CN116342906A - Cross-domain small sample image recognition method and system - Google Patents

Abstract

The invention provides a cross-domain small sample image recognition method and a system, comprising the following steps: determining a trained small sample recognition network; the small sample recognition network is used for carrying out matching classification on samples of different categories; the training process of the small sample recognition network requires an image generation network, and the image generation network comprises: the system comprises a variable self-encoder module and a style conversion module, wherein the variable self-encoder module is used for extracting and reconstructing characteristic distribution of a target domain sample, and the variable self-encoder comprises intermediate parameters which are used for gradient rising to reconstruct the target domain sample into a sample with a more complex style; the style conversion module is used for stylizing the source domain samples to obtain labeled training samples belonging to the target domain data distribution; and inputting the target domain sample to be identified into a trained small sample identification network to conduct prediction classification, so as to obtain an image identification result. According to the method, the target domain information is effectively introduced in the training stage, and the generalization capability of the model to the target domain is improved.

Description

A cross-domain small-sample image recognition method and system

technical field

The invention belongs to the field of computer vision and image processing, and more specifically relates to a cross-domain small-sample image recognition method and system.

Background technique

Currently, deep learning-based methods have achieved excellent performance on many computer vision tasks. However, these powerful deep learning models often rely on large-scale data sets and long-term training processes, requiring more human resources, time costs, and expensive computing costs. In this case, the problem of small-sample image recognition has become one of the key research directions in the field of deep learning. Compared with the traditional image recognition driven by a large number of samples, small sample image recognition can achieve accurate prediction and classification of new categories of samples under the condition of limited sample size, which is more in line with real application scenarios. However, in more practical applications, it is difficult to collect samples from the same domain to complete a large number of small sample classification tasks. When the distribution of training and testing data is inconsistent, that is, the distribution of source domain data is different from that of target domain data. There is a domain shift between them, and the model is more difficult to generalize. In order to distinguish it from the problem setting of traditional small-sample learning, the small-sample learning problem in this scenario is defined as "cross-domain small-sample learning".

The previous method is based on model fine-tuning. This type of method is based on the idea of meta-learning "learning to learn", which aims to enable the model to learn task-independent knowledge during training, so that the model does not overfit the training data and can generalize to new tasks faster. For example, the article "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, Proceedings of the International Conference on Machine Learning (ICML), 2017." proposes the idea of meta-learning to effectively solve the problem of cross-domain small sample classification.

The existing fine-tuning-based cross-domain small-sample recognition has obvious limitations. For each input of a new task, the operation of fine-tuning the model is very complicated, especially for large-scale models, parameter fine-tuning will consume a lot of calculations cost and time, therefore, the above technique is not an essential solution.

Contents of the invention

Aiming at the defects of the prior art, the purpose of the present invention is to provide a cross-domain small-sample image recognition method and system, aiming at solving the problem that the existing cross-domain small-sample recognition technology based on fine-tuning needs to be performed once for each new input task. The complexity of parameter fine-tuning operations; and the existing small sample recognition methods cannot solve the problem of domain gap and poor generalization performance.

In order to achieve the above object, in the first aspect, the present invention provides a cross-domain small-sample image recognition method, comprising the following steps:

Determine the trained small-sample recognition network; the small-sample recognition network is used to match and classify samples of different categories; the training process of the small-sample recognition network needs to use an image generation network, and the image generation network includes: variable Divided into a self-encoder module and a style conversion module, the variational self-encoder module is used to extract and reconstruct the feature distribution of the target domain samples, and the variational self-encoder includes intermediate parameters for gradient ascent to reconstruct the target domain samples into a sample with a more complex style; the style conversion module is used to stylize the source domain sample to obtain a labeled training sample belonging to the target domain data distribution; wherein, different image styles correspond to different sample categories or data domains;

Input the target domain samples to be recognized to the trained small-sample recognition network for predictive classification and image recognition results.

In a possible implementation manner, the small sample recognition network includes: a feature extraction module and a metric matching module;

The feature extraction module is used to extract the features of the input samples;

The metric matching module is used for matching and classifying samples of different categories.

