CN110414611A

CN110414611A - Image classification method and device, feature extraction network training method and device

Info

Publication number: CN110414611A
Application number: CN201910702994.4A
Authority: CN
Inventors: 邓晗; 张学森; 伊帅; 闫俊杰
Original assignee: Beijing Sensetime Technology Development Co Ltd
Current assignee: Beijing Sensetime Technology Development Co Ltd
Priority date: 2019-07-31
Filing date: 2019-07-31
Publication date: 2019-11-05

Abstract

This disclosure relates to a kind of image classification method and device, feature extraction network training method and device, described image classification method, comprising: image to be processed input fisrt feature is extracted network, obtains the fisrt feature of image to be processed；Fisrt feature is inputted to M second feature respectively and extracts network, obtains M second feature；According to M second feature, the classification results of the image to be processed are obtained.Image classification method according to an embodiment of the present disclosure can extract network by multiple second feature and extract the various features in sample image respectively, keep characteristic information richer, improve the accuracy rate of classification processing.

Description

Image classification method and device, feature extraction network training method and device

Technical field

This disclosure relates to field of computer technology more particularly to a kind of image classification method and device, feature extraction network Training method and device.

Background technique

In the related art, neural network is when extracting the feature of the target object in image, using target object as one A entirety carries out feature extraction, but the appearance of target object, posture, in the picture proportion, complex degree of background, target pair As the factors such as integrated degree in the picture may influence to extract as a result, if extracting multiple features of target object, mind More through network training process parameter, training difficulty is larger.

Summary of the invention

The present disclosure proposes a kind of image classification method and devices, feature extraction network training method and device.

According to the one side of the disclosure, a kind of image classification method is provided, comprising:

Image to be processed input fisrt feature is extracted into network, obtains the fisrt feature of the image to be processed；

The fisrt feature is inputted into M second feature respectively and extracts network, obtains M second feature；

According to the M second feature, the classification results of the image to be processed are obtained.

Image classification method according to an embodiment of the present disclosure can extract network by multiple second feature and extract sample respectively Various features in this image keep characteristic information richer, improve the accuracy rate of classification processing.

In one possible implementation, the fisrt feature is the default network in fisrt feature extraction network The output feature of level, it is described according to the M second feature, obtain the classification results of the image to be processed, comprising:

By the fisrt feature extract network described in preset network level subsequent level to the fisrt feature into Row feature extraction processing, obtains third feature；

The classification results of the image to be processed are obtained according to the third feature and the M second feature.

According to the one side of the disclosure, a kind of feature extraction network training method is provided, comprising:

Sample image input fisrt feature is extracted into network, obtains the first training characteristics of the sample image；

First training characteristics are inputted into M second feature respectively and extract network, obtain M the second training characteristics；

According to the M the second training characteristics, M second feature of training extracts network.

Neural network training method according to an embodiment of the present disclosure can extract network by multiple second feature and mention respectively Take the various features in sample image, improve the performance of neural network, and can the multiple second feature of station work extract networks, make Training difficulty decline, improves the training effectiveness of neural network.

In one possible implementation, according to the M the second training characteristics, M second feature of training extracts net Network, comprising:

According to the M the second training characteristics, station work N group second feature extracts network, and every group of second feature is extracted Network includes that the M second feature extracts at least one of network, and M, N are integer and 1 < N≤M.

In one possible implementation, according to the M the second training characteristics, station work N group second feature is mentioned Take network, comprising:

The markup information that corresponding second training characteristics of network and sample image are extracted according to i-th group of second feature, determines I-th group of second feature extracts the first network loss of network, and i is integer and 1≤i≤N；

The first network loss of network is extracted according to i-th group of second feature, training i-th group of second feature extracts net Network.

In one possible implementation, according to the M the second training characteristics, station work N group feature extraction net Network, comprising:

According to first training characteristics and the M the second training characteristics, N group second feature described in station work is extracted Network and the fisrt feature extract network.

In one possible implementation, in the training process that i-th group of second feature extracts network, k-th of training Characteristic distance between second training characteristics in period be greater than the feature between the second training characteristics of -1 cycle of training of kth away from From k is the integer greater than 1.

