based on the channel mask marking table, calculate the running cost of each mask combination scheme, construct a cost function based on the multi-objective optimization strategy, calculate the comprehensive cost function value of each mask combination scheme, and select the mask combination scheme with the smallest comprehensive cost function value as the optimal mask combination scheme; The mask execution module is used to perform reversible pruning operations on the target deep neural network according to the optimal mask combination scheme.
2. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: The channel pool building module also includes: A channel parameter extraction unit is used to obtain the channel weight parameters of each convolutional layer in the target deep neural network and calculate the L2 norm of the channel based on the channel weight parameters; An activation feature statistics unit is configured to input a plurality of preset sample images into a target deep neural network and extract an activation feature map output by each channel; perform statistical processing on the activation feature map to obtain an activation mean, an activation standard deviation, and a non-zero activation ratio, wherein the non-zero activation ratio represents the proportion of non-zero elements in the activation feature map; The channel feature fusion unit is used to combine the L2 norm of each channel with the corresponding activation feature to form a channel feature vector, and aggregate all channel feature vectors to form a channel pool; The channel score predictor construction unit is used to train a channel score predictor for scoring based on the channel feature vector and its corresponding training label.
3. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 2 is characterized in that: The training labels are obtained by pruning each channel separately and calculating the decrease in the accuracy of the target deep neural network after pruning, wherein the accuracy is calculated by the ratio between the number of correctly recognized samples of the target deep neural network on the preset sample image and the total number of samples of the preset sample image.
4. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: The resource status information includes: the maximum available computing power of the processor, the current remaining power and the communication bandwidth.
5. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: The computing load ratio is obtained by comparing the preset computing power requirement with the maximum available computing power of the processor. The power tension is obtained by comparing the preset power safety threshold with the current remaining power. The communication resource ratio is obtained by comparing the preset communication bandwidth requirement with the current communication bandwidth. Finally, the computing load ratio, power tension and communication resource ratio are combined to obtain the initial pruning threshold.
6. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 2 is characterized in that: The importance score includes a structural domain score and an activation domain score, wherein the structural domain score is represented by the L2 norm of the channel, and the activation domain score is obtained by calculating the standard deviation of the pixel values of the activation feature map output by the channel in the input image and averaging them; further, the importance score is obtained by weighted summation of the structural domain score and the activation domain score.
7. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: The specific steps of the mask optimization module include: S201, perturb the current channel mask marking table and sample to generate multiple mask combination schemes, each mask combination scheme represents a configuration vector of a different channel retention or pruning state; S202, constructing a cost function based on a multi-objective optimization strategy; S203, calculating the comprehensive cost function value of each mask combination scheme according to the cost function; S204: Select the solution with the smallest comprehensive cost function value as the optimal mask combination solution.
8. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 7 is characterized in that: The calculation formula of the cost function is: ； in, Represents the comprehensive cost function value of the mth mask combination scheme, m represents the mask combination scheme index, 、 and They represent the time weight coefficient, power consumption weight coefficient and accuracy weight coefficient respectively, and the sum of the three is 1. represents the prediction inference time of the mth mask combination scheme, represents the power consumption prediction value of the mth mask combination scheme, Represents the prediction accuracy of the mth mask combination scheme.
9. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: The reversible pruning operation includes performing a freezing process on the channel marked as pruned by the mask according to the optimal mask combination scheme, and the freezing process includes skipping the convolution calculation of the channel and retaining its convolution kernel parameters.
10. The single-stage dynamic channel pruning system for low-altitude UAVs according to claim 1 is characterized in that: When the accuracy of the target deep neural network is lower than a preset accuracy threshold, the gradient L2 norm of the channel pruned by the reversible pruning operation is calculated, and when the gradient L2 norm is greater than the preset norm threshold, the corresponding channel is restored to a retained state.