CN116091891A

CN116091891A - Image recognition method and system

Info

Publication number: CN116091891A
Application number: CN202310009782.4A
Authority: CN
Inventors: 尹芳; 邓小宁; 金剑; 马杰; 袁园
Original assignee: North Health Medical Big Data Technology Co ltd
Current assignee: North Health Medical Big Data Technology Co ltd
Priority date: 2023-01-04
Filing date: 2023-01-04
Publication date: 2023-05-09

Abstract

The invention provides an image recognition method and system, wherein the method comprises the following steps: determining a first image similar to a target image in each image according to the similarity between the target image and each image in a public data set, wherein the target image is obtained by processing an original image based on federal learning; and inputting the first image into a target GAN, and outputting a recognition result of the first image, wherein the target GAN is obtained after training the GAN based on the public data set. The system performs the method. According to the invention, the recognition result of the first image similar to the target image processed by the original image by utilizing the federal learning is output based on the trained target GAN, and the reduction degree detection of the original image can be realized by comparing the recognition result with the original image.

Description

Image recognition method and system

Technical Field

The present invention relates to the field of image processing technologies, and in particular, to an image recognition method and system.

Background

The medical industry promotes digital transformation, and a large amount of medical data is orderly accumulated, but the difficulty in acquiring the data is a great difficulty in clinical research of each medical institution. In one aspect, the medical data relates to private information of the patient; on the other hand, the limited sample size accumulated by a single medical institution is insufficient to support evidence-based research in modern medicine. The federal study can safely utilize and develop data resources, can realize the medical data safety statistical analysis and medical simulation and prognosis under the data protection, and can bring creative changes for clinical diagnosis and treatment, drug research and development, health monitoring, public health and the like, and continuously mine and release data value.

In the prior art, methods such as block chain combination federal learning, federal learning fusion and the like (such as methods of utilizing neural network vectors, a cooperative mechanism and the like) are mainly adopted to reconstruct images in federal learning, and the block chain combination federal learning method needs federal learning and block chain fusion, so that the realization is more complex, and the accuracy is low.

Disclosure of Invention

The image recognition method and the system provided by the invention are used for solving the problems of complex realization and low accuracy of image reduction degree detection for federal learning in the prior art, and can improve the accuracy of reduction degree detection of an original image and reduce the realization complexity of image reduction degree detection for federal learning.

The invention provides an image recognition method, which comprises the following steps:

determining a first image similar to a target image in each image according to the similarity between the target image and each image in a public data set, wherein the target image is obtained by processing an original image based on federal learning;

and inputting the first image into a target generation type countermeasure network GAN, outputting a recognition result of the first image, and training the GAN based on the public data set by the target GAN.

According to the image recognition method provided by the invention, the first image similar to the target image in each image is determined according to the similarity between the target image and each image in the public data set, and the method comprises the following steps:

acquiring a second image with the maximum similarity between each image and the target image;

and acquiring the first image according to the second image.

According to the image recognition method provided by the invention, the step of acquiring the first image according to the second image comprises the following steps:

in the case that the function used by the federal learning is a linear function, taking the second image as the first image;

and under the condition that the function used by the federal learning is a nonlinear function, optimizing the second image based on a gradient-free optimization algorithm, and taking the optimized second image as the first image.

According to the image recognition method provided by the invention, in the case that the function used by the federal learning is a nonlinear function, the second image is optimized based on a gradient-free optimization algorithm, and the optimized second image is used as the first image, and the method comprises the following steps:

and optimizing the second image based on a Bayesian optimization algorithm and/or a covariance matrix adaptive algorithm, and taking the optimized second image as the first image.

According to the image recognition method provided by the invention, the target image acquisition method comprises the following steps:

processing the original image based on the federal learning to obtain a third image;

partitioning the third image to obtain a fourth image;

performing label repair on the fourth image to obtain a fifth image;

and optimizing the fifth image based on a seed optimization algorithm to obtain the target image.

According to the image recognition method provided by the invention, the optimization is performed on the fifth image based on the seed optimization algorithm to obtain the target image, and the method comprises the following steps:

and optimizing the fifth image based on a global optimality optimization algorithm and a particle swarm algorithm to obtain the target image.