In this way, each second feature can be made to extract the different features that network extracts sample image respectively, reduced Multiple second feature extract a possibility that network extracts the same or similar second training characteristics, reduce multiple characteristic informations Between information redundancy, improve treatment effeciency, and it is richer to make multiple second feature extract the feature that networks extract, favorably In the accuracy for improving classification results.

In one possible implementation, described according to the M the second training characteristics, station work N group second After feature extraction network, the method also includes:

It is true respectively according to the markup information of first training characteristics, the M the second training characteristics and sample image The fixed fisrt feature extracts network and the M second feature extracts the third network losses of network；

Network is extracted according to the fisrt feature and the M second feature extracts the third network losses of network, adjustment The fisrt feature extracts network and the M second feature extracts the network parameter of network.

In one possible implementation, the method also includes:

Network handles processing image is extracted at least through the fisrt feature after the training and carries out feature extraction processing, is obtained Obtain the fourth feature of the image to be processed；

According to the fourth feature, the classification results to the target object in the image to be processed are obtained.

In one possible implementation, network handles processing figure is extracted at least through the fisrt feature after the training As carrying out feature extraction processing, the fifth feature of the image to be processed is obtained, comprising:

At least one second feature after extracting network and the training by the fisrt feature after the training extracts net Network carries out feature extraction processing to image to be processed, obtains at least one fourth feature of the image to be processed.

According to the one side of the disclosure, a kind of image classification device is provided, comprising:

First extraction module obtains the image to be processed for image to be processed input fisrt feature to be extracted network Fisrt feature；

Second extraction module extracts network for the fisrt feature to be inputted M second feature respectively, obtains M the Two features；

Module is obtained, for obtaining the classification results of the image to be processed according to the M second feature.

In one possible implementation, the fisrt feature is the default network in fisrt feature extraction network The output feature of level, second extraction module are further configured to:

According to the one side of the disclosure, a kind of feature extraction network training device is provided, comprising:

Third extraction module obtains the of the sample image for sample image input fisrt feature to be extracted network One training characteristics；

4th extraction module extracts network for first training characteristics to be inputted M second feature respectively, obtains M A second training characteristics；

Training module, for according to the M the second training characteristics, M second feature of training to extract network.

In one possible implementation, the training module is further configured to:

In one possible implementation, described device further include:

Determining module, for the mark according to first training characteristics, the M the second training characteristics and sample image Information is infused, determines the third network losses that the fisrt feature extracts network and the M second feature extracts network respectively；

Module is adjusted, for extracting the third of network and M second feature extraction network according to the fisrt feature Network losses adjust the network parameter that the fisrt feature extracts network and the M second feature extracts network.

In one possible implementation, described device further include:

5th extraction module, at least through after the training fisrt feature extract network handles handle image into Row feature extraction processing, obtains the fourth feature of the image to be processed；

Categorization module, for obtaining the classification to the target object in the image to be processed according to the fourth feature As a result.

In one possible implementation, the 5th extraction module is further configured to:

According to the one side of the disclosure, a kind of electronic equipment is provided, comprising:

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to: execute the above method.

According to the one side of the disclosure, a kind of computer readable storage medium is provided, computer program is stored thereon with Instruction, the computer program instructions realize the above method when being executed by processor.

It should be understood that above general description and following detailed description is only exemplary and explanatory, rather than Limit the disclosure.

According to below with reference to the accompanying drawings to detailed description of illustrative embodiments, the other feature and aspect of the disclosure will become It is clear.

Detailed description of the invention

The drawings herein are incorporated into the specification and forms part of this specification, and those figures show meet this public affairs The embodiment opened, and together with specification it is used to illustrate the technical solution of the disclosure.

Fig. 1 shows the flow chart of the image classification method according to the embodiment of the present disclosure；

Fig. 2 shows the flow charts according to the feature extraction network training method of the embodiment of the present disclosure；

Fig. 3 shows the application schematic diagram of the neural network training method according to the embodiment of the present disclosure；

Fig. 4 shows the block diagram of the image classification device according to the embodiment of the present disclosure；

Fig. 5 shows the block diagram of the feature extraction network training device according to the embodiment of the present disclosure；

Fig. 6 shows the block diagram of the electronic device according to the embodiment of the present disclosure；

Fig. 7 shows the block diagram of the electronic device according to the embodiment of the present disclosure.

