An unmanned aerial vehicle photoelectric pod small target detection model optimization method based on cooperative knowledge distillation
By optimizing the small target detection model of UAV optoelectronic pods using a collaborative knowledge distillation method, the problems of limited target information and difficulty in extracting small target information were solved, thereby improving the accuracy and speed of small target recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN AEROSPACE PRECISION ELECTROMECHANICAL INST
- Filing Date
- 2024-11-29
- Publication Date
- 2026-05-29
AI Technical Summary
Existing target recognition and detection algorithms face challenges in mobile applications such as limited target information and difficulty in extracting information from small targets when used in UAV electro-optical pods, resulting in low accuracy in small target recognition.
A collaborative knowledge distillation-based approach is adopted, in which a teacher model guides a student model to perform Gaussian receptive field modeling and bounding box assignment training. By combining local and global collaborative distillation networks, the model is trained on the foreground and background of the target, thereby optimizing the small target detection model.
Without changing the model structure, the recognition accuracy and detection speed of small targets are improved, thus enhancing the overall accuracy of target detection.
Smart Images
Figure CN122116189A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for small target recognition and detection, specifically to an optimization method for a small target detection model for UAV optoelectronic pods based on collaborative knowledge distillation. Background Technology
[0002] Currently, deep learning models have been successfully applied in multiple fields, particularly in target detection and recognition for UAV electro-optical pods. Thanks to the support of deep learning, target detection technology for UAV electro-optical pods has experienced rapid development. Traditional target recognition for UAV electro-optical pods combines image processing and machine learning algorithms with feature matching. Target recognition primarily relies on manually designed features; however, the theoretical analysis of such manually designed features is difficult, and the training methods require considerable experience and skill, necessitating experienced algorithm engineers. Deep learning, on the other hand, eliminates the need for manually designed features and offers better performance and robustness, thus holding broad application prospects in the field of machine vision.
[0003] Currently, most deep learning models for object recognition operate in edge scenarios. Unlike servers with abundant computing resources, mobile devices have low computing power and limited memory. Therefore, model compression methods from object recognition and detection algorithms are needed to compress and accelerate deep models trained on servers. With the rapid development of drones in recent years, more and more real-world scenarios require deployment on mobile devices such as drone pods. Traditional model compression methods face numerous difficulties and limitations when used in increasingly complex scenarios.
[0004] First, there is a lack of target information. This lack of target information has two aspects: firstly, small targets constitute a small percentage of the dataset; secondly, small targets are relatively small in size in the original image. In general object detection models, a basic backbone neural network (such as ResNet) is typically used for multiple downsampling processes, resulting in small targets appearing as only a few pixels in size on the feature map. This increases the difficulty for the object detection classifier to detect small targets, leading to a decline in its detection performance.
[0005] Secondly, extracting information from small targets is difficult. When using deep convolutional neural networks to extract features from images, conventional receptive fields only focus on the correspondence between a single high-level feature and the original pixels. As the number of convolutional network layers increases, the range of original image pixels reflected by a single feature becomes larger, and the microscopic information decreases accordingly. Ultimately, the feature information of small targets will be submerged in background noise. For example, a 10*10 pixel small target may only have a 1*1 feature after convolution, or even multiple 10*10 pixel small targets may merge into a single feature after multiple convolution kernels, making the small target unrecognizable.
[0006] Knowledge distillation, as a model compression method, mainly uses a pre-trained teacher model to guide the training of a student model, transferring the knowledge learned by the teacher model to the student model with relatively weaker learning ability. By utilizing the supervision information of the teacher model, better performance and accuracy can be achieved. Therefore, it is necessary to develop an optimization method for small target detection models based on existing knowledge distillation methods and apply it to UAV optoelectronic pods to overcome the above problems. Summary of the Invention
[0007] The purpose of this invention is to address the technical problem that existing target recognition and detection algorithms suffer from low accuracy in small target recognition due to insufficient target information and difficulty in extracting small target information when applied to mobile devices such as drones, by using model compression methods. The invention provides an optimization method for small target detection models of drone optoelectronic pods based on collaborative knowledge distillation.
[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0009] An optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation is characterized by the following steps:
[0010] S1 downloads the small object detection dataset and converts the labels of the small object detection dataset into COCO format. The small object detection dataset includes a training dataset for training and a validation dataset for evaluating performance.
