Lightweight method and device for multi-type target detection of sar image combining channel pruning with knowledge distillation

Abstract

The application provides a channel pruning and knowledge distillation combined SAR image multi-type target detection lightweight method and device. The method comprises the following steps: step 1: training a complex model on a SAR image dataset, judging the importance of each channel in the complex model, and pruning the secondary channel by using channel pruning, and obtaining a pruned and optimized complex model through fine-tuning training; step 2: taking the pruned and optimized complex model as a teacher network, taking a lightweight model as a student network, constructing a knowledge distillation framework, training on the SAR image dataset, continuously updating the parameters of the lightweight model, and obtaining a lightweight SAR target detection weight model; and step 3: using the lightweight SAR target detection weight model as a SAR target detection network to detect targets in the SAR image to be detected.

Description

Technical Field

[0001] This invention relates to the field of intelligent target detection and recognition technology for synthetic aperture radar images, and in particular to a lightweight method and apparatus for multi-type target detection in SAR images that combines channel pruning and knowledge distillation. Background Technology

[0002] Driven by the rapid acquisition of massive amounts of SAR data and practical tasks, traditional SAR image target detection methods can no longer meet the current demands for timeliness and accuracy. Many experts and scholars at home and abroad have introduced computer vision and deep learning algorithms into this field. The most representative one is the Convolutional Neural Network (CNN), which learns shallow features of target texture and contour, as well as deep abstract semantic information of the target to achieve end-to-end target detection tasks. Compared with traditional methods, it is more suitable for detection tasks in complex scenes and has gradually become the mainstream method for SAR target detection at present.

[0003] Currently, deep learning-based target detection algorithms typically require powerful computing capabilities. In practical SAR target detection applications, these algorithms often need to meet the requirements of small size, easy portability, high accuracy, and high efficiency. Therefore, lightweight SAR target detection methods have become a current research focus. Researchers have mainly made lightweight improvements through network pruning, knowledge distillation, and network structure modification. Regarding network pruning, Zhou Zhou et al.'s paper "An Optimized Algorithm for Ship Detection in SAR Images Based on Channel Pruning," published in Volume 2, Issue 37 of *Shanghai Aerospace*, introduced channel pruning into the SAR target detection network, effectively reducing the number of model parameters, computational load, and inference time, but at the expense of detection accuracy. Regarding knowledge distillation, Fu Xiaoya's dissertation "Research on Ship Target Detection Algorithm in SAR Images Based on Scene Classification and Knowledge Distillation" used MobileNetV2 as the feature extraction network for the SSD detection algorithm, achieving lightweight improvements and using it as a student network. Knowledge distillation was also used to improve the performance of the lightweight network, but the detection accuracy still lagged behind the original network model.

[0004] Among the methods mentioned above, network pruning and knowledge distillation optimize network performance from the aspects of model compression and performance improvement, respectively. Some scholars have also improved the overall performance of the model from the perspective of integrating lightweight methods. Chen Shiqi et al.'s paper "Learning Slimming SAR Ship Object Detector Through NetworkPruning and Knowledge Distillation" published in Volume 14 of the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing combines the advantages of the above two methods. It uses network pruning to construct a lightweight network model and then uses knowledge distillation to improve the detection accuracy of the pruned lightweight network model. Similar work includes Zhan Ronghui et al.'s "SAR Ship Target Detection Method Based on Network Pruning and Knowledge Distillation".

[0005] However, in practical applications, channel pruning compression often comes at the cost of sacrificing detection accuracy. Excessive pruning can cause a serious drop in model accuracy, and pruning operations have a lower limit on the compression of model volume, which cannot fully meet the requirements of lightweight and high-precision SAR target detection. Although knowledge distillation can improve the detection accuracy of lightweight models to some extent, when the model capacity difference is too large, the student network cannot fully learn the complex knowledge of the teacher network, which limits the effect of accuracy improvement. Summary of the Invention

[0006] This invention addresses the issue of low accuracy in existing lightweight target detection algorithms for SAR images by designing a lightweight method and apparatus for multi-type target detection in SAR images that combines channel pruning and knowledge distillation. This method sparsifies the scaling factor of the batch normalization layer in a complex network during training, determines the importance of corresponding feature channels, prunes secondary channels, and uses these pruned channels as a teacher network after fine-tuning. A knowledge distillation framework is then constructed to guide the training of the lightweight model, thereby improving the detection accuracy of the lightweight model.

