Distillation method, device, equipment and medium of calling detection model
By distilling the small model in the intelligent driving assistance system, the intermediate layer characteristics and results of the teacher model are used, combined with soft labels and hard labels, the call detection accuracy of the small model is improved, and the contradiction between real-time and accuracy is solved.
Patent Information
- Application Number
- CN202411608032.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-02-25
AI Technical Summary
In the intelligent driving assistance system, how to improve the call detection accuracy of small models while ensuring real-time performance has become an urgent problem.
By entering the call hand image dataset into the student model, the classification teacher model and the segmented teacher model, the intermediate layer features and results of the teacher model are obtained, and combined with soft labels, hard labels and feature-level knowledge, the student model is distilled to improve its call detection accuracy.
It has achieved the accuracy of calling detection of student models while keeping the computing resources unchanged, and is close to or even reaching the level of teacher models, solving the problem of low detection accuracy of small models.
Smart Images
Figure CN119516520A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of intelligent driving and deep learning technology, and in particular to a distillation method, device, equipment and medium for a phone call detection model. Background Art
[0002] In intelligent driving assistance systems, accurately identifying whether the driver is using a mobile phone while driving is crucial to preventing traffic accidents.
[0003] In recent years, deep learning models have made significant progress in this field. However, on the one hand, high-performance models are often computationally intensive and difficult to run in real time on resource-constrained devices; on the other hand, although small models are more suitable for edge computing environments, their detection accuracy is usually lower than that of large models.
[0004] Therefore, how to improve the detection accuracy of small models while ensuring real-time performance has become an urgent problem to be solved. Summary of the invention
[0005] The present application provides a distillation method, device, equipment and medium for a phone call detection model, which is used to solve the problem in the prior art of how to improve the detection accuracy of a small model while ensuring real-time performance.
[0006] In a first aspect, the present application provides a distillation method for a phone call detection model, comprising:
[0007] Input the phone-making hand image dataset into the student model, the classification teacher model and the segmentation teacher model, wherein the classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on the hand area, the mobile phone area and the background area in the input image;
[0008] Obtaining the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model;
[0009] The classification result is used as a soft label, the phone call detection annotation information corresponding to the phone call hand image dataset is used as a hard label, and the classification intermediate layer features and the segmentation intermediate layer features are used as feature-level knowledge;
[0010] The student model is distilled according to the soft labels, the hard labels and the feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
[0011] In a second aspect, the present application provides a distillation device for a telephone detection model, comprising:
[0012] An input module is used to input a data set of images of a hand making a phone call into a student model, a classification teacher model and a segmentation teacher model, wherein the classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on the hand area, the mobile phone area and the background area in the input image;
[0013] An acquisition module, used to acquire the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model;
[0014] a label determination module, used to use the classification result as a soft label, use the phone call detection annotation information corresponding to the phone call hand image data set as a hard label, and use the classification intermediate layer feature and the segmentation intermediate layer feature as feature-level knowledge;
[0015] A distillation module is used to distill the student model according to the soft label, the hard label and the feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
[0016] In a third aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the distillation method of the phone call detection model as described in any one of the embodiments of the present application is implemented.
[0017] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a distillation method for a phone call detection model as described in any of the embodiments of the present application.
[0018] In a fifth aspect, the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the distillation method of the phone call detection model as described in any of the embodiments of the present application.
[0019] The distillation method, device, equipment and medium of the phone call detection model provided in the present application, in addition to allowing the student model to be supervised by the annotation information corresponding to the phone call hand image dataset, also uses the classification results of the phone call hand image dataset by the classification teacher model to supervise the training of the student model. At the same time, it also adopts the mechanism of multiple teacher models, and uses the respective intermediate layer features of the classification teacher model and the segmentation teacher model as additional supervision signals to supervise the learning of the intermediate layer of the student model, so that not only the output distribution of the student model is close to the classification teacher model, but also the student model can imitate the method of extracting features by the classification teacher model and the segmentation teacher model, thereby improving the student model's ability to recognize phone call behavior and improving the student model's phone call detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.
[0021] Figure 1 A flow chart of a distillation method for a phone call detection model provided in an embodiment of the present application;
[0022] Figure 2 A specific flow chart of distilling a student model in the distillation method of a phone call detection model provided in an embodiment of the present application;
[0023] Figure 3 A flow chart of another distillation method for a phone call detection model provided in an embodiment of the present application;
[0024] Figure 4 A structural schematic diagram of a distillation device for a phone call detection model provided in an embodiment of the present application;
[0025] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.