In a possible implementation manner, the training process of the small sample recognition network is as follows:

将生成样本/>

输入小样本识别网络中进行预测识别计算得到损失L_T，更新小样本识别网络F的参数同时，更新上述图像生成网络G的中间参数以生成更复杂的目标域风格样本，继续循环对小样本识别网络进行训练，直至训练好的小样本识别网络满足需求。When training the small-sample recognition network F, for each input task T, randomly select N types of images from M types of images in a loop, determine K samples for each type of image as the support set S, and determine q samples for each type of image The sample is used as a query set Q, T=(S, Q), and the generated sample is obtained after inputting the image generation network G

will generate a sample />

Input the small-sample recognition network to perform predictive recognition calculations to obtain the loss L _T , update the parameters of the small-sample recognition network F, and at the same time update the intermediate parameters of the above-mentioned image generation network G to generate more complex target domain style samples, and continue to cycle through small-sample recognition The network is trained until the trained small sample recognition network meets the requirements.

In a possible implementation manner, the style transformation module uses the stylization parameters output by the variational autoencoder as style features to stylize, so as to retain the content characteristics of the original samples and generate the new samples.

In a possible implementation manner, the metric matching module is used to match and classify the query set and support set of the current task example, specifically including: calculating the category center for N types of support set samples, and calculating the category center for each sample of the query set The distance to the N category centers, classify each sample in the query set to the nearest category center to complete the identification and classification of the query set samples.

In a possible implementation manner, the overall objective function of the small sample recognition network is:

计算的分类损失，α表示小样本识别网络的参数，A表示基于风格化支持集/>

和对应小样本识别网络参数α所选择的分类器，ω为小样本识别网络的输出结果。Among them, L _ω represents the stylized query set

The calculated classification loss, α represents the parameters of the small sample recognition network, and A represents the stylized support set/>

And the classifier selected by the corresponding small sample recognition network parameter α, ω is the output result of the small sample recognition network.

In the second aspect, the present invention provides a cross-domain small-sample image recognition system, including:

The recognition network determination unit is used to determine the trained small-sample recognition network; the small-sample recognition network is used to match and classify samples of different categories; the training process of the small-sample recognition network needs to use an image generation network, so The image generation network includes: a variational self-encoder module and a style conversion module, the variational self-encoder module is used to extract and reconstruct the feature distribution of the target domain samples, and the variational self-encoder includes intermediate parameters for Gradient ascent reconstructs the target domain samples into samples with more complex styles; the style conversion module is used to stylize the source domain samples to obtain labeled training samples belonging to the target domain data distribution; where different image styles correspond to different sample class or data field;

The sample recognition unit is used to input the target domain samples to be recognized to the trained small-sample recognition network to perform prediction classification and obtain image recognition results.

In a possible implementation manner, the small sample recognition network includes: a feature extraction module and a metric matching module; the feature extraction module is used to extract features of input samples; the metric matching module is used to match different types of samples Classification, specifically: matching and classifying the query set and support set of the current task example, specifically including: calculating the category center for N-type support set samples, calculating the distance from each sample in the query set to N category centers, and querying Collect each sample and classify it to the nearest category center to complete the identification and classification of the query set samples.

将生成样本/>

输入小样本识别网络中进行预测识别计算得到损失L_T，更新小样本识别网络F的参数同时，更新上述图像生成网络G的中间参数以生成更复杂的目标域风格样本，继续循环对小样本识别网络进行训练，直至训练好的小样本识别网络满足需求。In a possible implementation manner, the system also includes: a recognition network training unit, for training the small-sample recognition network F, for each input task T, cyclically randomly selects N types of images from M types of images, Determine K samples for each type of image as the support set S, determine q samples for each type of image as the query set Q, T=(S, Q), and generate samples after inputting the image generation network G

will generate a sample />

Input the small-sample recognition network to perform predictive recognition calculations to obtain the loss L _T , update the parameters of the small-sample recognition network F, and at the same time update the intermediate parameters of the above-mentioned image generation network G to generate more complex target domain style samples, and continue to cycle through small-sample recognition The network is trained until the trained small sample recognition network meets the requirements.