Specific embodiment

Various exemplary embodiments, feature and the aspect of the disclosure are described in detail below with reference to attached drawing.It is identical in attached drawing Appended drawing reference indicate element functionally identical or similar.Although the various aspects of embodiment are shown in the attached drawings, remove It non-specifically points out, it is not necessary to attached drawing drawn to scale.

Dedicated word " exemplary " means " being used as example, embodiment or illustrative " herein.Here as " exemplary " Illustrated any embodiment should not necessarily be construed as preferred or advantageous over other embodiments.

The terms "and/or", only a kind of incidence relation for describing affiliated partner, indicates that there may be three kinds of passes System, for example, A and/or B, can indicate: individualism A exists simultaneously A and B, these three situations of individualism B.In addition, herein Middle term "at least one" indicate a variety of in any one or more at least two any combination, it may for example comprise A, B, at least one of C can indicate to include any one or more elements selected from the set that A, B and C are constituted.

In addition, giving numerous details in specific embodiment below to better illustrate the disclosure. It will be appreciated by those skilled in the art that without certain details, the disclosure equally be can be implemented.In some instances, for Method, means, element and circuit well known to those skilled in the art are not described in detail, in order to highlight the purport of the disclosure.

Fig. 1 shows the flow chart of the image classification method according to the embodiment of the present disclosure, as shown in Figure 1, which comprises

In step s 11, image to be processed input fisrt feature is extracted into network, obtains the first of the image to be processed Feature；

In step s 12, the fisrt feature is inputted into M second feature respectively and extracts network, obtains M second spy Sign；

In step s 13, according to the M second feature, the classification results of the image to be processed are obtained.

In one possible implementation, described image classification method can be held by terminal device or other processing equipments Row, wherein terminal device can be user equipment (User Equipment, UE), mobile device, user terminal, terminal, honeycomb Phone, wireless phone, personal digital assistant (Personal Digital Assistant, PDA), handheld device, calculate equipment, Mobile unit, wearable device etc..Other processing equipments can be server or cloud server etc..In some possible realization sides In formula, this method can be realized in such a way that processor calls the computer-readable instruction stored in memory.

In one possible implementation, in step s 11, the fisrt feature is extracted network and be can be with multiple The deep learning neural network of network level, such as convolutional neural networks etc. are also possible to other nets that can realize feature extraction Network, it is not limited here.It may include one or more target objects in the image to be processed, the target object can be People, vehicle, article etc..

In one possible implementation, image to be processed can be inputted fisrt feature and extracts network, fisrt feature mentions Feature extraction processing can successively be carried out by taking the multiple network layers grade of network, for example, first network level can to sample image into Row feature extraction, second network level can carry out feature extraction etc. to the feature that first network level obtains.In this example, As network level is deepened, the size for the feature that network level obtains reduces, and receptive field increases.

In one possible implementation, the fisrt feature is the default network in fisrt feature extraction network The output feature of level, that is, fisrt feature can be extracted to the feature that the default level in network obtains and be determined as first spy Sign, for example, the default level can be the intermediate level that the fisrt feature extracts network, for example, the fisrt feature mentions Taking network altogether includes ten network levels, can be using the 5th network level as the default level, and by the 5th network layer The feature that grade obtains is determined as the fisrt feature.The disclosure to fisrt feature extract the quantity of the network network level that includes with And default level is with no restrictions.

In one possible implementation, the fisrt feature can continue to input default in fisrt feature extraction network The subsequent network level of level is handled, and can get the feature of subsequent level output, for example, third feature in this example can Acquisition is more smaller than fisrt feature, but the feature that receptive field is bigger, for example, the port number of feature is more than fisrt feature, but every The aspect ratio fisrt feature in a channel is small.

In one possible implementation, in step s 12, the fisrt feature, which extracts network, can extract sample image A certain category feature, for example, the features such as key point of target object in extractable sample image.But in the appearance of target object Shoot that unintelligible, posture is complicated, proportion is smaller in the picture, background is more complex or target object is imperfect etc. in the picture In the case of, a single category feature is difficult to obtain higher classification accuracy.

In one possible implementation, network being extracted by multiple second feature, spy is carried out to sample image respectively Sign is extracted, and can extract the multiclass feature of sample image.For example, net can be extracted by M (M is greater than 1 integer) a second feature Network G₁,G₂...G_M, feature extraction processing is carried out to the fisrt feature of sample image respectively, obtains M second feature.

