A lightweight sensitive target recognition method based on network pruning and transfer learning

Abstract

The application discloses a kind of light sensitive target identification method based on network pruning and transfer learning, comprising: step 1: relevant battlefield actual image set is collected, and target image set is obtained by using cyclic counter network for cross-domain amplification;Step 2: the target image set is used as input to train YOLO-F, until the loss function of YOLO-F converges, and the trained YOLO-F is obtained, and the trained YOLO-F is used to realize target identification;The application ensures that when battlefield target detection is carried out through embedded device in related technology, the model size and operation times are greatly reduced under the condition that the accuracy tends to be stable interval, and it is suitable for light weight platform and identification scene with more demanding operation requirements.

Description

Technical Field

[0001] This invention belongs to the fields of image processing and artificial intelligence, and in particular relates to a lightweight sensitive target recognition method based on network pruning and transfer learning. Background Technology

[0002] Smart mines are a novel concept in munitions that have been widely discussed in recent years. They detect targets using acoustic, vibration, and infrared sensors, and then autonomously and intelligently destroy them. With the rapid development of next-generation information technology, various miniaturized, low-power composite sensors are being used in smart mines, enabling them not only to detect ground targets but also to autonomously identify and track near-ground aerial targets. Smart mine systems overcome the passive detonation of traditional landmines, actively detecting, tracking, and selectively destroying high-value targets. In particular, when combined with network communication to form a smart minefield, multiple smart mine nodes can autonomously network, sharing target information, coordinating attack timing, reporting battlefield situational information, and linking with the combat command and control system to maximize destructive effectiveness. This makes them unmanned, networked, intelligent area denial weapons.

[0003] For intelligent minefield weapon systems, their active and passive detection capabilities and network coordination are crucial to maximizing their combat effectiveness. Currently, most active intelligent minefield weapon systems in various countries primarily utilize acoustic or acoustic-vibration composite sensors for early warning and detection, while infrared sensors are mostly used for secondary homing. However, target identification and localization methods based on sound or vibration cannot meet the demands of the complex and ever-changing modern battlefield environment: vibration signals caused by moving targets on the battlefield are coupled with geological features and various environmental sounds, making extraction difficult; infrared and millimeter-wave radar detectors are significantly affected by the atmosphere and rain, and consume excessive power, making it difficult to meet the operational requirements of long-term unattended operation in intelligent minefields; the sound and vibration signal characteristics of different types of targets vary significantly with the number of targets and their payloads; in peacetime, the sound, infrared, and vibration characteristics of different ground and air combat units are difficult to obtain in advance due to military secrecy; furthermore, the use of various stealth and shielding materials allows the characteristics of hostile targets to be rapidly altered, further increasing the difficulty of target identification and tracking.

[0004] In the field of mobile computing, some new methods and technologies have provided solutions to the aforementioned problems. On the one hand, with the improvement of the computing power of system-on-a-chip (SoC), edge devices have gradually acquired the ability to perform some simple feature extraction and target recognition, avoiding the adverse effects on recognition results when there are deviations in feature extraction and feature selection in traditional pattern recognition methods. On the other hand, thanks to the maturity of methods such as transfer learning and deep learning, the recognition accuracy and speed of edge devices will continue to improve with the continuous improvement and supplementation of battlefield data, facilitating the "flexible" upgrade of weapon and equipment performance. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight sensitive target recognition method based on network pruning and transfer learning, in order to solve the problems that existing traditional detection network models are large in size and difficult to apply to edge devices with limited computing power; while lightweight detection networks are limited by the design of the network model, resulting in poor detection capabilities and accuracy.

[0006] This invention adopts the following technical solution: a lightweight sensitive target recognition method based on network pruning and transfer learning, comprising:

[0007] Step 1: Collect relevant battlefield image sets and use a recurrent adversarial network to perform cross-domain augmentation to obtain the target image set;

[0008] Step 2: Use the target image set as input to train YOLO-F until the loss function of YOLO-F converges and obtain the trained YOLO-F. Use the trained YOLO-F to achieve target recognition.