In a possible implementation, the style conversion module in the image generation network uses the stylization parameters output by the variational autoencoder as style features for stylization, so as to retain the content characteristics of the original samples and generate A new sample of the distribution characteristics of the data.

In a third aspect, the present application provides an electronic device, including: at least one memory for storing programs; at least one processor for executing the programs stored in the memory, and when the programs stored in the memory are executed, the processor is used for executing the first The method described in one aspect or any possible implementation of the first aspect.

In a fourth aspect, the present application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program runs on a processor, the processor executes any one of the first aspect or the first aspect. A possible implementation of the described method.

In a fifth aspect, the present application provides a computer program product, which, when the computer program product runs on a processor, causes the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

Generally speaking, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:

The present invention provides a cross-domain small-sample image recognition method and system. When training a small-sample recognition network model, an image to be tested is input, and a feature extraction network is used to realize feature extraction for input support set and query set samples. Since the feature extraction network has fully learned the distribution information about the target domain during the training phase, even if there is a domain gap between the test sample and the training sample, the network can effectively extract the category-related features in the sample during the feature extraction phase, which is helpful. to distinguish different types of samples.

The present invention provides a cross-domain small-sample image recognition method and system. The matching measurement module is selected to train the small-sample recognition model. By training a flexible measurement method and matching the similarity of features between different types of samples, the difference between the samples can be effectively found. The corresponding relationship between them, and then effectively match and identify.

The present invention provides a method and system for cross-domain small-sample image recognition. The generated network model based on confrontation includes a variational autoencoder module and a style conversion module, which can effectively introduce target domain information in the training phase and improve the accuracy of the model. The generalization ability of the domain enables the small-sample recognition model to achieve more accurate recognition on the target domain.

Description of drawings

FIG. 1 is a flowchart of a cross-domain small-sample image recognition method provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of the architecture of cross-domain small sample recognition provided by the embodiment of the present invention;

Fig. 3 is a schematic diagram of stylized generation provided by an embodiment of the present invention;

Fig. 4 is an architecture diagram of a cross-domain small-sample image recognition system provided by an embodiment of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The term "and/or" in this article is an association relationship describing associated objects, which means that there can be three relationships, for example, A and/or B can mean: A exists alone, A and B exist simultaneously, and B exists alone These three situations. The symbol "/" in this document indicates that the associated object is an or relationship, for example, A/B indicates A or B.

The terms "first" and "second" and the like in the specification and claims herein are used to distinguish different objects, rather than to describe a specific order of objects. For example, the first response message and the second response message are used to distinguish different response messages, rather than describing a specific order of the response messages.

In the embodiments of the present application, words such as "exemplary" or "for example" are used as examples, illustrations or illustrations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application shall not be interpreted as being more preferred or more advantageous than other embodiments or design schemes. Rather, the use of words such as "exemplary" or "such as" is intended to present related concepts in a concrete manner.

In the description of the embodiments of the present application, unless otherwise specified, "multiple" means two or more, for example, multiple processing units refer to two or more processing units, etc.; multiple A component refers to two or more components or the like.

First, technical terms involved in the embodiments of the present application are introduced.

(1) Stylization

Image stylization can also be called style migration, which means transferring the style of an image with artistic characteristics to an ordinary image, so that the original image retains the original content while having a unique artistic style, such as cartoons, comics, etc. , oil painting, watercolor, ink and so on.

(2) domain

Different domains refer to different sample data distributions. For example, images with different artistic styles have different data distributions, that is, images with different artistic styles belong to different domains.

Next, the technical solutions provided in the embodiments of the present application are introduced.

Fig. 1 is a flowchart of a cross-domain small-sample image recognition method provided by an embodiment of the present invention; as shown in Fig. 1 , it includes the following steps:

S101, determine the trained small-sample recognition network; the small-sample recognition network is used to match and classify samples of different categories; the training process of the small-sample recognition network needs to use an image generation network, and the image generation network includes : Variational autoencoder module and style conversion module, the variational autoencoder module is used to extract and reconstruct the feature distribution of the target domain samples, the variational autoencoder includes intermediate parameters, used for gradient ascent to convert the target domain The sample is reconstructed into a sample with a more complex style; the style conversion module is used to stylize the source domain sample to obtain a labeled training sample belonging to the target domain data distribution; wherein, different image styles correspond to different sample categories or data domains ;

S102. Input the target domain samples to be recognized to the trained small-sample recognition network to perform prediction classification and obtain image recognition results.

