WO2020186887A1

WO2020186887A1 - Target detection method, device and apparatus for continuous small sample images

Info

Publication number: WO2020186887A1
Application number: PCT/CN2019/130818
Authority: WO
Inventors: 乔宇; 陈贤煜; 王亚立
Original assignee: 深圳先进技术研究院
Priority date: 2019-03-15
Filing date: 2019-12-31
Publication date: 2020-09-24
Also published as: CN110033026B; CN110033026A

Abstract

Disclosed is a target detection method for continuous small sample images, comprising: decoupling a target detector by means of decoupling a target-background classifier into a background classifier and a target classifier; obtaining a single-stage target detector after parameters are initialized and decoupled according to the weight rubbing strategy by using the previous detection task framework; obtaining supervision parameters prior to the current detection task by a knowledge distillation method, and training or detecting the target by the supervision parameters via the initialized and decoupled target detector. The learning complexity is favorably reduced by decoupling; the training requirements for continuous small sample images can be effectively adapted by the initialization of rubbing strategy; the knowledge of previous detection tasks can be effectively utilized by means of knowledge distillation so as to effectively prevent old knowledge from being forgotten and to achieve a balance between the forgetting of old knowledge and the acquisition of new knowledge.

Description

Target detection method, device and equipment for continuous small sample images

Technical field

This application belongs to the field of image processing, and in particular relates to a target detection method, device and equipment for continuous small sample images.

Background technique

Through deep learning to detect samples, it can effectively complete the target detection and recognition of the image. Moreover, with the development of in-depth detection technology, target detection mainly includes two detection frameworks, namely, a two-step detection framework and a single-step detection framework. These frameworks have achieved significant results in major international public data sets.

However, depth detection is generally driven by a large-scale data set with complete bounding box calibration. In the actual image target detection process, collecting such data requires a lot of manpower and material resources. When training or detecting continuous small samples, new tasks usually continue to come. These tasks will face new detection tasks or new object types. Each target detection task only has a small sample of bounding box labels, using random The modernized detection architecture will easily lead to serious over-fitting problems, and may suffer serious forgetting when focusing on the current detection task, and cannot effectively maintain the performance of the previous task.

technical problem

In view of this, the embodiments of the present application provide a target detection method, device and equipment for continuous small-sample images to solve the problem of over-fitting and may suffer from the detection of continuous small-sample images in the prior art. The problem of serious forgetting.

Technical solutions

The first aspect of the embodiments of the present application provides a target detection method for continuous small sample images, and the target detection method for small sample images includes:

Decouple the target detector, and decouple the target-background classifier into a background classifier and a target classifier;

According to the weighted rubbing strategy and using the previous detection task framework to obtain the single-order target detector after parameter initialization and decoupling;

The supervised parameters before the current detection task are obtained by the knowledge distillation method, and the target is trained or detected by combining the supervised parameters with the decoupled target detector after initialization.

With reference to the first aspect, in the first possible implementation of the first aspect, the step of decoupling the target detector and decoupling the target-background classifier into a background classifier and a target classifier includes:

The single-stage target detector is decoupled, and the target-background classifier is decoupled into a background classifier and a target classifier. The score vector of the background classifier is obtained as:

The score vector of the target classifier is:

among them,

Is the prediction vector before the background classifier is processed by the normalized exponential function, p _obj /p _bg is the probability of the target/background,

Is the prediction vector before the target classifier is processed by the normalized exponential function, and q _i is the probability of belonging to the i-th target i=1, 2, ..., C.

With reference to the first aspect, in a second possible implementation manner of the first aspect, the steps of the weighted rubbing strategy and the use of a previous detection task framework to obtain a single-order target detector after parameter initialization and decoupling include:

Obtain and detect the target category in the task;

According to the target category, obtain a plurality of a priori boxes containing the target category in the source domain;

Respectively calculating the prediction vector of the target category in the detection task according to the acquired a priori boxes in the multiple source domains;

The multiple prediction vectors are summed and averaged, and then normalized to obtain the parameter vector corresponding to the target category.

In combination with the second possible implementation manner of the first aspect, in the third possible implementation manner of the first aspect, when the detection task includes multiple target categories, according to the parameter vectors corresponding to the multiple target categories, obtain The initialization parameter matrix of the target classifier.

With reference to the first aspect, in a fourth possible implementation of the first aspect, the steps of the weighted rubbing strategy and the use of the previous detection task framework to obtain a single-order target detector after parameter initialization and decoupling include:

Initialize the backbone, bounding box regression and background classifiers of the current detection task according to the backbone parameters, bounding box regression parameters, and background classifier parameters used in the detection task T _k-1 before the current detection task T _k .