[0026] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION
[0027] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only the parts related to the present application, rather than all structures, are shown in the accompanying drawings.
[0028] It should be noted that similar numbers and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. The acquisition, storage, use, processing, etc. of data in the technical solution of this application comply with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. The purpose is only to illustrate the feasibility of the implementation of the technical solution of this application, but it does not mean that the applicant has or must use the relevant content of the solution.
[0029] Figure 1 A flowchart of a distillation method for a phone call detection model provided in an embodiment of the present application. This embodiment is applicable to the case where knowledge distillation of a driver phone call detection student model is performed using multiple teacher models. The method can be performed by a distillation device for a phone call detection model, which can be implemented in software and / or hardware, preferably configured in an electronic device, such as a computer device or a server. Figure 1 As shown, the method specifically includes:
[0030] S101. Input the phone-making hand image dataset into the student model, the classification teacher model and the segmentation teacher model.
[0031] Among them, the hand image dataset of making a phone call can be pre-acquired training data, and each image data includes a hand image of a driver who is making a phone call and holding a mobile phone. The classification teacher model and the segmentation teacher model are both pre-trained large models. The classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on the hand area, mobile phone area and background area in the input image. Compared with the classification teacher model, the segmentation teacher model can provide more delicate hand and mobile phone features. The student model is the model for phone call detection that currently needs to be distilled.
[0032] S102, obtaining the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model.
[0033] The data set is input into the student model, classification teacher model and segmentation teacher model. The input image data is subjected to feature extraction through the input layer and intermediate layer of each model, and the final prediction result is output through the output layer based on the extracted features.
[0034] The classification intermediate layer features of the classification teacher model are the features extracted from the intermediate layer of the network structure of the classification teacher model. The classification result is the detection result of the classification teacher model on whether the driver in the input image is making a phone call. For example, the classification teacher model can output the probability of predicting a phone call, and when the probability value exceeds a preset threshold, it can indicate that the driver in the image is making a phone call.
[0035] The segmentation middle layer features of the segmentation teacher model are the features extracted from the middle layer of the network structure of the segmentation teacher model. Since the goal of the segmentation teacher model is to perform image segmentation on three different areas of the input image, namely, the hand, mobile phone and background, compared with the classification teacher model, the middle layer of the segmentation teacher model extracts detailed features that can represent the hand, mobile phone and background. Using these features as supervision signals for the student model can enable the student model to learn richer representations in the image, especially the representations of different areas of the driver's hand, mobile phone and background, thereby helping to improve the student model's ability to recognize the driver's phone call behavior.
[0036] S103: Use the classification result as a soft label, use the phone call detection annotation information corresponding to the phone call hand image dataset as a hard label, and use the classification intermediate layer features and the segmentation intermediate layer features as feature-level knowledge.
[0037] S104. Distill the student model according to the soft labels, hard labels and feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
[0038] The process of distilling the student model through hard labels is the process of supervised training of the student model based on the phone call detection annotation information. The soft label is the classification result output by the pre-trained classification teacher model. The soft label is used to supervise the training of the student model. Through this knowledge distillation process, the student model can be guided to learn the decision boundary of the classification teacher model, thereby improving the classification accuracy of the student model. At the same time, the feature-level knowledge composed of the classification intermediate layer features and the segmentation intermediate layer features is used to guide the training and learning of the student model, so that the student model can learn richer representations, thereby improving the distillation effect of the student model as a whole.
[0039] It should be noted that in the task of detecting the driver making a phone call, simply relying on the hard label transfer knowledge of the output layer often cannot fully utilize all the information of the teacher model, resulting in limited performance improvement of the student model. The technical solution of the present application combines hard labels, soft labels and feature-level knowledge to supervise the distillation of the driver's phone call detection student model, that is, in addition to allowing the student model to be supervised by the annotation information corresponding to the phone call hand image data set, the classification results of the phone call hand image data set are also used by the classification teacher model to supervise the training of the student model. At the same time, a multi-teacher model mechanism is also adopted, and the intermediate layer features of the classification teacher model and the segmentation teacher model are used as additional supervision signals to supervise the learning of the intermediate layer of the student model, so that not only the output distribution of the student model is close to the classification teacher model, but also the student model can imitate the method of extracting features by the classification teacher model and the segmentation teacher model, thereby improving the student model's ability to recognize phone calls and improving the student model's phone call detection accuracy. In addition, the student model can be configured on the device side to maintain its lightweight characteristics and meet real-time application requirements.