The invention also provides an image recognition system, comprising: the data acquisition module and the image recognition module;

the data acquisition module is used for determining a first image similar to the target image in each image according to the similarity between the target image and each image in the public data set, wherein the target image is obtained by processing an original image based on federal learning;

the image recognition module is used for inputting the first image into a target generation type countermeasure network GAN, outputting a recognition result of the first image, and obtaining the target GAN after training the GAN based on the public data set.

The invention also provides an electronic device comprising a processor and a memory storing a computer program, the processor implementing the image recognition method as described in any one of the above when executing the program.

The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the image recognition method as described in any of the above.

The invention also provides a computer program product comprising a computer program which, when executed by a processor, implements the image recognition method as described in any one of the above.

According to the image recognition method and system provided by the invention, the GAN is trained by using the public data set, the recognition result of the first image similar to the target image processed by the original image by using the federal learning is output based on the trained target GAN, the reduction degree detection of the original image can be realized by comparing the recognition result with the original image, compared with the image reduction degree detection method for federal learning by using the neural network vector in the prior art, the accuracy of the reduction degree detection of the image for federal learning is improved, and the complexity of the image reduction degree detection for federal learning is reduced compared with the image reduction degree detection method for federal learning by using the cooperative mechanism in the prior art.

Drawings

In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of an image recognition method according to the present invention;

FIG. 2 is a schematic illustration of a segmented lung image provided by the present invention;

FIG. 3 is a second flowchart of an image recognition method according to the present invention;

FIG. 4 is a schematic diagram of an image recognition system according to the present invention;

fig. 5 is a schematic diagram of the physical structure of the electronic device provided by the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Fig. 1 is a schematic flow chart of an image recognition method provided by the present invention, as shown in fig. 1, the method includes:

step 110, determining a first image similar to a target image in each image according to the similarity between the target image and each image in a public data set, wherein the target image is obtained by processing an original image based on federal learning;

step 120, inputting the first image into a target generation type countermeasure network (Generative Adversarial Network, GAN), outputting a recognition result of the first image, and training the GAN based on the public data set to obtain the target GAN.

It should be noted that, the execution subject of the above method may be a computer device.

Alternatively, the target image may specifically refer to an image obtained by processing an original image by using federal learning, and the original image may specifically be a medical image of a patient, and more specifically may be a lung image, an eye image, or the like of the patient. The common dataset may in particular be a set of images of the same features as the original image in particular, for example, when the original image is a lung image, the common dataset may be a lung image dataset and when the original image is an eye image, the common dataset may be an eye image dataset.

Along with the strict supervision of operation entities related to data, the premise of reasonably and legally mining the value of the data between enterprises and various institutions or individuals is to solve the problem of data privacy protection. In the current age, data is taken as an asset, all the participants flow right, the right to use flows on the premise of guaranteeing the ownership of the data, the data is guaranteed to be available and invisible to all the participants, and at least one party of data in a traditional mode commonly adopted at present is taken out of a database, so that leakage risks exist.

The privacy calculation is a technical set for realizing data analysis and calculation on the premise of protecting the data from external leakage, so as to achieve the purpose of being 'available and invisible' for the data; on the premise of fully protecting data and privacy safety, the conversion and release of data value are realized. The key technologies of privacy computation can be divided into three categories, the first category is federal learning, the second category is multiparty security computation, and the third category is privacy computation technology based on trusted hardware, which is represented by trusted execution environment.

The invention adopts federal learning to process the original image so as to protect the data security.

Alternatively, the similarity may be specifically a cosine similarity between the target image and each image in the common dataset, and the first image similar to the target image is found by calculating the cosine similarity between the target image and each image in the common dataset.

The cosine similarity is calculated as follows:

wherein similarity represents cosine similarity, A and B are vector expression forms of the target image and any image in the public data set respectively, and A= (a) ₁ ,a ₂ ,...,a _n )，B＝(b ₁ ,b ₂ ,...,b _n )。

Specifically, the size of any one of the target image and the public data set is adjusted to be the same size, a gray histogram of any one of the target image and the public data set is obtained, any one of the target image and the public data set is partitioned, for example, every 3 or 5 equal gray levels of any one of the target image and the public data set can be partitioned into one region, and a summation operation is performed on 3 or 5 values of each region to obtain a value, the value is taken as a vector expression form of any one of the target image and the public data set, cosine similarity of the two vectors is calculated, similarity of any one of the target image and the public data set is judged, and a first image similar to the target image is found.