[0011] S2 uses the training dataset of the small object detection dataset to pre-train the teacher model, trains the teacher model for at least 10 training cycles to obtain the trained teacher model, and saves the optimal weights during the training of the teacher model, which are then used as weights to guide the training of the student model.
[0012] S3 employs a collaborative knowledge distillation-based algorithm. It utilizes a pre-trained teacher model and its optimal weights to train the student model using a training dataset for Gaussian receptive field modeling, enabling the student model to learn the teacher model's bounding box assignment method. Simultaneously, it uses a collaborative knowledge distillation-based algorithm to train the foreground and background of the target. After training, a corresponding weight is generated.
[0013] S4 uses the student model trained in step S3 to test on the validation dataset, and uses the test results to score the weights in step S3.
[0014] S5 returns to step S3 and repeats at least 10 times to find the weight with the highest score as the optimal weight for training the student model. The corresponding student model is a detection model suitable for small targets, thus completing the optimization of the small target detection model.
[0015] Further, step S3 specifically includes:
[0016] S3.1 Gaussian modeling is performed on the effective receptive fields of the teacher model and the student model to obtain the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model.
[0017] S3.2 calculates the KL divergence between the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model. By minimizing the KL divergence, the student model learns the assignment method of the annotation box from the teacher model.
[0018] S3.3 simultaneously utilizes an algorithm based on collaborative knowledge distillation to train the foreground and background of the target;
[0019] After training is complete, S3.4 generates a corresponding weight.
[0020] Further, step S3.1 specifically includes:
[0021] S3.1.1 Based on the kernel size and stride of each convolutional layer in the student model, the effective receptive field r at each feature point in each convolutional layer is derived. n ;
[0022]
[0023] Where: k n and s i This represents the kernel size and stride of the nth convolutional operation;
[0024] S3.1.2 Using the location (x) of the feature points of each student model n ,y n μ is the average vector of a standard two-dimensional Gaussian distribution. e This will effectively allow you to experience the wild. n The square of Σ is used as the covariance matrix of the two-dimensional Gaussian distribution. e The standard two-dimensional Gaussian distribution N of the effective receptive field of the student model is obtained. e (μ e ,Σ e );
[0025]
[0026] S3.1.3 Using the center point (x) of the bounding box in the training dataset trained by each teacher model g ,y g μ is the average vector of a standard two-dimensional Gaussian distribution. g The covariance matrix ∑ is the square of the half-side length of the label box. gThe standard two-dimensional Gaussian distribution N of the teacher model is obtained. g (μ g ,∑ g );
[0027]
[0028]
[0029] Where: w g with h g These are the half-lengths of the annotation box;
[0030] Step S3.2 specifically involves calculating the KL divergence D between the standard two-dimensional Gaussian distribution of the effective receptive field of the student model and the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model. KL (N e ||N g By minimizing the KL divergence D KL (N e ||N g This allows the student model to learn the method of assigning annotation boxes from the teacher model;
[0031]
[0032] Further, step S3.3 specifically involves: calculating the loss L of local and global synergistic distillation. fea It is used to evaluate the consistency between the feature points extracted by the student model and the feature points extracted by the teacher model, and to train the foreground and background of the target.
[0033]
[0034] Where: M is the mask for extracting the foreground, A S and A C F represents the spatial attention mask and channel attention mask of the teacher detector built into the teacher model, respectively. T and F S Let F represent the feature maps of the teacher detector and the student detector built into the student model, respectively. f is the feature transformation that makes the feature map F of the student detector... S Feature map F of the teacher-specific detector T In the same dimension, C is the number of channels in the teacher model and the student model, H and W are the height and width of the feature maps of the teacher model and the student model, respectively, α and β are hyperparameters that balance the foreground and background losses, i,j are the pixel coordinates, k is the Kth channel, S is the student model, and T is the teacher model.
[0035] Furthermore, in step S2, the teacher model adopts ResNet-101;
[0036] In step S3, the student model uses ResNet-18 or ResNet-50.
[0037] Furthermore, in step S2, the teacher model is trained for 12 training cycles.
[0038] Further, in step S5, return to step S3 and repeat 12 times.
