Remote sensing image ship detection method based on multiple teachers and feature distillation

Abstract

The invention discloses a remote sensing image ship detection method based on multi-teacher and feature distillation, and belongs to the field of knowledge distillation. The method specifically comprises the following steps: constructing a remote sensing ship image data set which comprises a training set and a verification set, and constructing at least two large teacher network models and a student network model; inputting the remote sensing ship image in the training set into a large teacher network and student network model, and performing ship target detection training in the remote sensing ship image; performing multi-teacher combined mask generation feature distillation training on the student model by using the multi-teacher model to obtain a small student model; simulating deployment of the student model on the unmanned aerial vehicle or the satellite terminal by limiting resources used by the student model, and detecting and verifying the student model by using the verification set to obtain a final student model; and performing ship detection on the to-be-detected remote sensing ship image by using the final student model. The remote sensing image ship detection is realized by deploying a student network model under the condition of limited resources.

Description

Technical Field

[0001] This invention relates to the fields of image processing and target detection, and more specifically, to a method for ship detection in remote sensing images based on multi-teacher and feature distillation. Background Technology

[0002] In recent years, my country's aerospace industry has flourished, and remote sensing satellite technology has matured significantly. Imaging remote sensing satellites, covering visible light and various electromagnetic spectrum bands, are continuously transmitting real-time Earth information. In particular, remote sensing images taken by optical remote sensing satellites equipped with visible light cameras are of great importance in the field of ship inspection. Significant progress has also been made in target detection technology within the field of computer vision.

[0003] The development of deep learning technology has enabled various object detection algorithms to achieve high levels of speed and accuracy with the support of high-performance GPUs. This is mainly due to the complex network structures and massive amounts of training data of these algorithms. However, this also brings with it a large number of parameters and computational demands on the network models, as well as the high power consumption and high computing power requirements of the hardware. This limits the deployment of these models on platforms with limited computing resources, such as mobile devices or embedded systems. Remote sensing image ship target detection algorithms, on the other hand, need to be deployed on platforms with limited computing resources, such as drones and satellites, and have a widespread need for embedded deployment. This presents new requirements for lightweight algorithm models.

[0004] Knowledge distillation, as an effective model optimization technique, can significantly improve model efficiency and reduce the number of model parameters without sacrificing too much performance. However, in the field of ship detection in remote sensing images, common knowledge distillation methods only use a single teacher model. This limits the generalization ability of the student model and makes it susceptible to being misled by incorrect knowledge from the teacher model, leading to a decrease in accuracy. Furthermore, since distillation is only performed on the output layer, the student model cannot learn the internal details and reasoning process of the teacher model, resulting in low learning efficiency.

[0005] Chinese patent application, publication number CN 116486285, describes a target detection method for aerial images based on category mask distillation. This method can achieve distillation of the effective region of a specified category, improve the learning of difficult samples, and enhance the detection performance and generalization ability of the student network under the guidance of the teacher network. However, it has the problem that using only one teacher model makes the student model susceptible to being misled by the teacher model's erroneous knowledge, leading to a decline in model performance. Summary of the Invention

[0006] 1. Technical problems to be solved

[0007] To address the shortcomings of the prior art, this invention provides a remote sensing image ship detection algorithm based on multi-teacher and feature distillation. By performing feature distillation and output layer distillation on the student model through multiple teachers, a lightweight student network model with good generalization ability is obtained, so as to realize the deployment of the student network model to achieve remote sensing image ship detection under limited resources.

[0008] 2. Technical Solution

[0009] The objective of this invention is achieved through the following technical solutions.

[0010] This invention provides a remote sensing image ship detection method based on multi-teacher and feature distillation, the specific steps of which are as follows:

[0011] Step S1: Construct a remote sensing ship image dataset, preprocess the images in the dataset, and divide the preprocessed images into a training set and a validation set. The preprocessed images have been labeled with ship targets. Construct at least two large teacher network models and one student network model. The student network model has the same type of network structure as the teacher network model, but the student network has fewer layers than the teacher network.

[0012] Step S2: Input the remote sensing ship images in the training set into the large teacher network and student network models to train ship target detection in the remote sensing ship images; acquire the output information of the teacher network model and student network model in parallel; extract the intermediate layer feature parameters and output layer parameters from the student network model and each teacher network model; evaluate the ship feature extraction capability, detection accuracy and recall of the teacher network model, and save the best parameters of the teacher model.