[0009] YOLO-F consists of: an input terminal, a backbone network, a feature fusion network, and an output terminal;

[0010] The YOLO-F feature fusion network takes as input the low-level, mid-level, and high-level features output from the input terminal and obtains three fused features at three detection depths. During fusion, two RepVGG modules are used to enhance the network's feature extraction capability.

[0011] Further, step 2 includes:

[0012] Step 201: Input the target image set into the input terminal of YOLO-F for preprocessing;

[0013] Step 202: Input the preprocessed image into the YOLO-F backbone network to extract low-level features, mid-level features, and high-level features;

[0014] Step 203: Input the low-level features, mid-level features, and high-level features into the YOLO-F feature fusion network to obtain three fused features at the three detection depths;

[0015] Step 204: Input the fused features into the output of YOLO-F, perform result prediction and filtering, and obtain the final prediction result to achieve target recognition.

[0016] Furthermore, in step 203, three fusion features are obtained at the three detection depths as follows:

[0017] The fusion features obtained at the first detection depth include: high-level features are passed through the first CBS module to obtain the first feature, the first feature is upsampled and sent to the fourth connection layer, and the middle layer features are fused in the fourth connection layer and then sequentially enter the fourth C3 module and the second CBS module to obtain the second feature, the second feature is upsampled and sent to the first connection layer, and the bottom layer features are fused in the first connection layer and then passed through the first C3 module again to obtain the first fused feature, and then output to the output layer;

[0018] The fusion features obtained at the second detection depth include: the first fusion feature is sent to the second connection layer through the first RepVGG module, and the second feature is fused in the second connection layer and then sent to the second C3 module to obtain the second fusion feature, and then output to the output layer;

[0019] The fusion features obtained at the third detection depth include: the second fusion feature is sent to the third connection layer through the second RepVGG module, and after being fused with the first feature in the third connection layer, it is sent to the third C3 module to obtain the third fusion feature, and then output to the output layer.

[0020] Furthermore, in each C3 module and CBS module, a pruning algorithm is used to prune the filters during convolution, and the pruned network is used as the student network to learn the characteristics of the teacher network using knowledge distillation.

[0021] Furthermore, the activation functions of the first RepVGG module and the second RepVGG module in step 2 are both SiLU functions.

[0022] The beneficial effects of this invention are:

[0023] This invention significantly improves the feature extraction capability of lightweight networks by introducing the RepVGG reparameterization module. Specifically, the RepVGG module implements a multi-branch structure during the training phase, enhancing the model's feature extraction ability. During the inference phase, network branch merging greatly improves inference speed. The SiLU activation function is used instead of the traditional ReLU activation function, making the gradient smoother and more continuous within the input range, improving the model's training performance and mitigating the gradient vanishing problem. Experiments show that, compared with the initial model A, model B using the RepVGG module maintains essentially the same detection accuracy, decreasing by less than 1%, while improving inference speed by approximately 14%.

[0024] This invention employs a filter-based pruning algorithm to effectively reduce the model size. By identifying and removing filters that do not contribute significantly to model performance, the number of feature maps generated in layer i and convolutional kernels in layer i+1 is reduced, thereby achieving more efficient computation and storage. This pruning method has significant advantages on edge devices and mobile devices, enabling faster inference while reducing latency. Experiments show that model C generated by the pruning algorithm is approximately 44% smaller than model B, with a 42% increase in inference speed and only a 4% decrease in accuracy, achieving lightweight modeling without significantly reducing detection accuracy.

[0025] This invention also applies a feature-based knowledge distillation algorithm, which significantly improves the performance of the simplified model by transferring knowledge from the teacher model (YOLOv5x) to the student model. By introducing channel-level information transfer, performance is maintained as much as possible while reducing the model size. In specific implementation, the final pruned model is used as the student model, and the original YOLOv5x model with the highest accuracy is used as the teacher model. Features are output using layers 17, 20, and 23 of the Head part, ensuring the accuracy and performance of the pruned model. Experimental results show that after knowledge distillation, the final model D (YOLO-F) recovers to a level similar to the original model in terms of accuracy, while the inference speed and model size are only half that of the original model.