[0039] Compared with the prior art, the beneficial effects of the present invention are:
[0040] This invention provides an optimization method for small target detection models of UAV optoelectronic pods based on collaborative knowledge distillation. First, the effective receptive fields of the teacher and student models are Gaussian-modeled. Then, by minimizing the KL divergence, the student model learns the bounding box assignment method from the teacher model. Next, a bi-branch distillation network is established to distill and train on the target foreground and background separately. After several rounds of training, the student model with the highest weight score on the validation dataset is the detection model suitable for small targets. Adding Gaussian modeling allows the student model to learn the bounding box assignment strategy of the teacher model more rationally, improving the problem that previous bounding box assignments often ignored small targets. The training on the target foreground and background balances the foreground and background, with local distillation focusing more on the detection of small targets and global distillation focusing more on medium- and large-scale targets. The synergy of both improves the target detection accuracy. This method, without changing the model structure, can not only improve the target detection speed but also improve the recognition accuracy of small targets. Attached Figure Description
[0041] Figure 1 This is a bar chart showing the number of each category in the small target detection dataset used in step S1 of an embodiment of the optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation of the present invention.
[0042] Figure 2 This is an example diagram of the small target detection dataset used in an embodiment of the present invention, which is an optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation. Detailed Implementation
[0043] The present invention will be further described below with reference to the accompanying drawings and exemplary embodiments.
[0044] This invention discloses an optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation. The method is based on the Linux operating system, and the training platform uses an Nvidia GeForce RTX 3090, an Intel(R) Xeon(R) CPU E5-2680 28-core CPU processor, and the code execution framework is PyTorch. The runtime environment is Python 3.7 and CUDA 11.4.
[0045] The small object detection dataset used is the AI-TOD dataset provided by the School of Electronic Information at Wuhan University, and its example image is shown below. Figure 2 As shown, the AI-TOD dataset contains 282,580 object instances in 14,018 aerial images. Compared to existing aerial image target detection datasets, the average object size in the AI-TOD dataset is 12.8 pixels, significantly smaller than other existing aerial image datasets. This dataset includes 11,214 training images and 2,804 validation images, divided into eight categories. The number of objects in each category is as follows: Figure 1 As shown.
[0046] This invention provides an optimization method for a small target detection model of an unmanned aerial vehicle (UAV) optoelectronic pod based on collaborative knowledge distillation, comprising the following steps:
[0047] S1 downloads the small object detection dataset and converts the labels of the small object detection dataset into COCO format. The small object detection dataset includes a training dataset for training and a validation dataset for evaluating performance.
[0048] S2 uses the training dataset of the small object detection dataset to pre-train the teacher model, trains the teacher model for 12 training cycles to obtain the trained teacher model, and saves the optimal weights of the teacher model during training, which are then used as weights to guide the training of the student model.
[0049] S3 employs a collaborative knowledge distillation-based algorithm. It utilizes a pre-trained teacher model and its optimal weights to train the student model using a training dataset for Gaussian receptive field modeling, enabling the student model to learn the teacher model's bounding box assignment method. Simultaneously, it uses a collaborative knowledge distillation-based algorithm to train the foreground and background of the target. After training, a corresponding weight is generated.
[0050] S4 uses the student model trained in step S3 to train on the validation dataset, and uses the test results to score the weights obtained in step S3.
[0051] S5 returns to step S3 and repeats it 12 times, finding the weight with the highest score as the optimal weight for training the student model. The corresponding student model is a detection model suitable for small targets, thus completing the optimization of the small target detection model.
[0052] Step S3 specifically involves:
[0053] S3.1 Gaussian modeling is performed on the effective receptive fields of the teacher model and the student model to obtain the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model.
[0054] S3.2 calculates the KL divergence between the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model. By minimizing the KL divergence, the student model learns the assignment method of the annotation box from the teacher model.
[0055] S3.3 simultaneously utilizes an algorithm based on collaborative knowledge distillation to train the foreground and background of the target;
[0056] After training is complete, S3.4 generates a corresponding weight.