In one possible implementation, deep copy can be carried out to fisrt feature, for example, can to the first training characteristics into Row M times deep copy, obtains the copy of M fisrt feature.And the copy of M fisrt feature is inputted into M second feature respectively and is mentioned It takes network to carry out feature extraction processing, obtains M second feature.

In one possible implementation, in step s 13, it can be determined according to M second feature described to be processed The classification results of image, in this example, the classification results can be used for indicating that the classification of image to be processed, the disclosure tie classification Fruit is with no restrictions.

In this example, image to be processed and reference picture can be inputted respectively to fisrt feature and extract network, and by each second Feature extraction network obtains the second feature of image to be processed and the fixed reference feature of reference picture respectively, and determines image to be processed Second feature and reference picture fixed reference feature characteristic similarity (for example, cosine similarity).For example, if figure to be processed The characteristic similarity of the fixed reference feature of the second feature and reference picture of picture is greater than or equal to similarity threshold, then image to be processed In target object and reference picture in target object be same class, that is, image to be processed and reference picture can be classified as together One kind, conversely, image to be processed and reference picture can be classified as inhomogeneity.Alternatively, using second feature by image to be processed with Multiple reference pictures are compared, for example, determining the spy of the second feature of image to be processed and the fixed reference feature of each reference picture Similarity is levied, with the determining highest object reference feature of characteristic similarity with classification results, and by image to be processed and target The corresponding reference picture of fixed reference feature is classified as same class.Or by the second feature of multiple images to be processed respectively with one Or the fixed reference feature of multiple reference pictures is compared, to classify to multiple images to be processed, for example, one can be determined respectively The classification of each video frame in section video (will not include target for example, one kind can be classified as the video frame including target object The video frame of object is classified as another kind of etc.).

In one possible implementation, the fisrt feature, which extracts network, can also pass through the subsequent of the default level Level handles fisrt feature, obtains classification results.The step S13 can include: net is extracted by the fisrt feature The subsequent level that network level is preset described in network carries out feature extraction processing to the fisrt feature, obtains third feature；Root The classification results of the image to be processed are obtained according to the third feature and the M second feature.

In this example, M second feature can be obtained by extracting network by the M second feature, be mentioned by fisrt feature Take the default network level of network subsequent level can get third feature, and can by second feature and third feature come Determine the classification results of image to be processed.For example, image to be processed and reference picture can be inputted respectively to fisrt feature extracts net Network, and by fisrt feature extract network obtain image to be processed third feature, by each second feature extract network obtain respectively The second feature of image to be processed, and network is extracted by fisrt feature extraction network and second feature and obtains reference picture respectively Multiple fixed reference features.Further, can be compared respectively with corresponding fixed reference feature by third feature and M second feature It is right, to determine characteristic similarity, and it can determine whether image to be processed and reference picture belong to together according to the characteristic similarity It is a kind of.

Image classification according to an embodiment of the present disclosure can extract network by multiple second feature and extract sample graph respectively Various features as in, and network is extracted by fisrt feature and extracts third feature, keep characteristic information richer, improves at classification The accuracy rate of reason.

Fig. 2 shows the flow charts according to the feature extraction network training method of the embodiment of the present disclosure, as shown in Fig. 2, described Method includes:

In the step s 21, sample image input fisrt feature is extracted into network, obtains the first training of the sample image Feature；

In step S22, first training characteristics are inputted into M second feature respectively and extract network, obtain M second Training characteristics；

In step S23, according to the M the second training characteristics, M second feature of training extracts network.

In one possible implementation, in the step s 21, sample image can be inputted fisrt feature and extracts network, The multiple network layers grade that fisrt feature extracts network can successively carry out feature extraction processing, and the network level may include convolution Layer, active coating etc., in this example, it may include multiple network layers grade that fisrt feature, which extracts network, wherein fisrt feature extracts network The exportable sample image of default level (such as intermediate some level) the first training characteristics.

In one possible implementation, in step S22, the first training characteristics can be inputted respectively to M second spy Sign extracts network, obtains M the second training characteristics.It is copied deeply for example, M times can be carried out to the first training characteristics in the manner described above Shellfish obtains the copy of M the first training characteristics.And the copy of M the first training characteristics is inputted into M second feature respectively and is extracted Network carries out feature extraction processing, obtains M the second training characteristics.