Specifically, first determine the pre-trained image generation network G and small sample recognition network F; the image generation network G includes: a variational autoencoder module and a style conversion module; the variational autoencoder module is used to extract And reconstruct the feature distribution of the target domain samples, the intermediate parameter ε is used for gradient ascent to generate more difficult style parameters, and the style conversion module is used to stylize the source domain samples to obtain labeled training samples belonging to the target domain data distribution ; The small sample recognition network F includes: a feature extraction module and a metric matching module; the feature extraction module is used to extract the characteristics of the input samples, and the metric matching module is used to match and classify different types of samples.

As shown in Figure 2, the framework of the cross-domain small sample identification method and system provided by the present invention includes two parts:

包括变分自编码器模块/>

和风格转换模块{F,g}；所述变分自编码器模块用于提取并重建目标域样本的特征分布，其中间参数ε用于梯度上升产生更难的风格参数，所述风格转换模块用于对源域样本进行风格化以得到属于目标域数据分布的带标签训练样本。(1) Image generation pre-training network

Includes Variational Autoencoder module />

and the style conversion module {F, g}; the variational autoencoder module is used to extract and reconstruct the feature distribution of the target domain samples, and the intermediate parameter ε is used for gradient ascent to generate more difficult style parameters, the style conversion module It is used to stylize the source domain samples to obtain labeled training samples belonging to the target domain data distribution.

(2) Small-sample recognition network T: includes a feature extraction module and a metric matching module; the feature extraction module is used to extract features of input samples, and the metric matching module is used to match and classify different types of samples.

In a specific embodiment, the present invention provides a cross-domain small sample identification method, comprising the following steps:

In another example, the present invention provides a cross-domain small sample recognition system, including:

Image generation pre-training module G: includes a variational self-encoder unit and a style conversion unit; the variational self-encoder unit is used to extract and reconstruct the feature distribution of the target domain samples, and the intermediate parameter ε is used for gradient ascent to generate more difficult The style parameters of the style conversion unit are used to stylize the source domain samples to obtain labeled training samples belonging to the target domain data distribution.

Small sample identification module T: includes a feature extraction unit and a metric matching unit; the feature extraction unit is used to extract features of input samples, and the metric matching unit is used to match and classify different types of samples.

输入模块G之后得到生成样本/>

将生成样本输入所述小样本识别模块中进行预测识别计算得到损失L_T，更新识别模块F的参数同时，更新上述生成模块G的中间参数以生成更难的目标域风格，继而完成上述小样本识别模块F的训练。When training the small-sample recognition module F, for each task T input, randomly select N types of images from M types of images in a loop, determine K samples for each type of image as a support set S, and determine for each type of image q samples are used as the query set Q, that is, T=(S,Q), where

After entering the module G to get the generated sample />

Input the generated samples into the small-sample recognition module for prediction and recognition calculation to obtain the loss L _T , update the parameters of the recognition module F, and at the same time update the intermediate parameters of the above-mentioned generation module G to generate a more difficult target domain style, and then complete the above-mentioned small sample The training of recognition module F.

Input the new category samples of the target domain to be recognized into the trained small sample recognition module to predict and classify the samples to be recognized, and output the corresponding recognition results.

In an optional example, the variational autoencoder unit is used to calculate the data distribution for a set of unlabeled unknown domain samples, and reconstruct the corresponding stylized parameters by sampling the anchor point ε from the distribution, and the parameters specifically include mean and variance.

In an optional example, the style conversion unit is used to stylize the example, select the current example as the content feature, the stylization parameters output by the above variational autoencoder are used as the style feature for stylization, and the output samples are retained The original content characteristics, so its labels are preserved, and at the same time it has the data distribution characteristics of the target domain.

