A continuous learning target detection method and system based on knowledge distillation

Abstract

This invention provides a continuous learning target detection method and system based on knowledge distillation. The method includes: training a neural network based on images and old category labels in a training dataset to obtain a teacher model, and obtaining the teacher model's process feature vector during training; training the neural network based on images, new category labels, and the teacher model's process feature vector in the training dataset to obtain a student model, and obtaining the student model's training loss; obtaining a feature distillation loss based on three pieces of information in the teacher model's process feature vector and their corresponding masks; obtaining a knowledge distillation loss based on the student model's training loss and the feature distillation loss; training the student model based on the knowledge distillation loss to obtain an updated student model; and using the updated student model to perform target detection. This invention can solve the technical problem of low accuracy in existing target detection methods.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and continuous learning technology, and in particular to a continuous learning target detection method and system based on knowledge distillation. Background Technology

[0002] Biological vision systems can continuously acquire, integrate, and optimize knowledge, and their learning pattern is sequential. In contrast, current object detection models require pre-prepared complete datasets for training before they can be used for inference. This learning paradigm assumes that the data distribution is fixed, while real-world data is a continuous, dynamic data stream with a non-stationary distribution. This makes current detection models unable to adapt to the data flow and business needs of real-world environments.

[0003] To address the aforementioned problems, neural networks need to learn incrementally and continuously from dynamic data streams. The key challenges are: (1) the catastrophic forgetting of old knowledge. When learning a new category, after the model updates its parameters based on the new category's data, its performance on previously learned categories typically drops sharply, resulting in forgetting of previously learned categories. (2) the generalization problem of new knowledge. When learning from new data of the same category, the new data may unpredictably alter the data distribution of that category, thus affecting the accuracy of learning that category. Therefore, the main task of continuous learning is to address the issues of model stability and plasticity, requiring neural networks to prevent forgetting previously learned knowledge while maintaining the ability to learn new knowledge.

[0004] Knowledge distillation is a promising method for continuous learning, which extracts knowledge from models trained on previous tasks and incorporates it into models trained on new tasks to reinforce previously learned knowledge. In previous research, knowledge distillation has been widely applied to image classification tasks, while research on continuous learning for object detection tasks is relatively limited. Furthermore, the high complexity of detection tasks further increases the difficulty of continuous learning for object detection.

[0005] Current feature distillation methods suffer from inconsistencies in knowledge categories between new and old tasks, leading to knowledge confusion. Student models may learn useless or erroneous knowledge from teacher models, resulting in low accuracy on new tasks. Furthermore, due to the structure of deep neural networks (DNNs), only low-level semantic features can be distilled, lacking knowledge transfer through inter-pixel interactions. Summary of the Invention

[0006] This invention provides a continuous learning target detection method and system based on knowledge distillation, which can solve the technical problem of low accuracy in existing target detection methods.

[0007] According to one aspect of the present invention, a continuous learning target detection method based on knowledge distillation is provided, the method comprising:

[0008] The neural network is trained based on the images and old category labels in the training dataset to obtain the teacher model, and the process feature vector of the teacher model is obtained during the training process. The process feature vector of the teacher model includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model.

[0009] The neural network is trained based on the images, new category labels, and teacher model process feature vectors in the training dataset to obtain the student model, and the training loss of the student model is obtained.

[0010] The feature distillation loss is obtained based on three pieces of information in the feature vector of the teacher model process and their corresponding masks.

[0011] The loss after knowledge distillation is obtained based on the student model training loss and feature distillation loss;

[0012] The student model is trained using the loss obtained through knowledge distillation to obtain an updated student model.

[0013] Object detection is achieved using the updated student model.

[0014] Preferably, the characteristic distillation loss is obtained by the following formula:

[0015]

[0016] In the formula, L feature (F S ,F T Characteristic distillation loss is ) This is the final feature vector of the teacher model. For the final feature vector of the student model, For the teacher model process feature vector, Mfg is the feature vector of the student model process. i,j Mbg is a mask for target information in the teacher model. i,j The mask for the correct background information in the teacher model, α and β are the first and second hyperparameter coefficients respectively, and H rnew W rnew The new category target box r new The length and width of the vector are H and W, respectively, and i and j are the position coordinates of the pixel.