[0057] Step S3.1 specifically involves:
[0058] S3.1.1 Based on the kernel size and stride of each convolutional layer in the student model, the effective receptive field r at each feature point in each convolutional layer is derived. n ;
[0059]
[0060] Where: k n and s i This represents the kernel size and stride of the nth convolutional operation;
[0061] S3.1.2 Using the location (x) of the feature points of each student model n ,y n μ is the average vector of a standard two-dimensional Gaussian distribution. e This will effectively allow you to experience the wild. n The square of Σ is used as the covariance matrix of the two-dimensional Gaussian distribution. e The standard two-dimensional Gaussian distribution N of the effective receptive field is obtained. e (μ e ,∑ e );
[0062]
[0063] S3.1.3 Using the center point (x) of the bounding box in the training dataset trained by each teacher model g ,yg μ is the average vector of a standard two-dimensional Gaussian distribution. g The covariance matrix Σ is the square of the half-side length of the annotation box. g The standard two-dimensional Gaussian distribution N of the effective receptive field of the teacher model was obtained. g (μ g ,Σ g );
[0064]
[0065] Where: w g with h g These are the half-lengths of the annotation box;
[0066] S3.2 specifically involves calculating the KL divergence D between the standard two-dimensional Gaussian distribution of the effective receptive field of the student model and the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model. KL (N e ||N g By minimizing the KL divergence D KL (N e ||N g This allows the student model to learn the method of assigning annotation boxes from the teacher model;
[0067]
[0068] S3.3 specifically involves calculating the loss L of local and global synergistic distillation. fea It is used to evaluate the consistency between the feature points extracted by the student model and the feature points extracted by the teacher model, and to train the foreground and background of the target.
[0069]
[0070] Where: M is the mask for extracting the foreground, A S and A C F represents the spatial attention mask and channel attention mask of the teacher detector built into the teacher model, respectively. T and F S Let F represent the feature maps of the teacher detector and the student detector built into the student model, respectively. f is the feature transformation that makes the feature map F of the student detector... S Feature map F of the teacher-specific detector T In the same dimension, C is the number of channels in the teacher model and the student model, H and W are the height and width of the feature maps of the teacher model and the student model, respectively, α and β are the hyperparameters for balancing the foreground and background losses, i,j are the pixel coordinates, k is the Kth channel, S is the student model, and T is the teacher model.
[0071] In step S2 of this invention, the teacher model uses ResNet-101, and in step S3, the student model uses ResNet-18 or ResNet-50. In step S2, the teacher model is trained for 12 training cycles. In step S5, step S3 is returned and repeated 12 times. Such a number of training cycles can ensure that the student model is effectively trained and can also prevent the student model from overfitting to the teacher model.
[0072] In summary, the optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation of the present invention specifically includes the following steps:
[0073] S1 downloads the small object detection dataset and converts the labels of the small object detection dataset into COCO format. The small object detection dataset includes a training dataset for training and a validation dataset for evaluating performance.
[0074] S2 uses the training dataset of the small object detection dataset to pre-train the teacher model. The teacher model adopts ResNet-101. The teacher model is trained for 12 training cycles to obtain the trained teacher model. The optimal weights of the teacher model during training are saved and used as weights to guide the training of the student model.
[0075] S3.1.1 Based on the kernel size and stride of each convolutional layer in the student model, the effective receptive field r at each feature point in each convolutional layer is derived. n The student model is either ResNet-18 or ResNet-50;
[0076]
[0077] Where: k n and s i This represents the kernel size and stride of the nth convolutional operation;
[0078] S3.1.2 Using the location (x) of the feature points of each student model n ,y n μ is the average vector of a standard two-dimensional Gaussian distribution. e This will effectively allow you to experience the wild. n The square of Σ is used as the covariance matrix of the two-dimensional Gaussian distribution. e The standard two-dimensional Gaussian distribution N of the effective receptive field is obtained. e (μ e ,∑ e );
[0079]
[0080] S3.1.3 Using the center point (x) of the bounding box in the training dataset trained by each teacher modelg ,y g μ is the average vector of a standard two-dimensional Gaussian distribution. g The covariance matrix Σ is the square of the half-side length of the annotation box. g The standard two-dimensional Gaussian distribution N of the effective receptive field of the teacher model was obtained. g (μ g ,Σ g );
[0081]
[0082] Where: w g with h g These are the half-lengths of the annotation box;
[0083] S3.2 calculates the KL divergence D between the standard two-dimensional Gaussian distribution of the effective receptive field of the student model and the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model. KL (N e ||N g By minimizing the KL divergence D KL (N e ||N g This allows the student model to learn the method of assigning annotation boxes from the teacher model;
[0084]
[0085] S3.3 calculates the loss L of local and global co-distillation. fea It is used to evaluate the consistency between the feature points extracted by the student model and the feature points extracted by the teacher model, and to train the foreground and background of the target.