[0026] This invention ensures that when performing battlefield target detection through embedded devices in related technologies, the accuracy tends to be within a stable range, while greatly reducing the model size and the number of calculations. It is suitable for lightweight platforms and recognition scenarios with demanding computational requirements. Attached Figure Description

[0027] Figure 1 This is a diagram of the YOLO-F network model structure of the present invention;

[0028] Figure 2 These are the recognition diagrams for each model in this invention;

[0029] Figure 3 This is the detection thermogram of YOLO-F in this invention;

[0030] Figure 4 This is a comparison chart of the algorithm accuracy of the present invention and existing technologies. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0032] This invention improves the YOLOv5 network structure by incorporating a RepVGG reparameterization module into the feature fusion network, enhancing the network's feature extraction capabilities and improving the recognition accuracy and generalization ability of the recognition network. Through a filter-based pruning algorithm, filters that do not contribute significantly to model performance are identified and removed to reduce the model size, thereby achieving more efficient computation and storage. Finally, a feature-based knowledge distillation method is used to enable the pruned network, acting as a student network, to focus more on important regions in each channel—that is, the prominent regions for each category recognition in the teacher network—thus improving the network model's recognition accuracy in dense predictions.

[0033] Therefore, this invention discloses a lightweight sensitive target recognition method based on network pruning and transfer learning, comprising:

[0034] Step 1: Collect relevant battlefield images and use a recurrent adversarial network to perform cross-domain augmentation to obtain the target image set.

[0035] In step 1, when collecting relevant actual battlefield image sets, the actual battlefield image sets include different types of oracle bone script images, including tanks, armored vehicles, helicopters, and combat personnel. These images vary in size, shape, environmental conditions, and background conditions. Using a recurrent adversarial network (RON), cross-domain augmentation is performed on the actual battlefield image sets to obtain a target image set containing these actual images. The target image set is then manually labeled using LabelMe software, indicating the type, content, and location of different battlefield targets in each image. The dataset is divided into training, validation, and test sets in a 4:1:1 ratio. The training set is used for specific training of the network model, the validation set is used to adjust the training progress and various hyperparameters of the network model, and the test set is primarily used to evaluate the specific performance and generalization ability of the network model.

[0036] In step 1, a Recurrent Adversarial Network (CycleGAN) was introduced to augment the dataset across domains. CycleGAN is an asymmetric network for image transfer that can generate synthetic images with battlefield features by learning mappings from one domain to another. Because it does not require paired data during training, it effectively reduces the network's requirements for image data, enhancing its applicability. Simultaneously, as an image generation technique, CycleGAN can effectively augment data by increasing the size of the actual battlefield image set, thereby improving the generalization performance of the deep learning model.

[0037] The unique feature of Generative Adversarial Networks (GANs) lies in their ability to learn unsupervised image transformations. CycleGAN, through mechanisms such as cycle consistency loss and identity mapping loss, ensures that generated images maintain semantic and structural consistency. For battlefield recognition tasks, this means that the generated images not only resemble real battlefield scenes in appearance but also retain important semantic information, such as terrain, buildings, and other tactical features. Integrating CycleGAN into battlefield recognition systems can expand limited real-world datasets and improve the performance of deep learning models across various scenarios. This has significant practical implications when collecting large amounts of real-world samples on the battlefield may be impractical or costly, providing richer and more diverse training data for model performance.