[0013] Step S3 involves using at least two teacher models trained in step S2 to perform multi-teacher combined mask-generated feature distillation training on the student model, resulting in a smaller student model. During the multi-teacher model distillation training of the student model, feature distillation and output layer distillation are implemented. Feature distillation employs a mask-generated feature distillation method; specifically, through a masking mechanism, the student model is forced to use some of its own features to generate all the features of the teacher models. Output layer distillation uses a non-target category normalization method.

[0014] Step S4: Simulate the deployment of the student model on a drone or satellite terminal by limiting the resources used by the student model, and use a validation set to test and validate the obtained student model to obtain the final student model.

[0015] Step S5: Use the final student model to perform ship detection on the remote sensing image under test.

[0016] Furthermore, the process of step S2 is specified as follows:

[0017] The teacher model is preferably designed with a network structure similar to the student model, but with a deeper network structure, more parameters, and better performance. The teacher model is trained on the training set and evaluated using the validation set. Its performance in accuracy, precision, and recall is recorded, and the optimal parameters of the teacher model are saved to disk.

[0018] Furthermore, the process described in step S3 can be visualized as follows:

[0019] Distillation training of the student model using a multi-teacher model involves two aspects: feature distillation and output layer distillation; the total loss function of the student model during training is also considered. all It can be represented as:

[0020]

[0021] Among them, Loss original Let ρ represent the original loss of the student model, Q represent the total number of teacher models, and ρ represent the original loss of the student model. 1,i The hyperparameter representing the balance feature loss of the i-th teacher model is usually taken as . Loss fea,i Let ρ represent the feature loss of the i-th teacher model. 2,i Loss represents the hyperparameter of the loss of the balanced output layer of the i-th teacher model. logits,i This represents the output layer loss of the i-th teacher model.

[0022] Among them, the hyperparameter ρ of the balanced output layer loss of the i-th teacher model 2,i During the calculation process, an adaptive attention mechanism is applied, using confidence level Z. i Adjusting the impact of different teacher models on student models; the specific calculation method is as follows:

[0023]

[0024] Among them, Z i This represents the confidence level of the i-th teacher model. Teachers with higher confidence levels have a greater impact on the student development model. The confidence level is calculated as follows:

[0025]

[0026] Among them, Z i Z represents the confidence level of the i-th teacher model. n Let f represent the confidence score of the nth teacher model, y represent the true label value, σ represent the softmax operation, and f i This represents the soft label value output by the i-th teacher model;

[0027] In the feature distillation process, feature distillation is performed on each layer of the feature map multiple times at different depths. The specific method for feature distillation on a particular layer of the feature map is as follows: a mask-based feature generation method is used, allowing the student model to generate features from the teacher model; the mask for the student model's features is generated randomly.

[0028]

[0029] in, The mask value represents the position (a, b) of the feature map of layer l, where a and b represent the x and y coordinates of the feature map of layer l, respectively. λ represents the random number at position (0,1) of (a,b) in the l-th layer feature map, and λ represents the hyperparameter of the mask coverage ratio, which is usually set to 0.65.

[0030] After generating the mask through the above steps, use the mask to cover the feature map of the student model, and then generate the teacher's feature map by convolution on the feature map covered by the mask:

[0031]

[0032] in, This represents the projection layer, which consists of three convolutional layers and two ReLU activation layers, f align This represents the adaptation layer, which consists of a 1×1 convolutional layer, S. l This represents the feature map of the l-th layer of the student model, ⊙ represents element-wise multiplication, and M... l T represents the mask value of the feature map of layer l. l This represents the feature map of the l-th layer of the teacher model, Covn. 3×3 This represents a 3×3 convolutional layer, where ReLU represents a corrected linear unit;

[0033] Using the above method, the loss of feature distillation for the i-th teacher model is... fea,i Represented as:

[0034]

[0035] Where L represents the total number of distillation layers, C represents the number of channels, H represents the height of the input feature map, and W represents the width of the input feature map. This represents the feature value of the teacher model at position l, channel k, height h, and width w. This represents the feature value of the student model at the position of layer l, channel k, height h, and width w.