[0007] On the one hand, this invention provides a lightweight method for multi-type target detection in SAR images that combines channel pruning and knowledge distillation, including:

[0008] Step 1: Perform sparsity training on the complex model on the SAR image dataset, determine the importance of each channel in the complex model, and prune secondary channels by channel pruning. After fine-tuning the training, the pruned and optimized complex model is obtained.

[0009] Step 2: Use the pruned and optimized complex model as the teacher network and the lightweight model as the student network to construct a knowledge distillation framework. Train the model on a SAR image dataset to continuously update the parameters of the lightweight model and obtain a lightweight SAR target detection weight model.

[0010] Step 3: Use a lightweight SAR target detection weight model as the SAR target detection network to detect targets in the SAR image to be detected.

[0011] Furthermore, in step 1, the complex model is trained sparsely on the SAR image dataset, specifically including:

[0012] Using L1 regularization, the set Γ of scaling factor γ values ​​corresponding to each channel in the convolutional layer is sparsified, and the regularization formula of set Γ is introduced into the network loss function to obtain the modified network loss function:

[0013]

[0014] Where l(f(x,W),y) is the loss function of the complex model, (x,y) are the input and target of training, W is the trainable weight, λ is the balance factor, g(γ) is the sparse penalty term for the scaling factor γ, and g(γ)=||γ||1;

[0015] Take the partial derivative of formula (3) with respect to γ. In the optimization process of subgradient descent, make some scaling factors γ continuously approach 0 to obtain the sparsified set Γ.

[0016] Furthermore, in step 1, the importance of each channel in the complex model is determined, and secondary channels are pruned using channel pruning, specifically including:

[0017] Based on the scaling factor γ after sparse training, it is connected to the corresponding channel in the feature map of the convolutional layer, and the absolute values ​​are sorted. The pruning threshold S is determined by setting the pruning ratio. If the γ corresponding to a channel is less than the pruning threshold S, it is determined as a secondary channel and pruning is applied.

[0018] Furthermore, step 2 specifically includes:

[0019] The softmax output of the student network is used as the hard prediction, and the warmed prediction result of the softmax output of the student network is used as the soft prediction; the warmed prediction result of the softmax output of the teacher network is used as the soft label.

[0020] Construct a student loss L between hard labels and hard predictions. student Constructing a distillation loss L between soft prediction and soft labeling distill Thus, the total loss L = αL is constructed.distill +(1-α)L student ;

[0021] By continuously updating the parameters of the lightweight model based on the total loss L, a lightweight SAR target detection weight model is obtained.

[0022] Furthermore, student losses L student and distillation loss L distill The construction process includes:

[0023] Formula (8) is used as the target loss function.

[0024]

[0025] in, The actual target value, the predicted value of the student network, and the predicted value of the teacher network are respectively used. and These are the student network target loss and the distillation target loss, respectively, λ D This is the balance coefficient;

[0026] Formula (9) is used as the category loss function.

[0027]

[0028] in, These represent the true class probability, the class probability predicted by the student network, and the class probability predicted by the teacher network, respectively. This represents the target predicted value of the teacher network. Classify the losses of students' online activities. For distillation classification loss, λ D This is the balance coefficient;

[0029] Formula (10) is used as the bounding box loss function.

[0030]

[0031] in, These represent the ground truth bounding boxes, the bounding boxes predicted by the student network, and the bounding boxes predicted by the teacher network, respectively. This represents the target predicted value of the teacher network. For student network bounding box loss, For the distillation bounding box loss, λ D This is the balance coefficient;

[0032] Ultimately, the student's loss L student and distillation loss L distill We then perform a weighted sum to obtain the total loss.