Alternatively, the target generated type of the countermeasure network GAN may be specifically obtained after training the generated type of the countermeasure network GAN based on the common data set, specifically:

the pre-trained generated countermeasure network GAN on the public data set is used as a priori model, and the expression closest to the gradient of the real medical image (namely the original image) is found in the space of the generated countermeasure network GAN, so that the quality of the identification result of the restored target image (such as the medical image) is improved.

In the training process, when the training task of the generated countermeasure network GAN is completed together with the federal learning parties, a terminal playing a leading role in federal learning needs to additionally train a countermeasure network, wherein a generator is used for simulating samples in a public data set, and a discriminator helps to improve the result of reconstructing an image, so that the reduction degree of the image reconstructed by the generator, such as a medical image, and an original image is continuously improved.

Compared with the existing method for reconstructing data only by relying on gradient information, the method provided by the invention utilizes learning of the target generation type antagonism network GAN from images (such as lung images, eye images and other medical images) in a public data set to compensate the gradient information loss of the medical images in the gradient degradation process.

The generated countermeasure network GAN may specifically include a generator neural network and a arbiter neural network:

minmaxV(D,G)＝EX～pdata(x)[logD(x)]+EZ～pz(z)[log(1-D(G(z)))]

the objective function of the generated countermeasure network GAN is V (D, G), G (Z) is generated, E in a discriminator D (X) represents entropy, X-pdata (X) represents that X is from pdata 'real data (namely original image) distribution', E in the discriminator D (X) represents mathematical expectation of the real data X and noise data Z, the real data distribution is approximated by fitting the noise data distribution pz (Z), the relative relation between the data is not changed by taking the logarithm of the data, and the log can be used for amplifying loss due to monotonicity of the log function and log addition, so that the calculation and optimization are convenient.

Optionally, the obtained first image is input into the target generation type countermeasure network GAN, a recognition result of the first image is output, and based on the recognition result and the original image, detection of the reduction degree of the original image after federal learning processing can be realized, specifically: the reduction degree of the original image can be determined according to the cosine similarity between the identification result and the original image, and the larger the cosine similarity between the identification result and the original image is, the higher the reduction degree of the original image is, and otherwise, the lower the reduction degree of the original image is.

According to the image recognition method provided by the invention, the GAN is trained by using the public data set, the recognition result of the first image similar to the target image processed by the original image by using the federal learning is output based on the trained target GAN, the reduction degree detection of the original image can be realized by comparing the recognition result with the original image, compared with the image reduction degree detection method for federal learning by using the neural network vector in the prior art, the accuracy of the reduction degree detection of the image for federal learning is improved, and compared with the image reduction degree detection method for federal learning by using the cooperative mechanism in the prior art, the complexity of the image reduction degree detection for federal learning is reduced.

Further, in one embodiment, the determining, according to the similarity between the target image and each image in the common dataset, the first image similar to the target image in each image includes:

and acquiring the first image according to the second image.

Optionally, the vector expression form of each image in the common data set and the target image is obtained, the cosine similarity between each image in the common data set and the target image is calculated, and the image (namely the second image) with the maximum cosine similarity between each image and the target image is found.

From the second image, a first image similar to the target image is acquired. The first image is the image closest to the gradient information of the target image, and the quality of the reconstructed image is improved by mining the gradient information of the first image.

For example, a first image that is more similar to the reconstructed medical image is found, the distance between transformed gradients of the reconstructed medical image is used to generate the reconstructed medical image and the gradient changes of the medical image are observed, and the distance metric gradient matching loss is calculated.

The gradient matching loss is calculated through distance measurement, specifically, according to the gray level histogram of the medical image and the image in the public data set, the medical image and the image in the public data set are converted into vector expression form, and the cosine similarity between the medical image and each image in the public data set is calculated through the cosine value between the two vectors.