[0040] Figure 2 This is a specific flow chart of distilling the student model in the distillation method of the phone call detection model provided in the embodiment of the present application. Figure 2 As shown, including:
[0041] S201, constructing an activation level loss function according to the soft labels and the detection results.
[0042] The soft label is the predicted probability distribution of the phone call output by the classification teacher model based on the input image. The activation level loss function can be, for example, the KL divergence, expressed as follows:
[0043]
[0044] Among them, D KL represents the KL divergence, P(x) represents the soft label output by the teacher model, Q(x) represents the predicted value / detection value output by the student model, and x represents the input hand image.
[0045] S202: Construct a supervision loss function based on the phone call detection annotation information and the detection results.
[0046] The phone call detection label information refers to the label true value of whether the driver is on the phone in the input image. The supervised loss function can be, for example, a cross entropy loss function, which is expressed as follows:
[0047]
[0048] Where H represents the cross entropy loss function; C is the total number of categories; y i,crepresents the Cth element in the true label vector of sample i. If sample i belongs to the Cth class, then y i,c =1, otherwise y i,c =0; p i,c It represents the probability that the model predicts that sample i belongs to the Cth class.
[0049] S203. Construct a first feature-level loss function based on the classification intermediate layer features and the student model intermediate layer features.
[0050] S204. Construct a second feature-level loss function based on the segmentation intermediate layer features and the student model intermediate layer features.
[0051] The intermediate layer features output by the intermediate layers of the classification teacher model and the segmentation teacher model are used as knowledge to supervise the intermediate layer of the student model for learning, so that the student model can more closely imitate the way the teacher model extracts features. The first feature level loss function and the second feature level loss function are, for example, L2 loss (mean square error) to minimize the feature difference between them and the intermediate layer features of the student model. The L2 loss is expressed as follows:
[0052]
[0053] Among them, L represents L2 loss, y i represents the intermediate layer features of the classification teacher model / segmentation teacher model, Represents the intermediate layer features of the student model.
[0054] S205, distilling the student model according to the activation level loss function, the supervision loss function, the first feature level loss function and the second feature level loss function.
[0055] Exemplarily, a total loss function can be constructed according to the activation level loss function, the supervision loss function, the first feature level loss function and the second feature level loss function, and the student model is distilled according to the total loss function. The constructed total loss function is used as the joint optimization target Total_loss to supervise the learning of the student model, and can be expressed as follows:
[0056] Total_loss=Hard_loss+α*Soft_loss+β*Feature_loss
[0057] Among them, Hard_loss represents supervision loss; Soft_loss represents activation-level loss; Feature_loss represents feature-level knowledge composed of the first feature-level loss function and the second feature-level loss function. For example, the first feature-level loss function and the second feature-level loss function are added together to form Feature_loss; α and β are hyperparameters used to control the proportion of Soft_loss and Feature_loss in Total_loss. The optimal combination can be found by trying different values, such as 0.1 and 0.13 respectively.
[0058] Figure 3 A flow chart of another distillation method for a phone call detection model provided in an embodiment of the present application. As shown in the figure, the hand image data of the driver holding a mobile phone to make a phone call is input into a pre-trained classification teacher model, a pre-trained segmentation teacher model and a student model, and during the distillation process, the network parameters of the classification teacher model and the segmentation teacher model are frozen. The pre-trained classification teacher model processes the input image and outputs the classification result through the normalized exponential function layer softmax (temperature coefficient T is t), such as the soft label soft label in the figure. The student model processes the input image and outputs the detection result through softmax (temperature coefficient T is t), such as the soft predict in the figure, and then calculates the activation level loss according to the soft label and soft predict, such as the Soft Loss in the figure. The hard predict of the detection result output by the student model through softmax (temperature coefficient T is 1) calculates the supervision loss with the hard label hard label, such as the Hard Loss in the figure. It should be noted here that hard predict is the original prediction result of the student model, and soft predict is the prediction result corresponding to the temperature coefficient T of t, so that the probability distribution of the output soft predict can be flattened, so as to better capture the relationship between different categories to be predicted by the student model, thereby improving the learning effect of the student model on the soft label supervision signal. In addition, the classification intermediate layer features of the classification teacher model and the segmentation intermediate layer features of the segmentation teacher model are respectively combined with the student model intermediate layer features of the student model to construct the first feature-level loss function Feature Loss A and the second feature-level loss function Feature Loss B. Finally, the total loss function is constructed based on Soft Loss, Hard Loss, Feature Loss A, and Feature Loss B, and the student model is distilled based on the total loss function.