[0033] On the other hand, the present invention provides a lightweight device for multi-type target detection in SAR images that combines channel pruning and knowledge distillation, comprising:

[0034] The channel pruning module is used to sparsely train complex models on SAR image datasets, determine the importance of each channel in the complex model, and prune secondary channels by channel pruning. After fine-tuning the training, the pruned and optimized complex model is obtained.

[0035] The knowledge distillation module is used to construct a knowledge distillation framework by using the pruned and optimized complex model as the teacher network and the lightweight model as the student network. The framework is trained on a SAR image dataset to continuously update the parameters of the lightweight model and obtain a lightweight SAR target detection weight model.

[0036] The SAR target detection module is used to detect targets in SAR images by using a lightweight SAR target detection weight model as the SAR target detection network.

[0037] The beneficial effects of this invention are:

[0038] (1) This invention can effectively avoid the performance collapse of complex SAR image target models due to excessive pruning during the lightweighting process using single-channel pruning; it can also alleviate the problem of limited knowledge distillation effect caused by the large size difference between the teacher model and the student model. Compared with the method of first constructing a lightweight SAR image target detection model using channel pruning and then using the source network model to perform knowledge distillation, this invention optimizes the complex SAR image target model by pruning, removes secondary feature information in the complex model, and uses the pruned and optimized complex model as the teacher network, so that it can more effectively guide the student network learning and training, and the transfer of feature knowledge in SAR images is more targeted.

[0039] (2) Fine-tuning the pruned complex model is intended to reduce the impact of feature channel loss on target detection and restore the detection accuracy of the pruned model.

[0040] (3) Experiments show that the present invention achieves high-precision detection while significantly surpassing YOLOv3 and YOLOv7 in terms of model volume and detection time. Attached Figure Description

[0041] Figure 1 A schematic diagram illustrating a lightweight method for multi-type target detection in SAR images that combines channel pruning and knowledge distillation, as provided in an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of channel pruning provided in an embodiment of the present invention;

[0043] Figure 3 The knowledge distillation structure provided for embodiments of the present invention;

[0044] Figure 4 The information expression of labels with different degrees of softening provided in the embodiments of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0046] Example 1

[0047] like Figure 1 As shown, this embodiment of the invention provides a lightweight method for multi-type target detection in SAR images that combines channel pruning and knowledge distillation, including the following steps:

[0048] S101: The complex model is trained on a SAR image dataset with sparsity, the importance of each channel in the complex model is determined, and secondary channels are pruned. After fine-tuning the training, the pruned and optimized complex model is obtained.

[0049] S102: The pruned and optimized complex model is used as the teacher network and the lightweight model is used as the student network to construct a knowledge distillation framework. The model is trained on a SAR image dataset to continuously update the parameters of the lightweight model and obtain a lightweight SAR target detection weight model.

[0050] S103: Use a lightweight SAR target detection weight model as the SAR target detection network to detect targets in the SAR image to be detected.

[0051] The detection method provided in this invention can effectively avoid the performance collapse of the model due to over-pruning during the lightweighting process using single-channel pruning; it can also alleviate the problem of limited knowledge distillation effect caused by the large size difference between the teacher model and the student model. Compared with the method of first constructing a lightweight model using channel pruning and then using the source network model to perform knowledge distillation, this invention optimizes the teacher network by pruning, removing secondary feature information from the teacher network, making it more effective in guiding the student network's learning and training, and making the transfer of target feature knowledge in SAR images more targeted.

[0052] Example 2

[0053] Based on the above embodiments, this embodiment of the invention conducts a lightweight detection experiment on the reconstructed SSDD ship and MSAR oil tank and bridge datasets using the YOLOv5 target detection algorithm.

[0054] Experimental environment: Windows 10 operating system and PyTorch deep learning framework; computer configuration: Intel Xeon Gold 6230, 2.10GHz; 128GB RAM; two NVIDIA GeForce RTX 2080Ti graphics cards, 22GB VRAM each.