Further, in an embodiment, the acquiring the first image according to the second image includes:

Further, in an embodiment, in a case that the function used by the federal learning is a nonlinear function, the optimizing the second image based on the gradient-free optimization algorithm, and taking the optimized second image as the first image includes:

Optionally, if the function used in federal learning is a simple outward linear function, the quality of the reconstructed image can be better improved by taking the second image with the maximum cosine similarity with the target image as the first image, but if the function used in federal learning is nonlinear, a certain difficulty is often brought. In order to solve the problem of the non-linear function used in federal learning, it is necessary to optimize the second image by using a gradient-free optimization algorithm, and take the optimized second image as the first image similar to the target image.

The gradient-free optimization algorithm may specifically include a bayesian optimization algorithm and a covariance matrix adaptive algorithm, and when the second image is optimized, the two algorithms of the bayesian optimization algorithm and the covariance matrix adaptive algorithm may be used in combination, or one algorithm may be selected from the two algorithms to optimize the second image. For example, the second image is optimized based on a Bayesian optimization algorithm, which is a global operation optimization method, and can be used for solving the problems of random noise and the like in a black box function modeled by a Gaussian process, so that the accuracy of image reduction detection in federal learning is improved.

It should be noted that, the bayesian optimization algorithm or the covariance matrix adaptive algorithm is implemented to optimize the second image through gradient information of the second image.

According to the image identification method provided by the invention, the second image with the maximum cosine similarity between each image in the public data set and the target image is found, the first image closest to the gradient information of the target image is found according to the second image, the first image is input into the target GAN, the identification result of the first image is obtained, the original image is compared with the original image based on the identification result, the reduction degree detection of the original image can be realized, and the accuracy of the reduction degree detection of the original image is improved.

Further, in one embodiment, the method for acquiring the target image may specifically include:

partitioning the third image to obtain a fourth image;

performing label repair on the fourth image to obtain a fifth image;

Optionally, the working principle of federal learning is: the client terminal downloads the existing prediction model from the central server, trains the prediction model by using local data, uploads the updated content of the prediction model to the cloud, in the training process, the prediction model is optimized by fusing the prediction model updates of different client terminals, the client terminal downloads the updated prediction model to the local, the process is repeated continuously, and in the whole process, the data of the client terminal are always stored locally, so that the risk of data leakage does not exist. Machine learning based on federal learning collaborative construction is almost lossless in performance compared with the machine learning model obtained by centralized training.

Processing an original image based on federation learning to obtain a third image, and performing block processing on the third image to obtain a fourth image, wherein the block processing specifically refers to dividing the third image into a plurality of areas, namely dividing the image into a plurality of image blocks, and each image block is regarded as a view for performing a subsequent task of medical image reduction degree under a federation learning scene. The segmentation rule may be a random segmentation, or may be a segmentation based on the difference between the organ position and the focal region of the medical image of the patient, which is not particularly limited in the present invention.

For example, fig. 2 is a schematic view of a segmented lung image according to the present invention, and as shown in fig. 2, the lung image is divided into 9 regions by mesh division of 3*3. The reason for the segmentation of the lung image is mainly to consider that most organ locations are similar, and only the focal areas are subject to inter-individual variation.

Optionally, label repairing is performed on the fourth image to obtain a fifth image, specifically:

detecting the restoration degree of the medical image in federal learning, firstly attempting to restore the label of the medical image, performing joint modeling learning by using a deep learning model in federal learning, restoring the batch label from the gradient of the fully-connected classification layer by adopting a label restoration method, and restoring the label of the medical image in federal learning. For example, in a medical image classification task, samples in a medical image can be recovered from a batch gradient in a convolutional neural network (Convolutional Neural Networks, CNN) and a recurrent neural network (Recurrent Neural Network, RNN), as shown in the following formula.

Wherein y is ₁ The label of the restored image is restored after the label is restored; index represents the index value of the returned data, which orders the data from small to large; w is the average weight of the batch, a "synthetic" input batch, initialized to random noise, and optimized towards true values; c is the number of color channels, b is the number of embedded features, a is the number of target classes, N is the true value gradient of the nth layer, and all layers are traversed. x is the fourth image, y is the label of the fourth image formulated, and h is the batch size. The N () function emphasizes that the gradient of the composite data (for the original loss of the network with weight w) matches the gradient provided.

Optionally, the fifth image is optimized, specifically, the fifth image may be optimized by using a seed optimization algorithm, so as to obtain the target image.