In an optional example, the feature extraction unit is used for feature extraction, specifically including performing feature extraction on samples obtained after stylization of the current example.

In an optional example, the metric matching unit is used to match and classify the query set and support set of the current task example, specifically including: calculating the category center for N types of support set samples, and calculating the category center for each sample of the query set The distance to the centers of N categories to complete the identification and classification of query set samples.

In an optional example, the overall objective function of the recognition module is:

计算的分类损失，α表示上述小样本识别模块的参数，A表示基于风格化支持集/>

和对应小样本识别模块参数α所选择的分类器，ω即为其对应的输出结果。Among them, L _T represents the stylized query set

The calculated classification loss, α represents the parameters of the above small sample recognition module, and A represents the support set based on stylization />

And the classifier selected by the corresponding small sample recognition module parameter α, ω is its corresponding output result.

和方差/>

以减弱由采样的偶然性带来的对网络训练的影响，其中/>

之后再计算高斯分布的统计量N(ψ,ξ)，其中/>

对于每个小样本任务T＝{S_T,Q_T}，我们首先从高斯分布采样一个向量ε₁，输入D_vae解码得到向量/>

作为后面AdaIN网络的风格特征输入；将支持集图像S_T和查询集图像Q_T输入E_VGG，并将获得的特征作为AdaIN的另一个输入，AdaIN输出一些新的风格图像，表示为/>

这些风格图像近似服从目标域分布。根据前几节的介绍，“ε₁”即为“风格锚点”，因为VAE网络可以重建给定的M个目标域图像的分布。生成的风格化图像进一步输入任务模型，以解决小样本分类问题。在本发明中，选择RelationNet作为任务模型，其他解决小样本问题的模型也同样适用。为了搜索到更多符合目标域分布的风格，我们采用基于对抗的方法生成更困难的风格化样本，并尝试从“风格锚点”开始出发，迭代挖掘更多不可见的目标域分布。如图3所示，我们在任务模型的查询集上计算分类损失L_T1，我们获取对ε₁的反馈/>

并为下一次迭代准备更困难的样本/>

然后通过最小化L_T更新任务模型参数，希望可以更准确地目标域图像进行分类，即使模型具有出色的泛化能力。Fig. 3 is a schematic diagram of stylized generation provided by the embodiment of the present invention. As shown in Fig. 3, first, we calculate the mean value in the feature space based on a small number of target domain samples X _T

and variance />

In order to weaken the impact on network training brought by the contingency of sampling, where />

Then calculate the statistics N(ψ,ξ) of the Gaussian distribution, where />

For each small-sample task T={S _T ,Q _T }, we first sample a vector ε ₁ from the Gaussian distribution, and input D _vae to decode the vector />

As the style feature input of the subsequent AdaIN network; the support set image _ST and the query set image _QT are input into _EVGG , and the obtained features are used as another input of AdaIN, and AdaIN outputs some new style images, expressed as />

These style images approximately obey the target domain distribution. According to the introduction in the previous sections, “ε ₁ ” is the “style anchor”, because the VAE network can reconstruct the distribution of given M target domain images. The generated stylized images are further fed into the task model to solve the few-shot classification problem. In the present invention, RelationNet is selected as the task model, and other models for solving small sample problems are also applicable. In order to search for more styles that fit the target domain distribution, we use an adversarial approach to generate more difficult stylized samples, and try to iteratively mine more invisible target domain distributions starting from the "style anchor". As shown in Figure 3, we compute the classification loss L _T1 on the query set of the task model, and we obtain feedback on ε ₁ />

and prepare more difficult samples for the next iteration />

Then update the task model parameters by minimizing _LT , hoping to classify the target domain image more accurately, even if the model has excellent generalization ability.