[0055] Dataset: The experimental dataset was constructed by combining oil tank and bridge targets from the MSAR dataset and ship targets from the SSDD dataset. The MSAR dataset contains 1250 images of oil tank targets and 1582 images of bridge targets, while the SSDD dataset contains 1160 images of ship targets. All target types were partitioned into training and validation sets in an 8:2 ratio.

[0056] Experimental parameter settings: epochs: 300; initial learning rate: 0.01; IoU threshold: 0.5; batch size: 16; input image size: 640×640; SGD optimizer was used. During channel pruning and knowledge distillation, the sparsity training parameter sr was set to 5×10⁻⁶. -4 The pruning ratios were set to 0.1–0.69; the knowledge distillation temperature parameter τ was 20; the knowledge distillation loss function was L2 distillation loss; and the balance factor α was 0.5.

[0057] Based on the above experimental setup, the detection process in this embodiment specifically includes the following steps:

[0058] S201: Sparsification training is performed on the training set for complex models based on convolutional neural networks. The purpose is to determine the importance of each channel in the convolutional layer and provide a reference index for channel pruning.

[0059] Specifically, let the output of the previous layer be x1, x2, ..., x m The normalized output is obtained according to formula (1).

[0060]

[0061] Where m is the number of training samples per batch, μ B and σ B Let be the mean and variance of each batch of data, and ε be the regularization parameter;

[0062] To improve the model's nonlinear feature extraction capability and overall expressive power, the data distribution is reconstructed using a learnable scaling factor γ and a translation factor β according to formula (2), resulting in the reconstructed result y. i :

[0063]

[0064] The above process describes the feature map processing of adjacent convolutional and batch normalization (BN) layers before pruning. Figure 2 As shown on the left.

[0065] In this embodiment, to avoid introducing additional computational overhead, the scaling factor γ in the BN layer is sparsified to determine the importance of the corresponding channel.

[0066] As one possible implementation, L1 regularization is used to sparsify the set Γ of γ values ​​corresponding to each channel in the convolutional layer. After sparsification, the feature channels with γ values ​​approaching 0 are the secondary channels. The regularization formula of Γ is then introduced into the network loss function:

[0067]

[0068] Where l(f(x,W),y) is the loss function of the original convolutional neural network, (x,y) are the input and target of training, W is the trainable weight, λ is the balance factor, and g(γ) is the sparse penalty term for the scaling factor, i.e., g(γ)=||γ||1.

[0069] Taking the partial derivative of equation (3) with respect to γ, we can see that the partial derivative is discontinuous at γ = 0. If the left and right partial derivatives have opposite signs, then the loss function L has a minimum point at this point. During the optimization process of subgradient descent, some γ will continuously move closer to 0, and the set Γ will become sparse.

[0070]

[0071] S202: Channel pruning is performed on complex models to optimize the feature knowledge carried by the complex models and reduce the size of the teacher network model. A diagram of channel pruning is shown below. Figure 2 As shown.

[0072] Specifically, based on the scaling factor γ after sparsification training, it is connected to the corresponding channel in the feature map of the convolutional layer, and the absolute values ​​are sorted. The pruning threshold S is determined by setting the pruning ratio. If the γ corresponding to the feature channel is less than the pruning threshold S, it is determined as a secondary channel and pruning is applied.

[0073] It should be noted that, to ensure the integrity of the network structure, the pruning threshold S should not exceed the largest γ value in any BN layer, to ensure matching with the dimensionality of the backbone network. Furthermore, structures with residual connections are not pruned to ensure that the dimensionality of the feature maps of diameter connections and residual layers is consistent.

[0074] The process of pruning branches in the passageway is as follows Figure 2 As shown. For example, in the feature map of the i-th convolutional layer of the original network, let each channel be connected to a sparsified scaling factor γ, and the feature channel C in the figure... i2 and C i4 The corresponding γ value is less than the threshold, and its contribution to the overall network performance is relatively small. Therefore, the corresponding previous layer convolutional kernel and feature channels are pruned to achieve the goal of compressing the model volume.