In combination with the first aspect, in a fifth possible implementation manner of the first aspect, the supervised parameter pair before the current detection task is obtained by the knowledge distillation method, and the supervised parameter pair is combined with the decoupled single-stage target detector after initialization. The steps to train the target include:

Calculate the total loss function of the current target detection task T _k according to the formula:

Standardize the target detector of the current detection task according to the total loss function, where the main loss

Including bounding box regression loss, background classifier loss, target plus background classifier loss, send the image of the current detection task to the target detector before the current detection task to obtain the supervision parameters

λ is the weight coefficient.

A second aspect of the embodiments of the present application provides a target detection device for continuous small sample images, and the target detection device for continuous small sample images includes:

The decoupling unit is used to decouple the target detector and decouple the target-background classifier into a background classifier and a target classifier;

The rubbing unit is used for rubbing the strategy according to the weight and using the previous detection task framework to obtain the single-stage target detector after parameter initialization and decoupling;

The distillation unit is used to obtain the supervision parameters before the current detection task through the knowledge distillation method, and train or detect the target through the decoupled target detector after initialization and the supervision parameters.

With reference to the second aspect, in the first possible implementation manner of the second aspect, the decoupling unit is configured to:

The score vector of the target classifier is:

among them,

The third aspect of the embodiments of the present application provides a target detection device for continuous small-sample images, including a memory, a processor, and a computer program stored in the memory and running on the processor. The processor When the computer program is executed, the steps of the target detection method for continuous small-sample images as described in any one of the first aspect are implemented.

The fourth aspect of the embodiments of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the continuous Steps of the target detection method for small sample images.

Beneficial effect

Compared with the prior art, the embodiment of the present application has the beneficial effect that by decoupling the target-background classifier into a background classifier and a target classifier, the decoupled classifier can use a weighted rubbing strategy or prior detection Task parameter initialization is beneficial to reduce the complexity of learning, and through the initialization of the rubbing strategy, it can effectively adapt to the training requirements of continuous small sample images. The knowledge distillation method can effectively use the knowledge of the previous detection task, which can effectively combat the old The forgetting of knowledge realizes the balance between the forgetting of old knowledge and the acquisition of new knowledge.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only of the present application. For some embodiments, for those of ordinary skill in the art, other drawings can be obtained from these drawings without creative labor.

FIG. 1 is a schematic diagram of the implementation process of a target detection method for continuous small sample images provided by an embodiment of the present application;

2 is a schematic diagram of the original framework of a single-stage detector provided by an embodiment of the present application;

3 is a schematic diagram of the framework of a decoupled single-stage detector provided by an embodiment of the present application;

4 is a schematic diagram of a target detection device for continuous small sample images provided by an embodiment of the present application;

FIG. 5 is a schematic diagram of an image detection device provided by an embodiment of the present application.

Embodiments of the invention

In the following description, for the purpose of illustration rather than limitation, specific details such as a specific system structure and technology are proposed for a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from obstructing the description of this application.

In order to illustrate the technical solutions described in the present application, specific embodiments are used for description below.

Figure 1 is a schematic diagram of the implementation process of a method for target detection of continuous small sample images provided by an embodiment of the application, and the details are as follows:

In step S101, the target detector is decoupled, and the target-background classifier is decoupled into a background classifier and a target classifier;

Specifically, the decoupling refers to decoupling the modules in the target detector into a series of sub-modules, thereby reducing the complexity of the task structure for continuous small sample detection. The target detector can select a single-stage target detector, so that the target can be detected efficiently and accurately. As shown in Figure 2, a curved single-stage target detector includes a backbone, bounding box regression, and target-background classifier.

For a new detection task, since the backbone and bounding box regression are shared by different target categories, they can be obtained from the initialization of the previous detection task. Conversely, the target-background classifier must be initialized randomly, because the new target category may not have been seen in previous tasks. However, this fact exacerbates the difficulty of the task of learning the target-background classifier, especially when the new task is a small sample. In order to reduce the training burden, this application proposes to decouple the target-background classifier into a background classifier and a target classifier, as shown in Figure 3. Suppose the score vector of the background classifier is

among them

Is the prediction vector before the background classifier is processed by the normalized exponential function softmax, and p _obj /p _bg is the probability of the target/background. The score of the target classifier is

among them

Is the prediction vector before the target classifier is processed by the normalized exponential function softmax, and q _i is the probability of belonging to the i-th target i=1, 2, ..., C. After that, we can directly calculate the target-background probability score vector, namely

v=[p _obj q ₁ , p _obj q ₂ ,..., p _obj q _c , p _bg ] (3)

Through decoupling, we only need to initialize the target and background classifiers separately instead of initializing a target-background classifier. For a new detection task, the background classifier can be initialized from the previous task. This is because the binary classifier can share parameters among different targets. The target classifier is a multi-classifier, and the rubbing method can be proposed in step S102 to solve the random initialization problem caused by the arrival of a new task.