[0017] Preferably, the three pieces of information and their corresponding masks in the feature vector of the teacher model process are obtained by the following formula:

[0018]

[0019]

[0020]

[0021] In the formula, Mfg i,j This serves as a mask for the target information in the teacher model. The target boxes for the old categories are r. old Length and width, Mnew i,j This serves as a mask for erroneous background information in the teacher model. The new category target box r new Length and width, Mbg i,j Nbg is the mask for the correct background information in the teacher model, and Nbg is the bounding box r that does not belong to the old category. old It also does not belong to the new category of target boxes. new The total number of pixels in the region.

[0022] Preferably, the loss after knowledge distillation is obtained by the following formula:

[0023] L′ S =L feature (F S ,F T )+L S ;

[0024] In the formula, L′ S For the loss due to knowledge distillation, L S The training loss for the student model.

[0025] Preferably, the neural network is trained based on images and old category labels in the training dataset to obtain a teacher model, and the feature vector of the teacher model during the training process is obtained, including:

[0026] The teacher model backbone network is used to extract deep image features and shallow image features of the teacher model from the images and old category labels in the training dataset. The deep image features and shallow image features of the teacher model are then concatenated to obtain a multi-scale feature map of the teacher model.

[0027] The teacher model encoder obtains the process feature vector of the teacher model based on the multi-scale feature map of the teacher model.

[0028] The teacher model decoder obtains the final feature vector of the teacher model based on the feature vector of the teacher model process.

[0029] The teacher model classification subgrid obtains the teacher model classification soft label based on the final feature vector of the teacher model.

[0030] The teacher model regression subgrid obtains the teacher model regression soft label based on the final feature vector of the teacher model;

[0031] The teacher model activation function obtains the teacher model category label based on the teacher model classification soft label and the teacher model location label based on the teacher model regression soft label, thereby completing the training of the neural network and obtaining the teacher model.

[0032] Preferably, the neural network is trained based on images, new category labels, and feature vectors from the teacher model process in the training dataset to obtain the student model, which includes:

[0033] The deep and shallow image features of the student model are extracted from the images and new category labels in the training dataset using the student model backbone network. The deep and shallow image features of the student model are then concatenated to obtain a multi-scale feature map of the student model.

[0034] The student model encoder obtains the process feature vector of the student model based on the multi-scale feature map of the student model and the process feature vector of the teacher model.

[0035] The student model decoder obtains the final feature vector of the student model based on the process feature vector of the student model.

[0036] The student model classification subgrid obtains the soft label for student model classification based on the final feature vector of the student model.

[0037] The student model regression subgrid obtains the student model regression soft label based on the final feature vector of the student model;

[0038] The student model activation function obtains the student model category label based on the student model classification soft label and the student model location label based on the student model regression soft label, thereby completing the training of the neural network and obtaining the student model.

[0039] According to another aspect of the present invention, a continuous learning target detection system based on knowledge distillation is provided, the system comprising a network construction module, a knowledge distillation module, and an image target detection module;

[0040] The network building module is used to train the neural network based on the images and old category labels in the training dataset to obtain the teacher model, and to obtain the teacher model process feature vector during the training process. The teacher model process feature vector includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model.

[0041] The knowledge distillation module is used to train the neural network based on images, new category labels, and teacher model process feature vectors in the training dataset to obtain a student model and to obtain the student model training loss; it is used to obtain the feature distillation loss based on three pieces of information in the teacher model process feature vector and their corresponding masks; it is used to obtain the loss after knowledge distillation based on the student model training loss and the feature distillation loss; and it is also used to train the student model based on the loss after knowledge distillation to obtain an updated student model.

[0042] The image target detection module is used to store the updated student model and to perform target detection based on the updated student model.

[0043] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a knowledge distillation-based continuous learning target detection program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the knowledge distillation-based continuous learning target detection program.

[0044] By applying the technical solution of this invention, and separating foreground features (target information in the teacher model) and background features (correct background information in the teacher model and incorrect background information in the teacher model), different parameters are set for features in different regions. This reduces the transmission of erroneous information from the teacher model, enabling the student model to learn correct and useful knowledge. Consequently, it reduces the catastrophic forgetting problem of old tasks in continuously learning new tasks. This invention improves the accuracy of the information transmitted by the teacher model, mitigates catastrophic and transient forgetting problems, reduces model computational resources and time, and increases target detection accuracy. Attached Figure Description

[0045] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0046] Figure 1 A flowchart of a continuous learning target detection method based on knowledge distillation according to an embodiment of the present invention is shown;

[0047] Figure 2 A block diagram of a knowledge distillation-based continuous learning target detection method according to an embodiment of the present invention is shown.