[0075] S203: Fine-tune the training of the pruned complex model. The purpose is to reduce the impact of feature channel loss on target detection and restore the detection accuracy of the pruned model.

[0076] Specifically, the teacher model after channel pruning is used as a pre-trained model. Using the same parameters, it is fine-tuned again in fewer training rounds to obtain the pruned and optimized complex model.

[0077] S204: Using the pruned and optimized complex model as the teacher network and the lightweight model as the student network, a knowledge distillation framework is constructed and trained on the training set to continuously update the parameters of the lightweight model, resulting in a lightweight SAR target detection weight model; specifically including the following sub-steps:

[0078] S2041: The constructed knowledge distillation framework, as follows Figure 3 As shown. The softmax output of the student network is used as the hard prediction, and the warmed prediction result of the softmax output of the student network is used as the soft prediction; the warmed prediction result of the softmax output of the teacher network is used as the soft label.

[0079] The output of Softmax without heating is shown in formula (5). The output of Softmax after heating is shown in formula (6).

[0080]

[0081]

[0082] To transfer the generalization ability of the teacher network to the student network in knowledge distillation, the network needs to comprehensively learn the features of positive and negative samples. The Softmax function in the network can map the output of neurons to the (0, 1) interval, which is equivalent to outputting the probability of the predicted class, as shown in Equation (5). However, when the probability distribution of the Softmax output is small, the value of the negative label will approach 0, and its contribution to the calculation of the loss function will be weakened. Therefore, a temperature variable τ is added to the Softmax function to amplify the information carried by the negative label, as shown in Equation (6), so that the student network can better learn the inter-class differences of different targets and improve the detection accuracy. Wherein, y i y is the output of the previous neuron. i ' represents the output of the Softmax function, and τ is the temperature variable.

[0083] S2042: Constructing a student loss L between hard labels and hard predictions student Constructing a distillation loss L between soft prediction and soft labeling distill Thus, the total loss L = αL is constructed. distill +(1-α)L student This improves the detection accuracy and robustness of student networks in responding to targets.

[0084] Specifically, in the original network, the loss function is composed of the target loss f obj Classification loss f cl and bounding box loss f bb The model consists of three parts, as shown in equation (7). Through continuous iterative training, the loss value gradually decreases and converges, and the model prediction value continuously approaches the true value.

[0085]

[0086] in, For the target, class probability, and bounding box predicted by the model, This corresponds to the actual value.

[0087] Based on this, in constructing the student network loss and distillation loss, the embodiments of the present invention also consist of target loss, classification loss and bounding box loss, respectively.

[0088] Since target loss is the main basis for judging background and target, and a prerequisite for further learning target type and location regression, the target loss function needs to be modified first. The softened teacher network is used to predict and use the real labels, while simultaneously guiding the student network to learn target features. The modified target loss function is shown in equation (8):

[0089]

[0090] in, The modified target loss function, The actual target value, the predicted value of the student network, and the predicted value of the teacher network are respectively used. and These are the student network target loss and the distillation target loss, respectively, λ D This is the balance coefficient.

[0091] Because YOLOv5 predicts both the target type and the bounding box, standard knowledge distillation would transfer the teacher network's prediction of the background box to the student network, thus affecting the student network's prediction of the target box. Therefore, to ensure that the student network only learns the class probability and bounding box information when the teacher network's target prediction is high, a factor related to the updated teacher network's target prediction is introduced into the distillation loss of the classification and bounding box loss functions. The product of . If the object in the prediction box is the background, The value of will be very low, causing the distillation loss to approach 0, which can effectively prevent students from learning background information incorrectly online. The modified category loss function is shown in equation (9):

[0092]

[0093] in, For the modified class loss function, These represent the true class probability, the class probability predicted by the student network, and the class probability predicted by the teacher network, respectively. This represents the target predicted value of the teacher network. Classify the losses of students' online activities. For distillation classification loss, λ D This is the balance coefficient.