It in one possible implementation, can be according to described M of M the second training characteristics training the in step S23 Two feature extraction networks.

In one possible implementation, there are many network parameter that M second feature extracts network, while the instruction of training White silk difficulty is higher, and operand is larger.Network can be extracted to M second feature to be grouped, and M second feature of station work mentions Take network, wherein step S13 can include: according to the M the second training characteristics, station work N group second feature extracts net Network, it includes that the M second feature extracts at least one of network that every group of second feature, which extracts network, and M, N are integer and 1 < N ≤M。

In this example, M second feature can be extracted to network and be divided into N (N is integer and 1 < N≤M) group, and be respectively trained each Group second feature extracts network and fisrt feature extracts network, for example, second feature is extracted network G₁,G₂..G_MIt is divided into N group, For example, first group is G₁₁、G₁₂……G_1Q(wherein, Q is positive integer), second group is G₂₁、G₂₂……G_2W(wherein, W is positive whole Number) ... i-th group is G_i1、G_i2……G_iE(wherein, E is positive integer) ... N group is G_N1、G_N2……G_NR(wherein, R is positive Integer).In this example, when training i-th (i is integer, and 1≤i≤N) group second feature extracts network, other groups second are special The network parameter that sign extracts network remains unchanged.

In one possible implementation, according to the M the second training characteristics, station work N group second feature is mentioned Take network can include: the markup information of corresponding second training characteristics of network and sample image is extracted according to i-th group of second feature, Determine that i-th group of second feature extracts the first network loss of network, i is integer and 1≤i≤N；It is mentioned according to i-th group of second feature The first network of network is taken to lose, training i-th group of second feature extracts network.

In one possible implementation, the fisrt feature that can be completed by training extracts network and extracts sample image First training characteristics, and each group second feature can be trained to extract network in batches.For example, the 1st group of second feature extraction network can It is trained in the 1st trained batch, the 2nd group of second feature is extracted network and can be trained in the 2nd trained batch ... N group second feature is extracted network and can be trained in n-th training batch.

In one possible implementation, by taking i-th of trained batch as an example, i-th of trained batch can train i-th group Two feature extraction networks.In this example, when i-th group of second feature of training extracts network, other group of second feature extracts network Network parameter remain unchanged.In this example, i-th of trained batch can extract network to i-th group of second feature and repeatedly be instructed Practice (for example, i-th of trained batch may include multiple cycles of training), that is, network and fisrt feature are extracted to i-th group of second feature Extract the adjusting that network carries out multiple network parameter.Each trained batch can extract net to the second feature that the batch to be trained Network carries out the adjusting of multiple network parameter.

In one possible implementation, corresponding second training characteristics of network can be extracted according to i-th group of second feature And the markup information of sample image, determine that i-th group of second feature extracts the network losses of i-th of trained batch of network.Showing In example, fisrt feature can be extracted to the feature of l (l is positive integer) layer of network as the first training characteristics h^l, the M A second feature extracts network can be respectively to the first training characteristics h^lIt is handled, obtains the second training characteristics f respectively₁、f₂…… f_M, wherein it is G that i-th group of second feature, which extracts network,_i1、G_i2……G_iE, the second instruction of i-th group of second feature extraction network acquisition Practicing feature is f_i1、f_i2……f_iE.The second training characteristics f that network obtains can be extracted according to i-th group of second feature_i1、f_i2…… f_iEAnd the markup information of sample image determines that i-th group of second feature extracts the network losses L of network.

In this example, the markup information of sample image can be the characteristic similarity of sample image and reference picture, show In example, if the target object in sample image and reference picture is same people, the feature phase of sample image and reference picture It is labeled for 1 like degree, if the target object in sample image and reference picture is not same people, sample image and reference The characteristic similarity of image is labeled for 0.

In this example, network can be extracted by fisrt feature and M second feature extracts network and extract reference picture respectively Feature, and determine i-th group each second feature extract network extract sample image the second training characteristics and reference picture Characteristic similarity between second training characteristics.And i-th group second is determined according to the difference of this feature similarity and markup information The network losses of feature extraction network.For example, can determine that i-th group of the corresponding feature of each second feature extraction network is similar respectively Difference between degree and the mark of sample image, and the difference is determined that i-th group of second feature extracts the first network of network Lose L.In this example, the characteristic similarity can be indicated with cosine similarity, for example, output feature and reference picture Export the cosine similarity between feature.The disclosure to the determination methods of network losses with no restrictions.