[0048] Figure 3A schematic diagram of a continuous learning target detection method based on knowledge distillation according to an embodiment of the present invention is shown.

[0049] Figure 4 A schematic diagram of the structure of a knowledge distillation-based continuous learning target detection system according to an embodiment of the present invention is shown. Detailed Implementation

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0052] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0053] like Figure 1 and Figure 2 As shown, this invention provides a continuous learning target detection method based on knowledge distillation, the method comprising:

[0054] S10. Train the neural network based on the images and old category labels in the training dataset to obtain the teacher model, and obtain the teacher model process feature vector during the training process. The teacher model process feature vector includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model. The incorrect background information in the teacher model is equivalent to the target information in the student model.

[0055] S20. Train the neural network based on the images, new category labels and teacher model process feature vectors in the training dataset to obtain the student model and obtain the student model training loss.

[0056] S30. Obtain feature distillation loss based on three pieces of information in the feature vector of the teacher model process and their corresponding masks;

[0057] S40. Obtain the loss after knowledge distillation based on student model training loss and feature distillation loss;

[0058] S50. Train the student model based on the loss after knowledge distillation to obtain the updated student model;

[0059] S60. Implement object detection using the updated student model.

[0060] This invention separates foreground features (target information in the teacher model) and background features (correct background information and incorrect background information in the teacher model), sets different parameters for features in different regions, reduces the transmission of erroneous information from the teacher model, and enables the student model to learn correct and useful knowledge. This reduces the catastrophic forgetting problem of old tasks when learning new tasks in successive learning. This invention improves the accuracy of the information transmitted by the teacher model, mitigates catastrophic and transient forgetting problems, reduces model computational resources and time, and increases target detection accuracy.

[0061] According to one embodiment of the present invention, in S10, the neural network is trained based on the images and old category labels in the training dataset to obtain a teacher model, and the process feature vector of the teacher model during the training process is obtained, including:

[0062] S11. Using the teacher model backbone network, extract deep image features and shallow image features of the teacher model from the images and old category labels in the training dataset, and concatenate the deep image features and shallow image features of the teacher model to obtain the multi-scale feature map of the teacher model.

[0063] S12. The teacher model encoder obtains the process feature vector of the teacher model based on the multi-scale feature map of the teacher model.

[0064] S13. The teacher model decoder obtains the final feature vector of the teacher model based on the process feature vector of the teacher model; wherein, the final feature vector of the teacher model includes semantic information;

[0065] S14. The teacher model classification subgrid obtains the teacher model classification soft label based on the final feature vector of the teacher model.

[0066] S15. The teacher model regression subgrid obtains the teacher model regression soft label based on the final feature vector of the teacher model.

[0067] S16. The teacher model activation function obtains the teacher model category label based on the teacher model classification soft label and the teacher model location label based on the teacher model regression soft label, thereby completing the training of the neural network and obtaining the teacher model.

[0068] In this embodiment, the teacher model activation function can be the softmax activation function, and the training dataset, validation dataset, and test dataset of the teacher model all contain only the old class labels.

[0069] According to one embodiment of the present invention, in S20, the neural network is trained based on images, new category labels, and teacher model process feature vectors in the training dataset to obtain a student model, including:

[0070] S21. Using the student model backbone network, extract deep image features and shallow image features of the student model from the images and new category labels in the training dataset, and concatenate the deep image features and shallow image features of the student model to obtain a multi-scale feature map of the student model.

[0071] S22. The student model encoder obtains the process feature vector of the student model based on the multi-scale feature map of the student model and the process feature vector of the teacher model.

[0072] S23. The student model decoder obtains the final feature vector of the student model based on the process feature vector of the student model; wherein, the final feature vector of the student model includes semantic information;

[0073] S24. The student model classification subgrid obtains the soft label of the student model classification based on the final feature vector of the student model.

[0074] S25. The student model regression subgrid obtains the student model regression soft label based on the final feature vector of the student model.