[0038] CycleGAN consists of two generators and two discriminators, used to perform asymmetric transformation between the source and target images. Its loss function mainly consists of adversarial loss, cycle consistency loss, and identity loss. The specific process is represented by equations (1)-(5):

[0039] L GAN (G,D B (A,B)=E b～Pdata(b) [logD B (b)]+E a～Pdata(a) [log(1-D B (G(a)))](1)

[0040] L GAN (F,D A (B,A)=E a～Pdata(a) [logD A (a)]+E b～Pdata(b) [log(1-D A (F(b)))] (2)

[0041] L cyc (G,F)=E a～Pdata(a) [||F(G(a))-a||1]+E b～Pdata(b) [||G(F(b))-b||1] (3)

[0042] L ID (G,F)=E b～Pdata(b) [||G(b)-b||1]+E a～Pdata(a) [||F(a)-a||1] (4)

[0043] L = L GAN (G,D B ,A,B)+L GAN (F,D B ,B,A)+L cyc(G,F)+L ID (G,F) (5)

[0044] Where L represents the adversarial loss function, E represents the expected value, A represents the source domain, a represents the image in the source domain, B represents the target domain, b represents the image in the target domain, G represents the generator from the source domain to the target domain, F represents the generator from the target domain to the source domain, and D... A DB represents the discriminator for the source domain, and DB represents the discriminator for the target domain.

[0045] Adversarial loss forces the generator to produce realistic images that can fool the discriminator, cycle consistency loss prevents the generator from overfitting, and identity loss is used to ensure that the generated new images have a high similarity to the source domain images, thereby generating a batch of enhanced images that can greatly improve the generalization ability.

[0046] Step 2: Use the target image set as input to train YOLO-F until the loss function of YOLO-F converges and obtain the trained YOLO-F. Use the trained YOLO-F to achieve target recognition.

[0047] Wherein, YOLO-F, as Figure 1 As shown, it includes: an input terminal, a backbone network, a feature fusion network, and an output terminal;

[0048] The YOLO-F feature fusion network takes as input the low-level, mid-level, and high-level features output from the input terminal and obtains three fused features at three detection depths. During fusion, two RepVGG modules are used to enhance the network's feature extraction capability.

[0049] Step 2 includes:

[0050] Step 201: Input the target image set into the input terminal of YOLO-F for preprocessing;

[0051] Step 202: Input the preprocessed image into the YOLO-F backbone network to extract low-level features, mid-level features, and high-level features;

[0052] Step 203: Input the low-level features, mid-level features, and high-level features into the YOLO-F feature fusion network to obtain three fused features at the three detection depths;

[0053] Step 204: Input the fused features into the output of YOLO-F, perform result prediction and filtering, and obtain the final prediction result to achieve target recognition.

[0054] In step 203, three fusion features are obtained at the three detection depths as follows:

[0055] The fusion features obtained at the first detection depth include: the high-level features are passed through the first CBS module to obtain the first feature, the first feature is upsampled and sent to the fourth connection layer, and the middle layer features are fused in the fourth connection layer and then sequentially enter the fourth C3 module and the second CBS module to obtain the second feature, the second feature is upsampled and sent to the first connection layer, and the bottom layer features are fused in the first connection layer and then passed through the first C3 module again to obtain the first fused feature, and then output to the output layer;

[0056] The fusion features obtained at the second detection depth include: the first fusion feature is sent to the second connection layer through the first RepVGG module, and after the second feature is fused in the second connection layer, it is sent to the second C3 module to obtain the second fusion feature, and then output to the output layer;

[0057] The fusion features obtained at the third detection depth include: the second fusion feature is sent to the third connection layer through the second RepVGG module, and after being fused with the first feature in the third connection layer, it is sent to the third C3 module to obtain the third fusion feature, and then output to the output layer.

[0058] The activation functions for both the first and second RepVGG modules are SiLU functions.

[0059] In each of the C3 and CBS modules, a pruning algorithm is used to prune the filters during convolution, and the pruned network is used as the student network to learn the characteristics of the teacher network using knowledge distillation.

[0060] To improve the recognition accuracy and generalization ability of lightweight networks, this invention enhances the feature extraction capability of the network by integrating the RepVGG reparameterization module. This module decouples the model training and inference stages through structural reparameterization, improving model accuracy while greatly increasing the inference speed of the network. It enables multi-branch training of the network during the training stage to strengthen the model's feature extraction capability, and merges network branches during the inference stage to accelerate the model's inference speed.