[0094] Similarly, the bounding box loss function has a similar expression, as shown in equation (10):

[0095]

[0096] in, For the modified bounding box loss, These represent the ground truth bounding boxes, the bounding boxes predicted by the student network, and the bounding boxes predicted by the teacher network, respectively. This represents the target predicted value of the teacher network. For student network bounding box loss, For the distillation bounding box loss, λ D This is the balance coefficient.

[0097] In summary, the total distillation loss used to update model parameters in the network is the sum of equations (8)-(10), as shown in equation (11).

[0098]

[0099] S2043: Based on the total loss L, the parameters of the lightweight model are continuously updated to obtain a lightweight SAR target detection weight model.

[0100] Since channel pruning reduces the number of model parameters and complexity, it also lowers the probability of overfitting. Therefore, channel pruning can be considered a form of regularization. A channel-pruned teacher network can provide a smoother "soft label" that is more easily noticed and learned by the student network, mitigating overfitting and noise caused by an overly complex teacher network. Figure 4 For labels with different levels of softening, hard labels only indicate whether the target is a ship or not, providing too little information. Softened labels can guide the student network to some extent regarding the likelihood that the target is a ship. Pruned soft labels offer a smoother information distribution, further amplifying the differences between latent classes of the target. During knowledge distillation, the student network can learn more "hidden knowledge," improving distillation efficiency and making the model more adaptable to new data. The information representation of labels with different levels of softening is as follows: Figure 4 As shown.

[0101] S205: With the correct environment and variables configured, input the SAR image to be detected, and use the lightweight SAR target detection weight model and forward propagation algorithm to regress and obtain the target type and location information in the image to be detected, ultimately achieving high-precision lightweight SAR target detection and localization.

[0102] Furthermore, to evaluate the performance of the method of this invention in terms of detection accuracy, model size, and detection efficiency, the optimal training model of this invention (30% pruning ratio + knowledge distillation) was compared with the training results of current mainstream deep learning object detection algorithms. During the experiment, the parameter configurations of each type of algorithm remained consistent with those described above. Table 1 shows the performance of different algorithms. The results show that the method of this invention has the best overall performance, with an average accuracy of 95%, which is 5.4%, 10.9%, and 2.8% higher than the lightweight object detection algorithms YOLOv7-tiny, YOLOv3-tiny, and YOLOv5-n, respectively. While achieving high-accuracy detection, it also significantly surpasses YOLOv3 and YOLOv7 in terms of model size and inference time.

[0103] Table 1. Performance Comparison Results of Different SAR Target Detection Algorithms

[0104]

[0105] Example 3

[0106] Corresponding to the above method, this embodiment of the invention provides a multi-type SAR image target detection device that combines channel pruning and knowledge distillation. The device includes a channel pruning module, a knowledge distillation module, and a SAR target detection module.

[0107] The channel pruning module is used to sparsely train the complex model on a SAR image dataset, determine the importance of each channel in the complex model, and prune secondary channels. After fine-tuning, the pruned and optimized complex model is obtained. The knowledge distillation module is used to construct a knowledge distillation framework by using the pruned and optimized complex model as the teacher network and the lightweight model as the student network. It is then trained on the SAR image dataset to continuously update the parameters of the lightweight model, resulting in a lightweight SAR target detection weight model. The SAR target detection module is used to use the lightweight SAR target detection weight model as a SAR target detection network to detect targets in the SAR images to be detected.

[0108] It should be noted that the detection device provided in the embodiments of the present invention is for implementing the above method embodiments, and its specific functions can be described in the above method embodiments, which will not be repeated here.