The process of optimizing the fifth image based on the seed optimization algorithm is essentially a consistency regularization process of the fifth image. Due to the fact that there is under-fitting and over-fitting in the label repair process for the fourth image using RNN or CNN. Under-fitting is a feature of the model that does not fit well, over-fitting is a feature that fits the model too well, resulting in too much attention being given to the details of each data, while ignoring the overall features of the data. The over-fitting and under-fitting problems present in the model can be solved by consistent regularization.

The consistency regularization treatment process specifically utilizes a plurality of seed optimization algorithms simultaneously in a common optimization mode to optimize input to match a target gradient, so that the quality of the restoration degree of the reconstructed medical image is improved. The same loss super-parameters are adopted in the optimization mode of the common optimization, different random seeds are given out and used for initializing the pixel level of the medical image, a local minimum value can be found in each optimization cycle, and the medical image features with correct semantics are distributed on all levels. The medical image pixel mean is improved by mixing information from all seeds in the group and feedback using consistent regularization (which makes the model produce the same output distribution when the input is disturbed, i.e. small disturbances to the input should not change the output of the model) while regularizing all inputs.

The improvement of the input reconstructed medical image is updated in an iterative mode, pixel-level random Gaussian noise is added in each update, the diversity is increased, and the gap between the reconstructed medical image and the original image is further reduced.

A simple consistency regularization method is that if RNN or CNN is insensitive to noise, noise is added to carry out neural network training when training is carried out, and noise items are put into a loss function.

Further, in an embodiment, the optimizing the fifth image based on the seed optimization algorithm to obtain the target image includes:

Optionally, the seed optimization algorithm may specifically include a global optimality optimization algorithm and a particle swarm algorithm, and when the fifth image is optimized, the two algorithms may be used in combination.

Since federal learning is not a local privacy protection strategy for data, it has been considered to solve the artificial intelligence calculation problem efficiently and protect the important direction of personal data, however, with the current deepening of data limitation, a method of back-pushing user data from gradient and model parameters is emerging.

In many cases, the basic information of the image and the like can still be reconstructed by using the blurred data and parameters in the machine learning process. The gradient shared training scheme is not absolutely safe, and the server can recover the user's local training data from the exchanged parameter gradients using gradient attacks, despite avoiding direct contact of the central server with the user data. The method comprises the steps of randomly generating virtual training data, generating virtual gradients based on the virtual training data, and repeatedly iterating through gradient descent by taking the difference between the reduced virtual gradients and the actual gradients as an optimization target, so that private data of a user can be restored. Such attacks are known as gradient leakage or gradient reversal.

In order to defend against such gradient attacks (gradient leakage or gradient reversal), some methods propose that users can add noise disturbance to gradient information or perform lossy transformation, such as gradient processing, before uploading to better ensure information security.

The invention defends the gradient attack (gradient leakage or gradient reversal) by carrying out adaptive attack training on the target generation type countermeasure network GAN, and the specific process is as follows:

the adaptive attack is an attack method designed for a defense mechanism by adopting a certain knowledge of an attacker on the proposed defense mechanism. The same transformation can be adopted in the optimization process to carry out the adaptive attack. The goal of an attacker is to reveal as much medical image information as possible from the degraded gradient, and the attacker may or may not know the potential defense strategy employed by the client. An adaptive attack is initiated by directly inputting this knowledge or by estimating the defense parameters through the observed gradients. Furthermore, it is assumed that the attacks can be facilitated and improved with knowledge extracted from the common dataset (broadly referring to some defending against empirical parameters, networks, etc.), with adaptive attacks, by estimating the variation (continuous monitoring and analysis is the core of the adaptive security architecture, meaning analyzing each attack differently, such as the loss function curves employed by the attack, etc.) and incorporating it into the optimization process of federally learned medical image restoration degree detection.

For example, fig. 3 is a second flowchart of an image recognition method according to the present invention, as shown in fig. 3, the method may specifically include:

step 1, inputting medical images corresponding to original images after federal learning processing;

step 2, the medical image is segmented;

step 3, performing label restoration on the segmented medical image;

step 4, carrying out consistency regularization treatment on the image after label repair;

step 5, finding a second image with the greatest similarity with the image subjected to the consistency regularization treatment from each image in the public data set;

step 6, judging whether the function used by federation learning is a linear function or not;

step 7, if the function is linear, inputting the second image into the target generation type countermeasure network GAN;

step 8, if the second image is a nonlinear function, performing gradient information optimization on the second image based on a Bayesian optimization algorithm and/or a covariance adaptive algorithm to obtain a first image, and inputting the first image to a target GAN;

and 9, performing adaptive attack training on the target GAN.