Through decoupling operation, the detection architecture can be decomposed into the backbone, bounding box regression, target classifier and background classifier. When new tasks come, you can reduce learning difficulties.

In step S102, a single-order target detector after parameter initialization and decoupling is obtained according to the weighted rubbing strategy and the previous detection task framework;

The weighted rubbing strategy refers to selecting the public data set as the source domain, searching for the a priori box corresponding to the target category in the detection task in the source domain, and calculating the prediction vector of the target in the detection task according to the a priori box. After the prediction vector is summed and averaged, it is normalized to obtain the parameter vector corresponding to the target category. After obtaining the parameter vector, the target classifier in the target detector can be initialized according to the parameter vector obtained by the weight rubbing strategy.

Since each detection task is a small sample, a good sub-model initialization scheme will be very important. This application proposes a simple but effective solution. First, this application can use an existing large-scale international public data set (such as COCO) as the source domain, and pre-train the decoupling architecture shown in FIG. 3. Due to the backbone, bounding box regression and background classifiers can be shared among different target categories. Accordingly, the detection task T _k-1 may be repeated to initialize the task of detection of T _k, where T ₀ = S is the source task.

For the initialization of the target classifier. Since the target types of the new detection task and the previous detection task are different, when the new task arrives, the classifier needs to be able to adapt to the new task. In order to avoid random initialization, this application can adopt a weight rubbing strategy to re-adjust the classifier to meet the requirements of target detection settings.

Suppose there are C _S types of source tasks, and T _k detection tasks have

Categories. The target classifier of the kth continuous small sample detection task focuses on finding a parameter matrix

It is used to build the mapping of the source domain prediction vector (before the normalized exponential function is processed by softmax) to the target domain prediction vector. In the next step, we can use softmax on the prediction vector of the target domain to get the final classification result.

Assume that there is a category of target A in the T _kth detection task. First, the photos containing target A can be sent to the single-stage detector backbone, and then reach the target classifier in the source domain. For the n-th prior box containing target A, the prediction vector (before softmax) can be obtained

And C _S is the number of target types in the source domain. Then, all selected a priori boxes with a higher confidence (greater than a certain threshold) of the a priori box about the target A, the confidence is determined by the intersection of the a priori box and the true bounding box (intersection of union, IoU ) Decide. We use L ₂ to normalize the prediction vectors of the source domain, and average these normalized prediction vectors, assuming there are N vectors in total. We further normalize this average vector and use it as a parameter vector about target A:

It can be found, for the first goal on A T _k detectors task, W _A summary of the rich and relevant

Representation information in the source domain. It can be regarded as a prototype of objective A and provides a reliable parameter initialization scheme. Finally, we use a similar scheme to perform similar operations on other categories on the T _k- th detection task to obtain the initialization parameter matrix

As the target classifier. Give a picture of the query,

The prediction vector of the a priori box from the source domain can be converted into a vector on the T _k- th detection task domain.

Then it is sent to the normalized exponential function softmax to achieve classification.

In summary, rubbing allows all the decoupled sub-models to be deployed quickly, which can effectively slow down the overfitting phenomenon under the task of continuous detection with small samples.

In step S103, the supervised parameter before the current detection task is obtained by the knowledge distillation method, and the target is trained or detected by combining the supervised parameter with the decoupled target detector after initialization.

After rubbing our decoupling architecture, we start to train the T _kth target detection task. However, focusing only on the current task will lead to catastrophic forgetting, that is, the detection framework will gradually forget the knowledge learned in the previous tasks, so the detection performance of these tasks will be greatly degraded. In order to avoid this phenomenon, when training the current task, we use the distillation method as a regularization term. Specifically, we record the total loss function of the T _kth target detection task as

Main loss

Including bounding box regression (smooth L ₁ ) loss, background classifier (mutual entropy) loss, target plus background classifier (mutual entropy) loss. It can be used to train the detection architecture of the T _kth target detection task.