[0109] This invention utilizes channel pruning technology to prune and optimize complex network models, and uses it as a teacher network to guide the training of lightweight student networks within a knowledge distillation framework. This improves the target detection accuracy of the lightweight network model, ultimately achieving high-precision, lightweight SAR target detection and localization.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for lightweight multi-type target detection of SAR images by combining channel pruning and knowledge distillation, characterized in that, Comprise: Step 1: the complex model is trained on the SAR image data set, the importance of each channel in the complex model is determined, and the channel pruning is used to cut the secondary channel, and the pruned and optimized complex model is obtained through fine-tuning training; Step 2: the pruned and optimized complex model is used as the teacher network, the lightweight model is used as the student network, the knowledge distillation framework is constructed, the lightweight model parameters are updated on the SAR image data set, and the lightweight SAR target detection weight model is obtained; specifically comprising: The Softmax output of the student network is used as the hard prediction, and the prediction result of the Softmax output of the student network after temperature rising is used as the soft prediction; the prediction result of the Softmax output of the teacher network after temperature rising is used as the soft label; Building student loss between hard label and hard prediction Building distillation loss between soft prediction and soft label Thus, the total loss is built ; based on total loss The lightweight model parameters are constantly updated to obtain a lightweight SAR target detection weight model. where student loss and distillation loss The construction process includes: Equation (8) is used as the target loss function : （8） wherein, , , respectively the target real value, the prediction value of the student network and the prediction value of the teacher network, and respectively the student network target loss and the distillation target loss, is a balancing coefficient; Equation (9) is used as the class loss function : （9） wherein, , , respectively represent the category probability true value, the category probability predicted by the student network, and the category probability predicted by the teacher network, represents the target prediction value of the teacher network, is the student network classification loss, is the distillation classification loss, is a balance coefficient; Equation (10) is used as the bounding box loss function : （10） wherein, , , respectively represent the real bounding box, the bounding box predicted by the student network, the bounding box predicted by the teacher network, represents the target prediction value of the teacher network, is the bounding box loss of the student network, is the distillation bounding box loss, is the balance coefficient; Finally, the student loss and distillation loss are weighted and summed to get the total loss; Step 3: using the lightweight SAR target detection weight model as the SAR target detection network, the target in the SAR image to be detected is detected.

2. The channel pruning and knowledge distillation combined SAR image multi-type target detection lightweight method according to claim 1, characterized in that, In step 1, the complex model is trained on the SAR image data set, specifically comprising: Adopting L1 regularization method, the set of scaling factor values corresponding to each channel in the convolution layer is sparsely expressed and the regularization formula of the set is introduced into the network loss function to obtain a modified network loss function:​ （3） in, For the loss function of complex models, For the input and target of training, For trainable weights, As a balance factor, scaling factor The sparse penalty term, ; Taking the partial derivative of equation (3) with respect to , the partial scaling factor is constantly approaching 0 in the process of sub-gradient descent optimization, and the sparse set is obtained.

3. The channel pruning and knowledge distillation combined SAR image multi-type target detection lightweight method according to claim 2, characterized in that, In step 1, the importance of each channel in the complex model is determined, and the channel pruning is used to cut the secondary channel, specifically comprising: Scaling factor based on sparsification training , and the absolute value is sorted, and the pruning threshold is determined by setting the pruning ratio , if the channel corresponding to is less than the pruning threshold , it is determined as a secondary channel and pruning is applied.

4. The channel pruning and knowledge distillation combined SAR image multi-type target detection lightweight device, applying the channel pruning and knowledge distillation combined SAR image multi-type target detection lightweight method of claim 1, characterized in that, Comprise: The channel pruning module is used for training the complex model on the SAR image data set, determining the importance of each channel in the complex model, and cutting the secondary channel by using channel pruning, and the pruned and optimized complex model is obtained through fine-tuning training; The knowledge distillation module is used for using the pruned and optimized complex model as the teacher network, using the lightweight model as the student network, constructing the knowledge distillation framework, training on the SAR image data set, updating the lightweight model parameters, and obtaining the lightweight SAR target detection weight model; The SAR target detection module is used for using the lightweight SAR target detection weight model as the SAR target detection network, and detecting the target in the SAR image to be detected.

Images