[0036] During output layer distillation, the non-target class probability distribution output by the student model is constantly changing and cannot be exactly equal to the non-target class probability distribution output by the teacher model. This results in unequal sums of the two non-target losses, hindering the student model from becoming closer to the teacher model. Therefore, normalization is performed on the two non-target loss distributions output by the student and teacher models respectively, forcing them to be equal. The output layer loss of the i-th teacher model is... logits,i Represented as:

[0037]

[0038] Among them, P i t O represents the predicted probability of the i-th teacher model for the target category t. t Let represent the student model's predicted probability for the target class t, γ represent the hyperparameter of the balancing loss, used to adjust the weights between the two loss terms, τ represent the temperature parameter for knowledge distillation, and N represent the total number of classes. P represents the normalization operation. i j,τ Let represent the prediction probability of the i-th teacher model for the non-target category j. This represents the student model's prediction probability for the non-target class j.

[0039] Preferably, the teacher model consists of three components: one with an improved YOLOv8 network structure, and the other two with YOLOv8l and YOLOv8m network structures. The student model uses a YOLOv8s network structure. The improved YOLOv8 network integrates an improved large kernel attention module into the YOLOv8 backbone network, and adds a spatial explicit vision module after the backbone network. The improved large kernel attention module is obtained by splitting the two-dimensional convolutional kernels of the original large kernel attention module's depthwise convolution and depthwise extended convolution into two cascaded one-dimensional separable convolutional kernels. The spatial explicit vision module includes two parallel modules: a lightweight MLP architecture and a learnable visual center.

[0040] Preferably, when generating a mask for the features of the student model, six layers from the teacher model are selected to perform feature distillation on the student model. Among these, three layers are the feature maps of the last three layers of the YOLO backbone network during feature extraction, and the other three layers are the feature maps of the three detector heads of YOLOv8.

[0041] 3. Beneficial effects

[0042] Compared with the prior art, the advantages of this invention are:

[0043] This invention effectively reduces the number of parameters in student models, enabling them to be deployed on platforms with limited computing resources, while maintaining the accuracy and generalization ability of the student models and reducing the resources required for training them. Its main contributions include:

[0044] 1. Using multiple teachers to distill students reduces the impact of errors in individual teacher models on student models, thereby improving the accuracy and generalization ability of student models.

[0045] 2. The distillation method using masked feature generation ensures that the student model does not simply imitate the teacher model. Instead, it forces the student model to use some of its own features to generate all the features of the teacher model, thus enabling the student model to obtain better feature extraction capabilities. In ship detection of remote sensing images, it can accurately extract the features of target ships in complex environments with noise and interference.

[0046] 3. The distillation method of normalized output layer is used to normalize the distribution of non-target loss, which ensures that the output distribution of the student model is better aligned with the distribution of the teacher model, improves the knowledge transfer effect during the distillation process, and enables the student model to better classify target ships.

[0047] 4. When calculating the total loss function of the student model during training, an adaptive attention mechanism is used so that teachers with higher confidence have a greater impact on training the student model. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method steps of the present invention;

[0049] Figure 2 A schematic diagram of multi-teacher characteristic distillation;

[0050] Figure 3 A schematic diagram of multi-teacher output layer distillation;

[0051] Figure 4 A distillation method for generating features from a mask. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0053] Example

[0054] This invention provides a remote sensing image ship detection method based on multi-teacher and feature distillation, the specific steps of which are as follows:

[0055] Step S1: Construct a remote sensing ship image dataset, preprocess the images in the dataset, and divide the preprocessed images into a training set and a validation set. The preprocessed images have been labeled with ship targets. Construct at least two large teacher network models and one student network model. The student network model has the same type of network structure as the teacher network model, but the student network has fewer layers than the teacher network.

[0056] Step S2: Input the remote sensing ship images in the training set into the large teacher network and student network models to train ship target detection in the remote sensing ship images; acquire the output information of the teacher network model and student network model in parallel; extract the intermediate layer feature parameters and output layer parameters from the student network model and each teacher network model; evaluate the ship feature extraction capability, detection accuracy and recall of the teacher network model, and save the best parameters of the teacher model.

[0057] Step S3: Using at least two teacher models trained in step S2, the student model is trained using multi-teacher combined mask-generated feature distillation to obtain a small student model. During the multi-teacher model's distillation training of the student model, feature distillation and output layer distillation are implemented. Feature distillation uses a mask-generated feature distillation method. Specifically, through a mask mechanism, the student model is forced to use some of its own features to generate all the features of the teacher models. Output layer distillation uses a non-target category normalization method.