[0075] S26. The student model activation function obtains the student model category label based on the student model classification soft label and the student model location label based on the student model regression soft label, thereby completing the training of the neural network and obtaining the student model.

[0076] In this embodiment, the student model can use the softmax activation function. Both the validation and test datasets for the student model contain only the new class labels, while the test dataset includes both the new and old class labels. During training, the student model is guided by the procedural feature vectors of the teacher model, enabling it to learn and converge better. This allows the student model to simultaneously identify both the new and old class labels on the test dataset, even when only learning the new class labels.

[0077] Since the teacher model does not include new class labels during training, it incorrectly treats these labels as background information. This incorrect background information (i.e., the erroneous background information in the teacher model) is information that the student model should discard during learning. To better guide the student model in learning important feature pixel regions, an attention mask is used to distinguish between the three types of information.

[0078] To mitigate the difference in distillation weight caused by varying target areas, the mask adjusts each target bounding box according to its own area. Specifically, the mask for target information in the teacher model is obtained using the following formula:

[0079]

[0080] The mask for the background information of errors in the teacher model is obtained using the following formula:

[0081]

[0082] Meanwhile, to reduce the imbalance between positive and negative samples caused by excessive background area due to too many pixels, and to prevent excessive attention to background information from interfering with the retention of old knowledge, the background was adjusted according to the total number of pixels. Specifically, the mask for the correct background information in the teacher model was obtained using the following formula:

[0083]

[0084] In the formula, Mfg i,j This serves as a mask for the target information in the teacher model. The target boxes for the old categories are r. old Length and width, Mnew i,j This serves as a mask for erroneous background information in the teacher model. The new category target box r new Length and width, Mbg i,j Nbg is the mask for the correct background information in the teacher model, and Nbg is the bounding box r that does not belong to the old category. old It also does not belong to the new category of target boxes. new The total number of pixels in the region.

[0085] According to one embodiment of the present invention, in S30, after separating the foreground and background information, it is still necessary to distill the valuable regions extracted from the foreground bounding box. To solve this problem, it is necessary to first calculate the difference between the semantic information generated by the teacher model after the decoder and the semantic information generated by the student model after the decoder, and select the regions with large differences as important feature locations that need to be focused on. Then, the important feature location information is added to the corresponding feature map in the form of a mask, and then feature distillation is performed at the corresponding locations to realize the imitation of the key features of the teacher model by the student model, and complete the semantic-guided distillation of continuous learning target detection. Specifically, the feature distillation loss is obtained by the following formula:

[0086]

[0087] In the formula, L feature (F S ,F T Characteristic distillation loss is ) This is the final feature vector of the teacher model. For the final feature vector of the student model, For the teacher model process feature vector, Mfg is the feature vector of the student model process. i,j Mbg is a mask for target information in the teacher model. i,j This serves as a mask for the correct background information in the teacher model, where α and β are the coefficients of the first and second hyperparameters, respectively. The new category target box r new The length and width of the vector are H and W, respectively, and i and j are the position coordinates of the pixel.

[0088] The student model focuses on the correct background information in the old category bounding boxes and the teacher model by different weights, while ignoring the incorrect background information in the teacher model.

[0089] According to one embodiment of the present invention, in S40 of the present invention, the loss after knowledge distillation is obtained by the following formula:

[0090] L′ S =L feature (F S ,F T )+L S ;

[0091] In the formula, L′ S For the loss due to knowledge distillation, L S The training loss for the student model.

[0092] According to one embodiment of the present invention, in S60 of the present invention, the updated student model can simultaneously detect both new category targets and old category targets.

[0093] To gain a further understanding of the present invention, the following description is provided in conjunction with... Figure 3 The present invention provides a detailed description of the continuous learning target detection method based on knowledge distillation.

[0094] In this embodiment, the task dataset consists of 80 target classes. The first task involves training, evaluating, and testing the first 20 target classes. The trained model is saved as a teacher model and used for knowledge distillation in subsequent tasks. The second task involves training and evaluating the 21st to 40th target classes, without providing any samples from Task 1. Guided by learning from the 21st to 40th target classes and the process feature vectors generated by the teacher model after Task 1 training, the model ultimately predicts both the old and new target classes (1-40) simultaneously in the test task. The trained model is saved as the teacher model for the first two tasks and used for knowledge distillation in subsequent Tasks 3 and 4. The implementation methods for Tasks 3 and 4 are the same as for Task 2 and will not be repeated here.