[0059] The student model obtained by distillation can be used as a phone call detection model to detect whether the driver is making a phone call. During implementation, the driver's image can be obtained in real time. For example, the driver's hand image can be obtained using the yolov6 detection model. Then, the driver's hand image is input into the phone call detection model for processing, and the phone call detection model directly outputs the prediction result of whether to make a phone call. For example, the output result is a probability value between 0 and 1, which indicates the probability that the current driver is making a phone call. The probability value is compared with a pre-configured threshold, and whether the driver is making a phone call can be determined based on the comparison result.
[0060] Experimental results show that the technical solution of the embodiment of the present application can significantly improve the performance of the student model in the driver's phone call detection task while keeping the computing resource consumption unchanged, approaching or even reaching the level of the teacher model, which fully proves the effectiveness and practicality of the solution of the present application.
[0061] Figure 4 A schematic diagram of the structure of a distillation device for a phone call detection model provided in an embodiment of the present application.
[0062] like Figure 4 As shown, the device 40 includes:
[0063] An input module 410 is used to input a data set of images of a hand making a phone call into a student model, a classification teacher model and a segmentation teacher model, wherein the classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on a hand area, a mobile phone area and a background area in the input image;
[0064] An acquisition module 420, used to acquire the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model;
[0065] A label determination module 430 is used to use the classification result as a soft label, the phone call detection annotation information corresponding to the phone call hand image data set as a hard label, and the classification intermediate layer feature and the segmentation intermediate layer feature as feature-level knowledge;
[0066] The distillation module 440 is used to distill the student model according to the soft label, the hard label and the feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
[0067] In some embodiments, the acquisition module 420 is further used to: acquire the student model intermediate layer features and detection results of the student model;
[0068] Accordingly, the distillation module 440 includes:
[0069] A first construction unit, configured to construct an activation level loss function according to the soft label and the detection result;
[0070] A second construction unit is used to construct a supervision loss function according to the phone call detection annotation information and the detection result;
[0071] A third construction unit is used to construct a first feature-level loss function according to the classification intermediate layer features and the student model intermediate layer features;
[0072] A fourth construction unit, used to construct a second feature-level loss function according to the segmentation intermediate layer features and the student model intermediate layer features;
[0073] A distillation unit is used to distill the student model according to the activation level loss function, the supervision loss function, the first feature level loss function and the second feature level loss function.
[0074] In some embodiments, the distillation unit is specifically used for:
[0075] A total loss function is constructed according to the activation-level loss function, the supervision loss function, the first feature-level loss function, and the second feature-level loss function, and the student model is distilled according to the total loss function.
[0076] In some embodiments, the student model is configured on the device side.
[0077] In some embodiments, the classification teacher model and the segmentation teacher model are both large models.
[0078] The distillation device of the phone call detection model provided in the embodiment of the present application can be used to implement the technical solution of the distillation method of the phone call detection model in the above embodiment. Its implementation principle and technical effect are similar and will not be repeated here.
[0079] It should be noted that it should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the distillation module 440 can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The function of the above distillation module 440. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in software form.
[0080] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 5 As shown, the electronic device may include: a transceiver 121 , a processor 122 , and a memory 123 .
[0081] The processor 122 executes the computer execution instructions stored in the memory, so that the processor 122 executes the scheme in the above embodiment. The processor 122 can be a general-purpose processor, including a central processing unit CPU, a network processor (network processor, NP), etc.; it can also be a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
[0082] The memory 123 is connected to the processor 122 via a system bus and completes communication between them. The memory 123 is used to store computer program instructions.
[0083] The transceiver 121 may be used to obtain tasks to be executed and configuration information of the tasks to be executed.
[0084] The system bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The transceiver is used to realize the communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory.
[0085] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.
[0086] An embodiment of the present application also provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the distillation method of the phone detection model in the above embodiment.