[0058] Step S4: Simulate the deployment of the student model on a drone or satellite terminal by limiting the resources used by the student model, and use the validation set to test and validate the obtained student model to obtain the final student model;

[0059] Step S5: Use the final student model to perform ship detection on the remote sensing ship image to be tested.

[0060] Furthermore, the process described in step S2 can be visualized as follows:

[0061] like Figure 2 , Figure 3 As shown, this example uses three teacher models: one is an improved YOLOv8 network structure, and the other two are YOLOv8l and YOLOv8m network structures. The teacher models have deeper network layers, more parameters, and better performance than the student models. The three teacher models are trained using the training set and evaluated using the test set. Their performance in terms of accuracy, precision, recall, etc., is recorded, and the best parameters of the three teacher models are saved to disk.

[0062] In this embodiment, one of the teacher models, namely the improved YOLOv8 network structure, specifically comprises an improved large kernel attention module integrated into the YOLOv8 backbone network. The improved large kernel attention module is connected to a cross-stage partial network (C2f) to obtain the improved cross-stage partial network (C2f_LSK) module. Furthermore, a spatial explicit vision module is added after the backbone network.

[0063] Furthermore, the improved large kernel attention module is obtained by splitting the two-dimensional convolutional kernels of the original large kernel attention module's depthwise convolution and depthwise extended convolution into two cascaded one-dimensional separable convolutional kernels. The improved large kernel attention module replaces the structure and function of the original large kernel attention module in the network.

[0064] Furthermore, the spatial vision module comprises two parallel modules: a lightweight MLP (Multilayer Perceptron) architecture and a Learnable Visual Center (LVC).

[0065] In fact, this invention does not limit the teacher network to any particular network structure, as long as it performs well in ship recognition tasks using remote sensing ship images.

[0066] Furthermore, the process described in step S3 can be visualized as follows:

[0067] In this example, the student model uses a YOLOv8s network structure. Training data is input into both the student and teacher models, resulting in the intermediate and output layer feature parameters for both models. There is no specific limit to the number of layers in the student network, but it is generally fewer than that of the teacher model. Ideally, the teacher and student networks should be of the same type, as a closer resemblance in network structure between the student and teacher models leads to better distillation results and facilitates better learning. However, teacher models typically have deeper layers, more parameters, and sometimes more attention mechanisms, while student models can only be similar, not identical. They generally share the same backbone network structure, such as a YOLO backbone or ResNet50.

[0068] Using the teacher model to distill the student model involves two aspects: feature distillation and output layer distillation; the total loss function of the student model during training is also considered. all It can be represented as:

[0069]

[0070] Among them, Loss original ρ represents the original loss of the student model, Q represents the total number of teacher models (3 in this example), and ρ represents the original loss of the student model. 1,i The hyperparameter representing the balance feature loss of the i-th teacher model is usually taken as . Loss fea,i Let ρ represent the feature loss of the i-th teacher model. 2,i Loss represents the hyperparameter of the loss of the balanced output layer of the i-th teacher model. logits,i Let represent the output layer loss of the i-th teacher model. For the hyperparameters balancing the output layer loss of output layer distillation, an adaptive attention mechanism is introduced to ensure that teachers with higher confidence have a greater impact on the student training model.

[0071]

[0072] Among them, Z i Z represents the confidence level of the i-th teacher model. n Let f represent the confidence score of the nth teacher model, y represent the true label value, σ represent the softmax operation, and f i This represents the soft label value output by the i-th teacher model;

[0073] like Figure 4 As shown, during feature distillation, a mask-based feature generation method is used, allowing the student model to generate features for the teacher model. When generating the mask for the student model's features, a random generation method is used. In this example, for the YOLO model, generally 6 layers from the teacher model are selected for feature distillation of the student model. Three of these layers are the feature maps from the last three layers of the YOLO backbone network during feature extraction, and the other three layers are the feature maps from the three detector heads of YOLOv8.

[0074]

[0075] in, The mask value at position (a, b) in the l-th layer feature map is used to blur the features generated by the student model; a and b represent the x and y coordinates of the l-th layer feature map, respectively. λ represents the random number at position (0,1) of (a,b) in the l-th layer feature map, and λ represents the hyperparameter of the mask coverage ratio. Preferably, λ is usually set to 0.65.