Fig. 4 is a cross-domain small-sample image recognition system architecture diagram provided by an embodiment of the present invention, as shown in Fig. 4, including:

将生成样本/>

输入小样本识别网络中进行预测识别计算得到损失L_T，更新小样本识别网络F的参数同时，更新上述图像生成网络G的中间参数以生成更复杂的目标域风格样本，继续循环对小样本识别网络进行训练，直至训练好的小样本识别网络满足需求。The recognition network training unit 410 is used to train the small-sample recognition network F, for each input task T, randomly select N types of images from M types of images, and determine K samples for each type of image as the support set S , determine q samples for each type of image as the query set Q, T=(S, Q), and generate samples after inputting the image generation network G

will generate a sample />

Input the small-sample recognition network to perform predictive recognition calculations to obtain the loss L _T , update the parameters of the small-sample recognition network F, and at the same time update the intermediate parameters of the above-mentioned image generation network G to generate more complex target domain style samples, and continue to cycle through small-sample recognition The network is trained until the trained small sample recognition network meets the requirements.

The recognition network determination unit 420 is used to determine the trained small-sample recognition network; the small-sample recognition network is used to match and classify samples of different categories; the training process of the small-sample recognition network requires the use of an image generation network, The image generation network includes: a variational self-encoder module and a style conversion module, the variational self-encoder module is used to extract and reconstruct the feature distribution of the target domain samples, the variational self-encoder includes intermediate parameters, with The target domain samples are reconstructed into samples with more complex styles based on gradient ascent; the style conversion module is used to stylize the source domain samples to obtain labeled training samples belonging to the target domain data distribution; where different image styles correspond to different The sample category or data domain of ;

The sample recognition unit 430 is configured to input the target domain samples to be recognized into the trained small-sample recognition network to perform prediction classification and obtain image recognition results.

It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment, and the corresponding program modules in the device have similar implementation principles and technical effects to the description in the above-mentioned method, and the working process of the device can refer to the corresponding program module in the above-mentioned method The process will not be repeated here.

Based on the methods in the foregoing embodiments, the embodiments of the present application provide an electronic device. The apparatus may comprise at least one memory for storing a program and at least one processor for executing the program stored in the memory. Wherein, when the program stored in the memory is executed, the processor is configured to execute the methods described in the above embodiments.

Based on the methods in the above-mentioned embodiments, the embodiments of the present application provide a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program runs on the processor, the processor executes the above-mentioned embodiment. Methods.

Based on the methods in the foregoing embodiments, the embodiments of the present application provide a computer program product that, when the computer program product runs on a processor, causes the processor to execute the methods in the foregoing embodiments.

It can be understood that the processor in the embodiment of the present application may be a central processing unit (central processing unit, CPU), and may also be other general processors, digital signal processors (digital signal processor, DSP), application specific integrated circuits (application specific integrated circuit, ASIC), field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. A general-purpose processor can be a microprocessor, or any conventional processor.

The method steps in the embodiments of the present application may be implemented by means of hardware, or may be implemented by means of a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (random access memory, RAM), flash memory, read-only memory (read-only memory, ROM), programmable read-only memory (programmable rom) , PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically erasable programmable read-only memory (electrically EPROM, EEPROM), register, hard disk, mobile hard disk, CD-ROM or known in the art any other form of storage medium. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be a component of the processor. The processor and storage medium can be located in the ASIC.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in or transmitted via a computer-readable storage medium. The computer instructions may be transmitted from one website site, computer, server, or data center to another website site by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) , computer, server or data center for transmission. The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (such as a floppy disk, a hard disk, or a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (solid state disk, SSD)) and the like.

It can be understood that the various numbers involved in the embodiments of the present application are only for convenience of description, and are not used to limit the scope of the embodiments of the present application.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (10)

1. The cross-domain small sample image recognition method is characterized by comprising the following steps of:

determining a trained small sample recognition network; the small sample recognition network is used for carrying out matching classification on samples of different categories; the training process of the small sample recognition network requires an image generation network, wherein the image generation network comprises: the system comprises a variable self-encoder module and a style conversion module, wherein the variable self-encoder module is used for extracting and reconstructing characteristic distribution of a target domain sample, and comprises an intermediate parameter used for reconstructing the target domain sample into a sample with more complex style by gradient rising; the style conversion module is used for stylizing the source domain sample to obtain a labeled training sample belonging to the target domain data distribution; wherein different image styles correspond to different sample categories or data fields;

and inputting the target domain sample to be identified into a trained small sample identification network to conduct prediction classification, so as to obtain an image identification result.