In order to alleviate forgetting, the method of knowledge distillation can be used, that is, the current detection task is sent to the previous target detector to obtain knowledge parameters including prediction boxes and item score vectors. In particular, we send the image of the T _kth target detection task into the T _k- 1th target detection frame. Then, we can obtain knowledge of the items up to T _k-1 from the previous task (for example, the prediction box and the item score vector). We use this knowledge as additional supervision (e.g.

), they can effectively normalize the corresponding sub-models of the T _k- th detection framework to retain the previously learned knowledge.

Therefore, the distillation method adaptively balances the forgetting of knowledge and the update of new tasks. Therefore, our detection architecture can maintain the previously learned knowledge while acquiring new knowledge, thereby achieving lifelong learning.

Through the target detection method of continuous small sample images described in this application, new tasks with small samples can be quickly deployed, and the performance of previously learned tasks can be effectively maintained.

It should be understood that the size of the sequence number of each step in the foregoing embodiment does not mean the order of execution. The execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiment of the present application.

FIG. 4 is a schematic structural diagram of a target detection device for continuous small sample images provided by an embodiment of the application. The target detection device for continuous small sample images includes:

The decoupling unit 401 is used to decouple the target detector, and decouple the target-background classifier into a background classifier and a target classifier;

The rubbing unit 402 is used for rubbing the strategy according to the weight and using the previous detection task framework to obtain the single-stage target detector after parameter initialization and decoupling;

The distillation unit 403 is configured to obtain the supervision parameters before the current detection task through the knowledge distillation method, and train or detect the target through the decoupled target detector after initialization and the supervision parameters.

Preferably, the decoupling unit is used for:

The score vector of the target classifier is:

among them,

The target detection device for continuous small sample images described in FIG. 4 corresponds to the target detection method for continuous small sample images described in FIG. 1.

FIG. 5 is a schematic diagram of a target detection device for continuous small sample images provided by an embodiment of the present application. As shown in FIG. 5, the target detection device 5 for continuous small sample images of this embodiment includes a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and running on the processor 50, For example, the target detection program of continuous small sample images. When the processor 50 executes the computer program 52, the steps in the foregoing embodiments of the target detection method for each continuous small sample image are implemented. Alternatively, when the processor 50 executes the computer program 52, the functions of the modules/units in the foregoing device embodiments are realized.

Exemplarily, the computer program 52 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 51 and executed by the processor 50 to complete This application. The one or more modules/units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 52 in the target detection device 5 of the continuous small sample image . For example, the computer program 52 can be divided into:

The target detection device 5 of the continuous small sample image may be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The target detection device of the continuous small sample image may include, but is not limited to, a processor 50 and a memory 51. Those skilled in the art can understand that FIG. 5 is only an example of the target detection device 5 for continuous small sample images, and does not constitute a limitation on the target detection device 5 for continuous small sample images, and may include more or less than that shown in the figure. Components, or a combination of some components, or different components, for example, the target detection device of the continuous small sample image may also include input and output devices, network access devices, buses, and the like.

The so-called processor 50 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.

The memory 51 may be an internal storage unit of the target detection device 5 for continuous small sample images, such as a hard disk or a memory of the target detection device 5 for continuous small sample images. The memory 51 may also be an external storage device of the target detection device 5 of the continuous small sample image, for example, a plug-in hard disk equipped on the target detection device 5 of the continuous small sample image, or a smart memory card (SmartMedia Card). ,SMC), Secure Digital (SD) card, Flash Card, etc. Further, the memory 51 may also include both the internal storage unit of the target detection device 5 of the continuous small sample image and an external storage device. The memory 51 is used to store the computer program and other programs and data required by the target detection device of the continuous small sample image. The memory 51 can also be used to temporarily store data that has been output or will be output.

Those skilled in the art can clearly understand that for the convenience and conciseness of description, only the division of the above-mentioned functional units and modules is used as an example. In practical applications, the above-mentioned functions can be allocated to different functional units and modules as required. Module completion means dividing the internal structure of the device into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist alone physically, or two or more units can be integrated into one unit. The above-mentioned integrated units can be hardware-based Formal realization can also be realized in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the protection scope of this application. For the specific working process of the units and modules in the foregoing system, reference may be made to the corresponding process in the foregoing method embodiment, which is not repeated here.

In the foregoing embodiments, the description of each embodiment has its own focus. For parts that are not detailed or recorded in a certain embodiment, reference may be made to related descriptions of other embodiments.

A person of ordinary skill in the art may be aware that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered beyond the scope of this application.

In the embodiments provided in this application, it should be understood that the disclosed apparatus/terminal device and method may be implemented in other ways. For example, the device/terminal device embodiments described above are only illustrative. For example, the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units. Or components can be combined or integrated into another system, or some features can be omitted or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

In addition, the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.