[0095] In Task 1, this embodiment trains a neural network to obtain a teacher model based on the training image dataset and old category labels, and saves the model's weights for knowledge transfer during subsequent task training.

[0096] In Task 2, this embodiment trains a student model based on the training image dataset and new category labels, while providing guidance through the process feature vectors generated by the previous teacher model.

[0097] like Figure 4 As shown, the present invention also provides a continuous learning target detection system based on knowledge distillation, the system comprising a network construction module, a knowledge distillation module, and an image target detection module;

[0098] The network building module is used to train the neural network based on the images and old category labels in the training dataset to obtain the teacher model, and to obtain the teacher model process feature vector during the training process. The teacher model process feature vector includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model.

[0099] The knowledge distillation module is used to train the neural network based on images, new category labels, and teacher model process feature vectors in the training dataset to obtain a student model and to obtain the student model training loss; it is used to obtain the feature distillation loss based on three pieces of information in the teacher model process feature vector and their corresponding masks; it is used to obtain the loss after knowledge distillation based on the student model training loss and the feature distillation loss; and it is also used to train the student model based on the loss after knowledge distillation to obtain an updated student model.

[0100] The image target detection module is used to store the updated student model and to perform target detection based on the updated student model.

[0101] This invention separates foreground features (target information in the teacher model) and background features (correct background information and incorrect background information in the teacher model), sets different parameters for features in different regions, reduces the transmission of erroneous information from the teacher model, and enables the student model to learn correct and useful knowledge. This reduces the catastrophic forgetting problem of old tasks when learning new tasks in successive learning. This invention improves the accuracy of the information transmitted by the teacher model, mitigates catastrophic and transient forgetting problems, reduces model computational resources and time, and increases target detection accuracy.

[0102] The present invention also provides a computer device, including a memory, a processor, and a knowledge distillation-based continuous learning target detection program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the knowledge distillation-based continuous learning target detection program.

[0103] In summary, this invention provides a continuous learning target detection method and system based on knowledge distillation, which has the following advantages compared with the prior art:

[0104] (1) It can be applied to target detection models with traditional convolutional networks and transformer structures, distilling the feature information after the encoder, and effectively distilling the interaction information between pixels while distilling the low-level detailed features (process feature vectors).

[0105] (2) This application distinguishes the feature information of the teacher model, clarifies the correct information and the incorrect information, and guides the feature distillation of valuable positions through semantic information, which alleviates the problems of catastrophic forgetting and instantaneous forgetting, enabling the student model to learn the existing knowledge of the teacher model more accurately, and reducing computational resources and time.

[0106] In specific deep learning model design, it is necessary to adaptively adjust the channel, height, and width through upsampling or adding auxiliary neural network modules during mathematical operations such as feature map summation and concatenation to conform to the mathematical constraints of matrix operations. However, the above solution omits these commonly used network structures and conventional operation methods in deep learning model construction. In actual model design, it may be necessary to add these conventional structures or implement these conventional operations to work together to achieve the normal operation of the solution in this application.

[0107] The parts of this invention not described in detail are techniques known to those skilled in the art.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A continuous learning target detection method based on knowledge distillation, characterized in that, The method includes: The neural network is trained based on the images and old category labels in the training dataset to obtain the teacher model, and the process feature vector of the teacher model is obtained during the training process. The process feature vector of the teacher model includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model. The neural network is trained based on the images, new category labels, and teacher model process feature vectors in the training dataset to obtain the student model, and the training loss of the student model is obtained. Feature distillation loss is obtained based on the masks corresponding to the process feature vector of the teacher model and the target information and correct background information in the teacher model. The loss after knowledge distillation is obtained based on the student model training loss and feature distillation loss; The student model is trained using the loss obtained through knowledge distillation to obtain an updated student model. Object detection is achieved using the updated student model; The characteristic distillation loss is obtained using the following formula: In the formula, Characteristic distillation loss, This is the final feature vector of the teacher model. For the final feature vector of the student model, For the teacher model process feature vector, For the student model process feature vector, This serves as a mask for the target information in the teacher model. This serves as a mask for the correct background information in the teacher model. These are the first and second hyperparameter coefficients, respectively. The target boxes for the new categories are respectively Length and width, ...