[0086]
[0087] Where: M is the mask for extracting the foreground, A S and A C F represents the spatial attention mask and channel attention mask of the teacher detector built into the teacher model, respectively. T and F S Let F represent the feature maps of the teacher detector and the student detector built into the student model, respectively. f is the feature transformation that makes the feature map F of the student detector... S Feature map F of the teacher-specific detector T In the same dimension, C is the number of channels in the teacher model and the student model, H and W are the height and width of the feature maps of the teacher model and the student model, respectively, α and β are hyperparameters that balance the foreground and background losses, i,j are the pixel coordinates, k is the Kth channel, S is the student model, and T is the teacher model.
[0088] S4 uses the student model trained in step S3 to test on the validation dataset, and uses the test results to score the weights obtained in step S3.
[0089] S5 returns to step S3 and repeats it 12 times, finding the weight with the highest score as the optimal weight for training the student model. The corresponding student model is a detection model suitable for small targets, thus completing the optimization of the small target detection model.
[0090] The following table shows the results of training the model on the AI-TOD dataset using the optimization method of the UAV electro-optical pod small target detection model based on collaborative knowledge distillation of the present invention and the directly trained model:
[0091]
[0092] Students 1 and 2 represent the results after training with ResNet50 and ResNet18 backbone networks respectively, using the method of this invention. ResNet50 and ResNet18 represent the results after direct training. As shown in the table above, compared to directly trained networks, the detection accuracy is significantly improved, while the computational cost is reduced. Specifically, for Student 1 with ResNet50 backbone network, the detection accuracy is improved by 59.8% and the computational cost is reduced by 23.1% compared to directly trained networks. For Student 2 with ResNet18 backbone network, the detection accuracy is improved by 39.0% and the computational cost is reduced by 57.25% compared to directly trained networks.
[0093] To verify the effectiveness of the Gaussian receptive field modeling training strategy and the target foreground and background training strategies in step S3 of this method, the following ablation experiment was conducted. First, both training strategies were removed and training was performed directly. Then, only the Gaussian receptive field modeling training strategy was retained for training. Next, the target foreground and background training strategies were retained. Finally, both the Gaussian receptive field modeling training strategy and the target foreground and background training strategies were used for training. The specific experimental results are shown in the table below:
[0094]
[0095] The experimental results shown in the table above indicate that, compared to directly trained models, both models showed improved detection accuracy after incorporating the Gaussian receptive field modeling training strategy. Specifically, Student 1, with a ResNet50 backbone, saw a 3.5% improvement in accuracy, while Student 2, with a ResNet18 backbone, improved by 4.8%. This is because the Gaussian receptive field modeling training strategy allows the student model to learn the teacher model's bounding box allocation strategy more effectively, addressing the previous issue of neglecting small targets in bounding box allocation. Similarly, incorporating a target foreground and background training strategy also improved the detection accuracy of both models compared to directly trained models. Student 1, with a ResNet50 backbone, saw a 7.1% improvement in accuracy, while Student 2, with a ResNet18 backbone, improved by 6.7%. This is because the target foreground and background training strategy balances the foreground and background, with local distillation focusing more on small targets and global distillation focusing more on medium- to large-scale targets, resulting in improved target detection accuracy. The table also shows that training with both strategies simultaneously achieved the highest detection accuracy and the fastest detection speed. Compared with existing technologies, this method improves the training method of the student model without changing the existing model structure, making it universal for different models.