If the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, this application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the program is executed by the processor, the steps of the foregoing method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electrical carrier signal, telecommunications signal, and software distribution media. It should be noted that the content contained in the computer-readable medium can be appropriately added or deleted in accordance with the requirements of the legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to the legislation and patent practice, the computer-readable medium Does not include electrical carrier signals and telecommunication signals.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still implement the foregoing The technical solutions recorded in the examples are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application, and should be included in Within the scope of protection of this application.

Claims

A target detection method for continuous small sample images is characterized in that the target detection method for small sample images includes:

Decouple the target detector, and decouple the target-background classifier into a background classifier and a target classifier;

According to the weighted rubbing strategy and using the previous detection task framework to obtain the single-order target detector after parameter initialization and decoupling;

The supervised parameters before the current detection task are obtained by the knowledge distillation method, and the target is trained or detected by combining the supervised parameters with the decoupled target detector after initialization.
The target detection method for continuous small sample images according to claim 1, wherein the step of decoupling the target detector and decoupling the target-background classifier into a background classifier and a target classifier comprises:

The single-stage target detector is decoupled, and the target-background classifier is decoupled into a background classifier and a target classifier. The score vector of the background classifier is obtained as:

The score vector of the target classifier is:

among them,
Is the prediction vector before the background classifier is processed by the normalized exponential function, p obj /p bg is the probability of the target/background,
Is the prediction vector before the target classifier is processed by the normalized exponential function, and q i is the probability of belonging to the i-th target i=1, 2, ..., C.
The target detection method for continuous small-sample images according to claim 1, wherein the step of using a weighted rubbing strategy and using a previous detection task framework to obtain a single-stage target detector after parameter initialization and decoupling comprises:

Obtain and detect the target category in the task;

According to the target category, obtain a plurality of a priori boxes containing the target category in the source domain;

Respectively calculating the prediction vector of the target category in the detection task according to the acquired a priori boxes in the multiple source domains;

The multiple prediction vectors are summed and averaged and then normalized to obtain a parameter vector corresponding to the target category.
The target detection method for continuous small-sample images according to claim 3, wherein when the detection task includes multiple target categories, the initialization of the target classifier is obtained according to the parameter vectors corresponding to the multiple target categories Parameter matrix.
The target detection method for continuous small-sample images according to claim 1, wherein the step of using a weighted rubbing strategy and using a previous detection task framework to obtain a single-stage target detector after parameter initialization and decoupling comprises:

Initialize the backbone, bounding box regression and background classifiers of the current detection task according to the backbone parameters, bounding box regression parameters, and background classifier parameters used in the detection task T k-1 before the current detection task T k .
The target detection method for continuous small-sample images according to claim 1, wherein the supervised parameter before the current detection task is obtained by the knowledge distillation method, and the decoupled single-stage target detector after initialization is combined with the supervised The steps of parameter training on the target include:

Calculate the total loss function of the current target detection task T k according to the formula:

Standardize the target detector of the current detection task according to the total loss function, where the main loss
Including bounding box regression loss, background classifier loss, target plus background classifier loss, send the image of the current detection task to the target detector before the current detection task to obtain the supervision parameters
λ is the weight coefficient.
A target detection device for continuous small-sample images is characterized in that the target detection device for continuous small-sample images includes:

The decoupling unit is used to decouple the target detector and decouple the target-background classifier into a background classifier and a target classifier;

The rubbing unit is used for rubbing the strategy according to the weight and using the previous detection task framework to obtain the single-stage target detector after parameter initialization and decoupling;

The distillation unit is used to obtain the supervision parameters before the current detection task through the knowledge distillation method, and train or detect the target through the decoupled target detector after initialization and the supervision parameters.
The target detection device for continuous small-sample images according to claim 7, wherein the decoupling unit is used for:

The single-stage target detector is decoupled, and the target-background classifier is decoupled into a background classifier and a target classifier. The score vector of the background classifier is obtained as:

The score vector of the target classifier is:

among them,
Is the prediction vector before the background classifier is processed by the normalized exponential function, p obj /p bg is the probability of the target/background,
Is the prediction vector before the target classifier is processed by the normalized exponential function, and q i is the probability of belonging to the i-th target i=1, 2, ..., C.
A target detection device for continuous small sample images, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program when the computer program is executed. The steps of realizing the target detection method of continuous small sample images according to any one of claims 1 to 6.
A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the continuous small sample image according to any one of claims 1 to 6 is realized The steps of the target detection method.