2. The method according to claim 1, characterized in that, The masks corresponding to the target information and the correct background information in the teacher model are obtained using the following formula: ； ； ； In the formula, This serves as a mask for the target information in the teacher model. The target boxes for the old categories are respectively Length and width, This serves as a mask for the correct background information in the teacher model. For all target boxes that do not belong to the old category It also does not belong to the new category of target boxes. The total number of pixels in the region.

3. The method according to claim 1 or 2, characterized in that, The loss after knowledge distillation is obtained using the following formula: ； In the formula, For the loss due to knowledge distillation, The training loss for the student model.

4. The method according to claim 1, characterized in that, The neural network is trained using images and old category labels from the training dataset to obtain a teacher model. The feature vectors of the teacher model during the training process are then obtained, including: The teacher model backbone network is used to extract deep image features and shallow image features of the teacher model from the images and old category labels in the training dataset. The deep image features and shallow image features of the teacher model are then concatenated to obtain a multi-scale feature map of the teacher model. The teacher model encoder obtains the process feature vector of the teacher model based on the multi-scale feature map of the teacher model. The teacher model decoder obtains the final feature vector of the teacher model based on the feature vector of the teacher model process. The teacher model classification subgrid obtains the teacher model classification soft label based on the final feature vector of the teacher model. The teacher model regression subgrid obtains the teacher model regression soft label based on the final feature vector of the teacher model; The teacher model activation function obtains the teacher model category label based on the teacher model classification soft label and the teacher model location label based on the teacher model regression soft label, thereby completing the training of the neural network and obtaining the teacher model.

5. The method according to claim 1, characterized in that, The neural network is trained based on images, new category labels, and feature vectors from the teacher model in the training dataset, resulting in the student model, which includes: The deep and shallow image features of the student model are extracted from the images and new category labels in the training dataset using the student model backbone network. The deep and shallow image features of the student model are then concatenated to obtain a multi-scale feature map of the student model. The student model encoder obtains the process feature vector of the student model based on the multi-scale feature map of the student model and the process feature vector of the teacher model. The student model decoder obtains the final feature vector of the student model based on the process feature vector of the student model. The student model classification subgrid obtains the soft label for student model classification based on the final feature vector of the student model. The student model regression subgrid obtains the student model regression soft label based on the final feature vector of the student model; The student model activation function obtains the student model category label based on the student model classification soft label and the student model location label based on the student model regression soft label, thereby completing the training of the neural network and obtaining the student model.

6. A continuous learning target detection system based on knowledge distillation, characterized in that, The system includes a network construction module, a knowledge distillation module, and an image target detection module; The network building module is used to train the neural network based on the images and old category labels in the training dataset to obtain the teacher model, and to obtain the teacher model process feature vector during the training process. The teacher model process feature vector includes the target information in the teacher model, the correct background information in the teacher model, and the incorrect background information in the teacher model. The knowledge distillation module is used to train the neural network based on images, new category labels, and teacher model process feature vectors in the training dataset to obtain a student model and to obtain the student model training loss; it is used to obtain the feature distillation loss based on the teacher model process feature vectors and the masks corresponding to the target information and correct background information in the teacher model; it is used to obtain the knowledge distillation loss based on the student model training loss and the feature distillation loss; and it is also used to train the student model based on the knowledge distillation loss to obtain an updated student model. The image target detection module is used to store the updated student model and to perform target detection based on the updated student model; The characteristic distillation loss is obtained using the following formula: In the formula, Characteristic distillation loss, This is the final feature vector of the teacher model. For the final feature vector of the student model, For the teacher model process feature vector, For the student model process feature vector, This serves as a mask for the target information in the teacher model. This serves as a mask for the correct background information in the teacher model. These are the first and second hyperparameter coefficients, respectively. The target boxes for the new categories are respectively Length and width, ...

7. A computer device, characterized in that, The method includes a memory, a processor, and a knowledge distillation-based continuous learning target detection program stored in the memory and executable on the processor, wherein the processor, when executing the knowledge distillation-based continuous learning target detection program, implements the method of any one of claims 1 to 5.

Images