[0087] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, it can implement the technical solution of the distillation method of the phone call detection model in the above embodiment.
[0088] In the process of implementation, the computer program product can be written in one or more programming languages or a combination thereof to perform the computer program code of the present application, and the programming language includes an object-oriented programming language, such as Java, Smalltalk, C++, and also includes a conventional procedural programming language, such as "C" language or similar programming language. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).
[0089] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.
Claims
1. A distillation method for a phone call detection model, characterized in that: include: Input the phone-making hand image dataset into the student model, the classification teacher model and the segmentation teacher model, wherein the classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on the hand area, the mobile phone area and the background area in the input image; Obtaining the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model; The classification result is used as a soft label, the phone call detection annotation information corresponding to the phone call hand image dataset is used as a hard label, and the classification intermediate layer features and the segmentation intermediate layer features are used as feature-level knowledge; The student model is distilled according to the soft labels, the hard labels and the feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
2. The method according to claim 1, characterized in that Also includes: Obtain student model intermediate layer features and detection results of the student model; Accordingly, the distilling the student model according to the soft label, the hard label and the feature-level knowledge includes: Constructing an activation level loss function according to the soft label and the detection result; Constructing a supervision loss function according to the phone call detection annotation information and the detection result; Constructing a first feature-level loss function according to the classification intermediate layer features and the student model intermediate layer features; Constructing a second feature-level loss function according to the segmentation intermediate layer features and the student model intermediate layer features; The student model is distilled according to the activation-level loss function, the supervision loss function, the first feature-level loss function, and the second feature-level loss function.
3. The method according to claim 2, characterized in that The distilling the student model according to the activation level loss function, the supervision loss function, the first feature level loss function and the second feature level loss function comprises: A total loss function is constructed according to the activation-level loss function, the supervision loss function, the first feature-level loss function, and the second feature-level loss function, and the student model is distilled according to the total loss function.
4. The method according to claim 1, characterized in that: The student model is configured on the device side.
5. The method according to claim 1, characterized in that The classification teacher model and the segmentation teacher model are both large models.
6. A distillation device for a telephone detection model, characterized in that: include: An input module is used to input a data set of images of a hand making a phone call into a student model, a classification teacher model and a segmentation teacher model, wherein the classification teacher model is used to classify and detect whether the target in the input image is making a phone call, and the segmentation teacher model is used to perform image segmentation on the hand area, the mobile phone area and the background area in the input image; An acquisition module, used to acquire the classification intermediate layer features and classification results of the classification teacher model, and the segmentation intermediate layer features of the segmentation teacher model; a label determination module, used to use the classification result as a soft label, use the phone call detection annotation information corresponding to the phone call hand image data set as a hard label, and use the classification intermediate layer feature and the segmentation intermediate layer feature as feature-level knowledge; A distillation module is used to distill the student model according to the soft label, the hard label and the feature-level knowledge, wherein the distilled student model is used to perform phone detection on the target in the image.
7. The device according to claim 6, characterized in that The acquisition module is also used to: acquire the student model intermediate layer features and detection results of the student model; Accordingly, the distillation module comprises: A first construction unit, configured to construct an activation level loss function according to the soft label and the detection result; A second construction unit is used to construct a supervision loss function according to the phone call detection annotation information and the detection result; A third construction unit is used to construct a first feature-level loss function according to the classification intermediate layer features and the student model intermediate layer features; A fourth construction unit, used to construct a second feature-level loss function according to the segmentation intermediate layer features and the student model intermediate layer features; A distillation unit is used to distill the student model according to the activation level loss function, the supervision loss function, the first feature level loss function and the second feature level loss function.
8. The device according to claim 7, characterized in that The distillation unit is specifically used for: A total loss function is constructed according to the activation-level loss function, the supervision loss function, the first feature-level loss function, and the second feature-level loss function, and the student model is distilled according to the total loss function.
9. The device according to claim 6, characterized in that The student model is configured on the device side.
10. The device according to claim 6, characterized in that The classification teacher model and the segmentation teacher model are both large models.
11. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the distillation method of the phone call detection model as described in any one of claims 1-5.
12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the distillation method of the phone call detection model as described in any one of claims 1-5.
13. A computer program product, characterized in that It comprises a computer program, which, when executed by a processor, implements the distillation method of the phone call detection model described in any one of claims 1-5.