[0076] After generating the mask through the above steps, the corresponding mask is used to cover the feature map of the student model. Then, the feature map covered by the mask is used to attempt to generate the feature map of the teacher through convolution.

[0077]

[0078] in, This refers to the projection layer, which is used in the knowledge distillation process but is not part of the network structure; it acts as an intermediate step in the distillation process. The projection layer consists of three convolutional layers and two ReLU activation layers, arranged in the following order: first a convolution, then a ReLU activation layer, repeated once, and finally a final convolution. align This represents the adaptation layer, which consists of a 1×1 convolutional layer, S. l This represents the feature map of the l-th layer of the student model, ⊙ represents element-wise multiplication, and M... l , representing the mask value of the feature map of the l-th layer, T l This represents the feature map of the l-th layer of the teacher model, Covn. 3×3 This represents a 3×3 convolutional layer, where ReLU represents a corrected linear unit;

[0079] Using the above method, the loss of feature distillation for the i-th teacher model is... fea,i Represented as:

[0080]

[0081] Where L represents the total number of distillation layers (6 layers in this example), C represents the number of channels, H represents the height of the input feature map, and W represents the width of the input feature map. This represents the feature value of the teacher model at position l, channel k, height h, and width w. This represents the feature value of the student model at the position of layer l, channel k, height h, and width w.

[0082] During output layer distillation, the probability distribution of non-target classes output by the student model is constantly changing and cannot be exactly equal to the probability distribution of non-target classes output by the teacher model. This results in unequal sums of the two non-target losses, hindering the student model from becoming closer to the teacher model. Therefore, normalization is performed on the two non-target loss distributions to force them to be equal. Normalizing the two non-target loss distributions to force them to be equal, so that the sum of the two non-target losses is equal, can make the student model closer to the teacher model. In multi-class classification problems in object detection, the model predicts all probabilities of which category the predicted box belongs to, and then uses non-maximum suppression to determine the category of the predicted box. The categories other than the true category of the predicted box are called non-target categories.

[0083] Loss of the output layer of the i-th teacher model logits,i Represented as:

[0084]

[0085] Among them, Pi t O represents the predicted probability of the i-th teacher model for the target category t. t Let represent the student model's predicted probability for the target class t, γ represent the hyperparameter of the balancing loss, used to adjust the weights between the two loss terms, τ represent the temperature parameter for knowledge distillation, and N represent the total number of classes. P represents the normalization operation. i j,τ Let represent the prediction probability of the i-th teacher model for the non-target category j. This represents the student model's prediction probability for the non-target class j.

[0086] Using the above method, the student model needs to undergo feature distillation and output layer distillation in each round of training. After repeated iterations, the student model achieves improved accuracy in detecting ships in remote sensing images and has strong generalization ability compared to the original model, which has the same number of parameters.

[0087] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. No reference numerals in the claims should limit the scope of the claims. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the scope of protection of the present invention. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Multiple elements stated in the product claims may also be implemented by a single element through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A remote sensing image ship detection method based on multi-teacher and feature distillation, characterized in that: The steps are as follows: Step S1, constructing a remote sensing ship image data set, preprocessing the images in the data set, dividing the preprocessed images into a training set and a verification set, wherein the preprocessed images have completed the annotation of ship targets; constructing at least two large teacher network models and one student network model; the student network model has the same type of network structure as the teacher network model, and the student network has fewer layers than the teacher network; Step S2, inputting the remote sensing ship images in the training set into the large teacher network and student network model to perform ship target detection training in the remote sensing ship images; obtaining the output information of the teacher network model and the student network model in parallel; extracting the intermediate layer feature parameters and the output layer parameters from the student network model and each teacher network model; evaluating the ship feature extraction capability, detection accuracy and recall rate indicators of the teacher network model, and saving the optimal parameters of the teacher model; Step S3, using at least two teacher models trained in step S2 to perform distillation training on the student model using multiple teachers combined with mask-generated features, to obtain a small student model; in the process of distillation training the student model using multiple teacher models, feature distillation and output layer distillation are implemented; feature distillation adopts a distillation method of mask-generated features, specifically, through a mask mechanism, the student model is forced to use part of its own features to generate all features of the teacher model; the output layer distillation adopts a non-target category normalization method; Step S4, simulating the deployment of the student model on the UAV or satellite terminal by limiting the resources used by the student model, and using the verification set to test and verify the obtained student model to obtain the final student model; Step S5: Use the final student model to perform ship detection on the remote sensing ship image to be tested.