[0061] Unlike previous network structures, the residual blocks in RepVGG are replaced by a single convolutional layer, which includes two convolutional operations: a batch normalization (BN) layer and a rectified linear unit (ReLU) activation function. This transformation from multiple computational models to a single convolutional layer reduces the number of network parameters without increasing the model size, making it easier to build lightweight networks.

[0062] The RepVGG architecture transitions from the training to the inference phase in three steps: First, the convolutional layers and BN layers are merged to obtain a 3×3 convolution, a 1×1 convolution, an Identity branch (used only when the output and input channels are the same), and three bias vectors; second, the three bias vectors are added together to obtain the final bias vector; finally, the two 1×1 convolution kernels are padded with zeros to become two 3×3 convolution kernels, resulting in three 3×3 convolution kernels. Then, the three 3×3 convolution kernels are added together using the additivity of convolution to obtain the final single 3×3 convolution.

[0063] Considering that the SiLU function exhibits smoother characteristics compared to the traditional ReLU activation function, and its derivative is continuous throughout the entire input range, it helps to propagate gradients more effectively, thereby improving the training effect of the model and alleviating the gradient vanishing problem. Therefore, this invention replaces the ReLU activation function in RepVGG with the SiLU function.

[0064] To reduce some floating-point calculations in redundant neural networks and thus improve network efficiency, a filter-based pruning algorithm is used for coarse-grained pruning. This achieves high compression ratios and reduces computation time during the testing phase without relying on any sparse convolution computation libraries or dedicated hardware.

[0065] By using the L1 norm as a regularizer, the weights generated by each filter in the current convolutional layer are calculated. When the weight value is low, the filter generally produces a feature map with the least impact on the output compared to other filters, i.e., the feature map with the smallest relative weight, as follows:

[0066]

[0067] In the formula, K l Here, ni represents the kernel weights in the filter, ni is the number of input channels in the i-th convolutional layer, and s is the L1 norm of the filter. j It affects how much effective information is retained after filtering the input information. The larger the value, the more effective information is retained after filtering, and the more important it is to the network prediction result; the smaller the value, the smaller the impact on the network prediction result, and the less important it is.

[0068] Furthermore, by subtracting the filter of layer i, and then subtracting part of the feature map generated by layer i and part of the convolution kernel of layer i+1, the total number of computations that can be reduced is shown in the following formula:

[0069] n i ×k 2 ×h i+2 ×w i+2

[0070] In the formula, k represents the convolution kernel size, h and w are the height and width of the input feature map, and i is the specific number of filter layers to be subtracted. This pruning method reduces the model size by identifying and removing filters that do not contribute significantly to model performance, thereby achieving more efficient computation and storage. This helps to achieve faster inference on edge devices, mobile devices, or applications requiring low latency, as well as a better balance between model performance and efficiency and adaptability to various application and hardware limitations.

[0071] Feature-based knowledge distillation differs from traditional transfer learning methods. In knowledge distillation, knowledge is typically transferred from the same dataset, and the two networks can be either homogeneous or heterogeneous. In contrast, transfer learning often involves transferring knowledge from datasets with different target audiences, and typically utilizes data knowledge from other domains on a single network. Specifically, this includes the following:

[0072] The pruned network model in this invention is selected as the student network, and the network model YOLOv5x with the highest accuracy during training on the same dataset is selected as the teacher network for knowledge distillation learning.