[0096] The embodiments described above are merely illustrative of specific implementations of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for optimizing a small target detection model for UAV optoelectronic pods based on collaborative knowledge distillation, characterized in that, Includes the following steps: S1 downloads the small object detection dataset and converts the labels of the small object detection dataset into COCO format. The small object detection dataset includes a training dataset for training and a validation dataset for evaluating performance. S2 uses the training dataset of the small object detection dataset to pre-train the teacher model, trains the teacher model for at least 10 training cycles to obtain the trained teacher model, and saves the optimal weights during the training of the teacher model, which are then used as weights to guide the training of the student model. S3 employs a collaborative knowledge distillation-based algorithm. It utilizes a pre-trained teacher model and its optimal weights to train the student model using a training dataset for Gaussian receptive field modeling, enabling the student model to learn the teacher model's bounding box assignment method. Simultaneously, it uses a collaborative knowledge distillation-based algorithm to train the foreground and background of the target. After training is complete, a corresponding weight is generated. S4 uses the student model trained in step S3 to test on the validation dataset, and uses the test results to score the weights obtained in step S3. S5 returns to step S3 and repeats at least 10 times to find the weight with the highest score as the optimal weight for training the student model. The corresponding student model is a detection model suitable for small targets, thus completing the optimization of the small target detection model.
2. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 1, characterized in that, Step S3 specifically involves: S3.1 Gaussian modeling is performed on the effective receptive fields of the teacher model and the student model to obtain the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model. S3.2 calculates the KL divergence between the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model and the standard two-dimensional Gaussian distribution of the effective receptive field of the student model. By minimizing the KL divergence, the student model learns the assignment method of the annotation box from the teacher model. S3.3 simultaneously utilizes an algorithm based on collaborative knowledge distillation to train the foreground and background of the target; After training is complete, S3.4 generates a corresponding weight.
3. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 2, characterized in that, Step S3.1 specifically includes: S3.1.1 Based on the kernel size and stride of each convolutional layer in the student model, the effective receptive field r at each feature point in each convolutional layer is derived. n ; Where: k n and s i This represents the kernel size and stride of the nth convolutional operation; S3.1.2 Using the location (x) of the feature points of each student model n ,y n μ is the average vector of a standard two-dimensional Gaussian distribution. e This will effectively allow you to experience the wild. n The square of Σ is used as the covariance matrix of the two-dimensional Gaussian distribution. e The standard two-dimensional Gaussian distribution N of the effective receptive field of the student model is obtained. e (μ e ,Σ e ); S3.1.3 Using the center point (x) of the bounding box in the training dataset trained by each teacher model g ,y g μ is the average vector of a standard two-dimensional Gaussian distribution. g The covariance matrix ∑ is the square of the half-side length of the label box. g The standard two-dimensional Gaussian distribution N of the effective receptive field of the teacher model was obtained. g (μ g ,∑ g ); Where: w g with h g These are the half-lengths of the annotation box; Step S3.2 specifically involves calculating the KL divergence D between the standard two-dimensional Gaussian distribution of the effective receptive field of the student model and the standard two-dimensional Gaussian distribution of the effective receptive field of the teacher model. KL (N e ||N g By minimizing the KL divergence D KL (N e ||N g This allows the student model to learn the method of assigning annotation boxes from the teacher model; 4. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 3, characterized in that, Step S3.3 specifically involves: calculating the loss L of local and global synergistic distillation. fea It is used to evaluate the consistency between the feature points extracted by the student model and the feature points extracted by the teacher model, and to train the foreground and background of the target. Where: M is the mask for extracting the foreground, A S and A C F represents the spatial attention mask and channel attention mask of the teacher detector built into the teacher model, respectively. T and F S Let F represent the feature maps of the teacher detector and the student detector built into the student model, respectively. f is the feature transformation that makes the feature map F of the student detector... S Feature map F of the teacher-specific detector T In the same dimension, C is the number of channels in the teacher model and the student model, H and W are the height and width of the feature maps of the teacher model and the student model, respectively, α and β are hyperparameters that balance the foreground and background losses, i,j are the pixel coordinates, k is the Kth channel, S is the student model, and T is the teacher model.
5. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 4, characterized in that: In step S2, the teacher model adopts ResNet-101; In step S3, the student model uses ResNet-18 or ResNet-50.
6. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 5, characterized in that: In step S2, the teacher model is trained for 12 training cycles.
7. The optimization method for small target detection model of UAV optoelectronic pod based on collaborative knowledge distillation according to claim 6, characterized in that: In step S5, return to step S3 and repeat 12 times.