2. According to claim 1, a remote sensing image ship detection method based on multi-teacher and feature distillation is characterized in that: In step S3, the multi-teacher model is used to perform distillation training on the student model, including feature distillation and output layer distillation; the total loss function Loss of the student model during training all It can be expressed as: Loss original represents the original loss of the student model, Q represents the total number of teacher models, and ρ 1,i represents the hyperparameter of the balanced feature loss of the i-th teacher model, Loss fea,i represents the feature loss of the i-th teacher model, ρ 2,i represents the hyperparameter of the balanced output layer loss of the i-th teacher model, Loss logits,i represents the output layer loss of the i-th teacher model.

3. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 2, characterized in that: The hyperparameter ρ of the balanced output layer loss of the i-th teacher model 2,i In the calculation process of , the adaptive attention mechanism is applied, through the confidence Z i Adjust the influence of different teacher models on the student model; the specific calculation method is: The confidence calculation method is: Z i represents the confidence of the i-th teacher model, Z n represents the confidence of the nth teacher model, y represents the true label value, σ represents the softmax operation, and f i Represents the soft label value output by the i-th teacher model.

4. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 3, characterized in that: In step S3, during the feature distillation process, feature distillation of one layer of feature maps is performed each time, and feature distillation is performed multiple times at different depths. The specific method for performing feature distillation on a specific layer of feature maps is as follows: a mask-generated feature distillation method is used to enable the student model to generate features of the teacher model by itself; a mask is generated by randomly generating features of the student model: in, Represents the mask value of the (a, b) position of the l-th layer feature map, where a and b represent the horizontal and vertical coordinates of the l-th layer feature map, respectively. represents the random number in the (a, b) position (0,1) of the feature map of the lth layer, and λ represents the hyperparameter of the mask coverage ratio.

5. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 4, characterized in that: Step S3 The characteristic distillation process also includes the following steps: After the mask is generated, the feature map of the student model is covered with the mask, and then the feature map covered by the mask is convolved to generate the teacher's feature map: in, represents the projection layer, which includes three convolutional layers and two ReLU activation layers. align represents the adaptation layer, including a 1×1 convolutional layer, S l represents the feature map of the first layer of the student model, ⊙ represents the point-by-point multiplication of elements, M l represents the mask value of the feature map of the lth layer, T l Represents the feature map of the first layer of the teacher model, Covn 3×3 represents a 3×3 convolutional layer, and ReLU represents a rectified linear unit.

6. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 5, characterized in that: Normalize the two non-target loss distributions output by the student model and the teacher model respectively to make them equal. The output layer loss Loss of the i-th teacher model is logits,i It is expressed as: Among them, P i t represents the predicted probability of the i-th teacher model for the target category t, O t represents the predicted probability of the student model for the target category t, γ represents the hyperparameter of the balanced loss, which is used to adjust the weight between the two loss terms, τ represents the temperature parameter of knowledge distillation, N represents the total number of categories, represents the normalization operation, P i j,τ represents the predicted probability of the i-th teacher model for non-target category j, represents the predicted probability of the student model for non-target category j.

7. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 6, characterized in that: There are three teacher models, one is an improved network structure, and the other two are YOLOv8l and YOLOv8m network structures. The improved YOLOv8 network integrates an improved large-core attention module into the YOLOv8 backbone network, and adds a spatial explicit vision module after the backbone network; the improved large-core attention module is obtained by splitting the two-dimensional convolution kernel of the deep convolution and deep dilation convolution of the original large-core attention module into two cascaded one-dimensional separable convolution kernels; The spatial explicit vision module includes two parallel modules: a lightweight multi-layer perceptron MLP architecture and a learnable visual center.

8. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 7, characterized in that: The student model uses the YOLOv8s network structure.

9. The method for ship detection in remote sensing images based on multi-teacher and feature distillation according to claim 8, characterized in that: When masking the features of the student model, 6 layers in the teacher model are selected to perform feature distillation on the student model, of which 3 layers are the last three layers of feature maps of the YOLO backbone network during feature extraction, and the other 3 layers are feature maps of the three detection heads of YOLOv8.

Images