[0073] To better extract knowledge from each channel of the network output, before calculating the differences between the student and teacher networks, the output activation values ​​are first converted into a probability distribution map at the channel dimension, and then the differences between the teacher and student networks are calculated. The channel-dimensional distillation loss can be expressed in a general formula as follows:

[0074]

[0075] In the formula, T represents the teacher network, S represents the student network, and y T y S These represent the feature maps generated by the two networks, respectively. φ is used to evaluate the differences in channel distribution between the teacher network and the student network, while φ is used to convert activation values ​​into a probability distribution. The specific representation is shown below:

[0076]

[0077] Where c is the number of input channels, c = 1, 2, ...; i represents the spatial index of a feature map channel; W and H are the length and width of the input image, respectively; and τ represents the temperature hyperparameter. The influence of different network scales and complexities is eliminated through a normalized exponential function (Softmax). The larger the Softmax function, the smoother the output probability distribution, meaning each channel focuses on a larger spatial region. When the number of channels in the student network and the teacher network is mismatched, a 1×1 convolution is used to adjust the number of channels in the student network, ensuring that the number of output channels is consistent, allowing the student network to learn the features from the teacher network normally.

[0078] Using information (KL) divergence as the loss function for knowledge distillation, as shown below:

[0079]

[0080] The core idea of ​​this method is to reduce the complexity between models by extracting knowledge from channels in one model and passing it to the corresponding channels in another model. This helps retain the feature learning capabilities of the original model in a streamlined model, thus maintaining performance as much as possible while reducing model size. Overall, feature-based knowledge distillation methods aim to transfer channel-level information between models to achieve the goal of reducing model size while maintaining performance. This approach can introduce an additional loss term during training to ensure knowledge transfer from the teacher model to the student model.

[0081] Based on the above steps, considering that the Head part of YOLOv5 contains three detection layers of different scales for predicting targets of different sizes, this invention uses layers 17, 20, and 23 in the network model (i.e., the Head part in the YOLO model) as feature output layers, thereby enabling the student network to better learn the knowledge features of the teacher network, so as to ensure that the pruned model can effectively maintain the accuracy and performance of the unpruned model.

[0082] Example 1

[0083] During training, specific environment variables were configured, the training cycle was set to 100 times, the batch training sample size was set to 16, the image input size was set to 640*640, the learning rate adjustment algorithm was set to cosine learning rate, and the initial learning rate was set to 0.01.

[0084] The model was trained using YOLO-F; the final results were evaluated using mean precision, parameter count, and F1 score as metrics. Mean precision (AP) is the integral of the precision (P) curve (vertical axis) and recall (R) curve (horizontal axis), calculated as follows:

[0085]

[0086]

[0087]

[0088] In the formula, TP represents a true instance, meaning both the actual and predicted cases are positive; FN represents a false negative, meaning the actual case is a positive case and the predicted case is a false case; FP represents a false positive, meaning the actual case is a negative case and the predicted case is a positive case. Since there is more than one type of military identification target in this invention, the mean AP (mean accuracy) should be calculated for each target, as shown in the following formula:

[0089]

[0090]

[0091] On the other hand, the lower the value of the parameter, the fewer the parameters there are, and the better the performance indicators are. The calculation formula is as follows:

[0092] Params=I×f×f×C+C

[0093] Where I is the input size, f is the kernel size, C is the output size, K is the number of target classes detected, and AP i This is the average accuracy for detecting the i-th target.

[0094] This invention compares the YOLO-F algorithm with other one-stage detection algorithms such as YOLOv5 and YOLOv7, along with their different model algorithms, on a near-ground battlefield dataset. The training results are as follows: Figure 4 As shown.

[0095] In addition, this invention designed four sets of ablation experiments based on a self-made battlefield target recognition dataset, such as... Figure 2 As shown, this is to further verify the impact of different modules and pruning algorithms on the performance of the model's target detection.

[0096]

[0097]

[0098] The ablation experiments are shown in the table above. Model A represents the YOLOv5s network model; Model B represents YOLOv5s with the 18th and 21st convolutional layers replaced by RepVGG modules, and the ReLU function in the traditional RepVGG module replaced by the SILU function; Model C represents YOLOv5s using RepVGG and pruning algorithms; Model D represents YOLO-F; the results are as follows:

[0099] Model A → Model B: By introducing the RepVGG convolution module, the average accuracy of the model fluctuated slightly from 0.865 to 0.86. The detection accuracy and model size decreased slightly, and due to the simplification of the inference network, the number of parameters decreased from 7.2G to 6.2G.