2. The method of claim 1, wherein the small sample identification network comprises: the feature extraction module and the measurement matching module;

the characteristic extraction module is used for extracting characteristics of an input sample;

the measurement matching module is used for carrying out matching classification on samples of different categories.

3. The method of claim 1, wherein the training process of the small sample recognition network is as follows:

when training the small sample recognition network F, for each input task T, circularly and randomly selecting N types of images from M types of images, determining K samples as a support set S for each type of images, determining Q samples as a query set Q, T= (S, Q) for each type of images, and obtaining a generated sample after inputting the images into the image generation network G

Will generate samples->

Inputting into a small sample recognition network to perform prediction recognition calculation to obtain loss L _T And updating parameters of the small sample recognition network F, and updating intermediate parameters of the image generation network G to generate more complex target domain style samples, and continuously training the small sample recognition network in a circulating way until the trained small sample recognition network meets the requirements.

4. A method according to any one of claims 1 to 3, wherein the stylistic conversion module stylizes the stylized parameters output by the variational self-encoder as stylized features to preserve the content features of the original samples and generate new samples with target domain data distribution features.

5. The method according to claim 2, wherein the metric matching module is configured to perform matching classification on the query set and the support set of the current task instance, and specifically includes: and calculating class centers for the N class support set samples, calculating distances from each sample of the query set to the N class centers, and classifying each sample of the query set to the class center closest to the sample of the query set to finish the identification and classification of the sample of the query set.

6. A method according to any one of claims 1 to 3, characterized in that the overall objective function of the small sample recognition network is:

wherein L is _T Representing a stylized query set

Calculated classification loss, alpha denotes parameters of the small sample recognition network, A denotes the stylized support set based +.>

And a classifier selected corresponding to the small sample identification network parameter alpha, wherein omega is an output result of the small sample identification network.

7. A cross-domain small sample image recognition system, comprising:

the identification network determining unit is used for determining a trained small sample identification network; the small sample recognition network is used for carrying out matching classification on samples of different categories; the training process of the small sample recognition network requires an image generation network, wherein the image generation network comprises: the system comprises a variable self-encoder module and a style conversion module, wherein the variable self-encoder module is used for extracting and reconstructing characteristic distribution of a target domain sample, and comprises an intermediate parameter used for reconstructing the target domain sample into a sample with more complex style by gradient rising; the style conversion module is used for stylizing the source domain sample to obtain a labeled training sample belonging to the target domain data distribution; wherein different image styles correspond to different sample categories or data fields;

the sample recognition unit is used for inputting the target domain sample to be recognized into the trained small sample recognition network so as to conduct prediction classification and obtain an image recognition result.

8. The system of claim 7, wherein the small sample recognition network comprises: the feature extraction module and the measurement matching module; the characteristic extraction module is used for extracting characteristics of an input sample; the measurement matching module is used for carrying out matching classification on samples of different categories, and specifically comprises the following steps: the method for matching and classifying the query set and the support set of the current task example specifically comprises the following steps: and calculating class centers for the N class support set samples, calculating distances from each sample of the query set to the N class centers, and classifying each sample of the query set to the class center closest to the sample of the query set to finish the identification and classification of the sample of the query set.

9. The system of claim 7, further comprising: the recognition network training unit is used for circularly and randomly selecting N types of images from M types of images for each input task T when training the small sample recognition network F, determining K samples as a support set S for each type of images, determining Q samples as a query set Q for each type of images, and obtaining a generated sample after inputting the small sample recognition network F into the image generation network G

Will generate samples->

Inputting into a small sample recognition network to perform prediction recognition calculation to obtain loss L _T And updating parameters of the small sample recognition network F, and updating intermediate parameters of the image generation network G to generate more complex target domain style samples, and continuously training the small sample recognition network in a circulating way until the trained small sample recognition network meets the requirements.

10. The system according to any one of claims 7 to 9, wherein the stylistic conversion module in the image generation network stylizes the stylized parameters output from the encoder as stylized features to preserve content features of original samples and generate new samples with target domain data distribution features.

Images