According to the image recognition method provided by the invention, the reduction degree judgment of the medical image in federal learning is realized through the methods of blocking the medical image, repairing the label, regularizing, deep mining gradient information, generating target GAN, adaptively attacking and the like, the manual intervention required in the existing method is reduced, the quality of the medical image in federal learning is improved, the reconstructed medical image with higher quality can still be restored when the gradient information with low fidelity and noise is faced, the reduction degree of the medical image in federal learning is detected, a certain reduction effect is still obtained under the measures of better guaranteeing information safety, confusing data characterization and the like by adopting gradient clipping or gradient thinning treatment, and the medical image information loss in the gradient degradation process is compensated by utilizing the method of generating an countermeasure network from the common data set, and the leakage degree of the medical image information in federal learning scene can be measured through the detection of the reduction degree so as to promote a better joint defense learning mechanism.

The image recognition system provided by the invention is described below, and the image recognition system described below and the image recognition method described above can be referred to correspondingly to each other.

Fig. 4 is a schematic structural diagram of an image recognition system provided by the present invention, as shown in fig. 4, including:

a data acquisition module 410 and an image recognition module 411;

the data acquisition module 410 determines a first image similar to the target image in each image according to the similarity between the target image and each image in the public data set, wherein the target image is obtained by processing an original image based on federal learning;

the image recognition module 411 is configured to input the first image to a target generation type countermeasure network GAN, output a recognition result of the first image, and obtain the target GAN after training the GAN based on the common dataset.

According to the image recognition system provided by the invention, the GAN is trained by using the public data set, the recognition result of the first image obtained after the original image is processed by utilizing the federal learning is output based on the trained target GAN, the reduction degree detection of the original image can be realized by comparing the recognition result with the original image, compared with the image reduction degree detection method for federal learning by adopting the neural network vector in the prior art, the accuracy of the reduction degree detection of the image for federal learning is improved, and compared with the image reduction degree detection method for federal learning by adopting the cooperative mechanism in the prior art, the complexity of the image reduction degree detection for federal learning is reduced.

Fig. 5 is a schematic physical structure of an electronic device according to the present invention, as shown in fig. 5, the electronic device may include: a processor (processor) 510, a communication interface (communication interface) 511, a memory (memory) 512 and a bus (bus) 513, wherein the processor 510, the communication interface 511 and the memory 512 communicate with each other via the bus 513. Processor 510 may invoke logic instructions in memory 512 to perform the following method:

Further, the logic instructions in the memory described above may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer power supply screen (which may be a personal computer, a server, or a network power supply screen, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

Further, the present invention discloses a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, are capable of performing the image recognition method provided by the above-mentioned method embodiments, for example comprising:

In another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the image recognition method provided in the above embodiments, for example, including:

The system embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer power screen (which may be a personal computer, a server, or a network power screen, etc.) to perform the method described in the various embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image recognition method, comprising:

2. The image recognition method according to claim 1, wherein the determining a first image similar to the target image in each image according to a similarity between the target image and each image in the common dataset comprises:

and acquiring the first image according to the second image.

3. The image recognition method according to claim 2, wherein the acquiring the first image from the second image includes:

4. The image recognition method according to claim 3, wherein, in the case where the function used by the federal learning is a nonlinear function, the optimizing the second image based on the gradient-free optimization algorithm, and taking the optimized second image as the first image, comprises:

5. The image recognition method according to any one of claims 1 to 4, wherein the target image acquisition method includes:

partitioning the third image to obtain a fourth image;

performing label repair on the fourth image to obtain a fifth image;

6. The image recognition method according to claim 5, wherein the optimizing the fifth image based on the seed optimization algorithm to obtain the target image includes:

7. An image recognition system, comprising: the data acquisition module and the image recognition module;

8. An electronic device comprising a processor and a memory storing a computer program, characterized in that the processor implements the image recognition method of any one of claims 1 to 6 when executing the computer program.

9. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the image recognition method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the image recognition method according to any one of claims 1 to 6.