[0100] Model B → Model C: Pruning operation is performed, the model speedup ratio is 2, the average accuracy value is stabilized at 0.827, the model size is reduced to 7.0G, and the number of parameters is reduced from 6.2G to 3.6G, which greatly compresses the model.

[0101] Model C → Model D: Knowledge distillation was performed, using the YOLOv5x model, which is also from the YOLOv5 series and has the highest accuracy, as the teacher network, and the pruned model C as the student network. After 100 rounds of learning, the recognition accuracy of the network model was significantly improved while the size and number of parameters of the original model remained basically unchanged. This greatly met the various requirements of this paper for lightweight target recognition in unmanned systems.

[0102] Figure 3 The heatmap represents the detection performance of YOLO-F, where cool colors indicate lower algorithm attention and warm colors indicate higher attention. As can be seen from the graph, the algorithm proposed in this invention, in terms of feature fusion, focuses on detecting the core parts of the target while also expanding on external information to obtain more comprehensive feature information.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lightweight sensitive target recognition method based on network pruning and transfer learning, characterized in that, include: Step 1: Collect relevant battlefield image sets and use a recurrent adversarial network to perform cross-domain augmentation to obtain the target image set; Step 2: Use the target image set as input to train YOLO-F until the loss function of YOLO-F converges and obtain the trained YOLO-F. Use the trained YOLO-F to achieve target recognition. The YOLO-F includes: an input terminal, a backbone network, a feature fusion network, and an output terminal; The YOLO-F feature fusion network takes as input the low-level, mid-level, and high-level features output from the input terminal and obtains three fused features at three detection depths. During fusion, two RepVGG modules are used to enhance the network's feature extraction capability. Among them, three fusion features were obtained at the three detection depths as follows: The fusion features obtained at the first detection depth include: the high-level features are passed through the first CBS module to obtain the first feature, the first feature is upsampled and sent to the fourth connection layer, and the middle layer features are fused in the fourth connection layer and then sequentially enter the fourth C3 module and the second CBS module to obtain the second feature, the second feature is upsampled and sent to the first connection layer, and the bottom layer features are fused in the first connection layer and then passed through the first C3 module again to obtain the first fused feature, and then output to the output layer; The fusion features obtained at the second detection depth include: the first fusion feature is sent to the second connection layer through the first RepVGG module, and after the second feature is fused in the second connection layer, it is sent to the second C3 module to obtain the second fusion feature, and then output to the output layer; The fusion features obtained at the third detection depth include: the second fusion feature is sent to the third connection layer through the second RepVGG module, and after being fused with the first feature in the third connection layer, it is sent to the third C3 module to obtain the third fusion feature, and then output to the output layer.

2. The lightweight sensitive target recognition method based on network pruning and transfer learning according to claim 1, characterized in that, Step 2 includes: Step 201: Input the target image set into the input terminal of YOLO-F for preprocessing; Step 202: Input the preprocessed image into the YOLO-F backbone network to extract low-level features, mid-level features, and high-level features; Step 203: Input the low-level features, mid-level features, and high-level features into the YOLO-F feature fusion network to obtain three fused features at the three detection depths; Step 204: Input the fused features into the output of YOLO-F, perform result prediction and filtering, and obtain the final prediction result to achieve target recognition.

3. The lightweight sensitive target recognition method based on network pruning and transfer learning according to claim 2, characterized in that, In each of the C3 and CBS modules, a pruning algorithm is used to prune the filters during convolution, and the pruned network is used as the student network to learn the characteristics of the teacher network using knowledge distillation.

4. The lightweight sensitive target recognition method based on network pruning and transfer learning according to claim 1, characterized in that, The activation functions for both the first and second RepVGG modules in step 2 are SiLU functions.

Images