In this example, the first network that network can be extracted according to i-th group of second feature loses L, i-th group of second feature of training Extract network.For example, it may be determined that the first network loss L extracts i-th group of second feature the gradient of each parameter of network(p indicates that i-th group of second feature extracts any network parameter of network), and according to the direction for minimizing network losses, lead to Gradient descent method is crossed to adjust each network parameter.

In one possible implementation, in addition to according to make network losses minimize direction regulating networks parameter it Outside, distance restraint can also be added.Characteristic distance between the second training characteristics of k-th of cycle of training is greater than -1 training of kth Characteristic distance between second training characteristics in period, k are the integer greater than 1.

It in this example, can be according to making i-th group of second feature extract network by taking i-th group of second feature extracts network as an example The direction that the characteristic distance between the second training characteristics obtained increases, carrys out regulating networks parameter.For example, i-th of trained batch Network is extracted to i-th group of second feature and has carried out multiple training (that is, multiple adjusting has been carried out to network parameter), wherein in kth After secondary regulating networks parameter (k-th of cycle of training), i-th group of second feature is extracted between the second training characteristics that network obtains Characteristic distance is greater than the characteristic distance between a preceding regulating networks parameter (- 1 cycle of training of kth) afterwards the second training characteristics. The characteristic distance may include Euclidean distance or COS distance etc., the disclosure to the definition of characteristic distance with no restrictions.Addition away from , can be by training from constraint, the parameter differences for extracting each second feature between network are increasing, and each second feature extracts net The difference between the second training characteristics that network extracts is increasing, that is, so that each second feature is extracted network and extracts sample graph respectively The different features of picture.

In one possible implementation, each cycle of training of i-th of trained batch can be determined through the above way Network losses, and network and i-th group of second feature are extracted to fisrt feature according to network losses in each cycle of training and mentioned It takes the network parameter of network to be adjusted, and in the case where the training condition of i-th of satisfaction trained batch, completes i-th of instruction The training for practicing batch, followed by the training of i+1 training batch, that is, i-th group of second feature extracts network and meet training When condition, the training that i-th group of second feature extracts network is completed, the training that i+1 group second feature extracts network is then carried out. Wherein, the training condition of described i-th trained batch may include frequency of training (that is, number cycle of training), for example, working as frequency of training When reaching preset times, meet the training condition of i-th of trained batch.Alternatively, the training condition of described i-th trained batch can Size including network losses holds back scattered property, for example, when the network losses are less than or equal to loss threshold value or converge on default When in section, meet the training condition of i-th of trained batch.The disclosure to training condition with no restrictions.

It in one possible implementation, can be identical by extracting the training method of network with i-th group of second feature Method carries out the training that multiple groups second feature extracts network, each group second feature can be extracted network training and completed.At every group In the training of two feature extraction networks, the network parameter that the corresponding second feature of the batch extracts network is only adjusted, other second The network parameter of feature extraction network remains unchanged.

In one possible implementation, fisrt feature can be extracted to network and extracts network with M second feature, jointly It is trained, wherein according to the M the second training characteristics, station work N group feature extraction network, comprising: according to described the One training characteristics and the M the second training characteristics, N group second feature extraction network and the fisrt feature described in station work Extract network.In this example, network can be extracted to N group second feature and the fisrt feature is extracted network and trained in batches, that is, Station work.For example, the 1st group of second feature extracts network and fisrt feature is extracted network and can be carried out in the 1st trained batch Training, the 2nd group of second feature extracts network and fisrt feature is extracted network and can be trained in the 2nd trained batch ... N Group second feature extracts network and fisrt feature is extracted network and can be trained in n-th training batch.

In one possible implementation, by taking i-th of trained batch as an example, net can be extracted according to i-th group of second feature The markup information of corresponding second training characteristics of network and sample image determines that i-th group of second feature extracts network and described first The network losses of the trained batch of i-th of feature extraction network.For example, fisrt feature extracts the subsequent of the default level of network Network level can continue with the first training characteristics h^l, can get the output feature h that fisrt feature extracts network^l+n(n first The quantity of network level after the default level of feature extraction network), it can be according to h^l+n, i-th group of second feature extract network and obtain The the second training characteristics f obtained_i1、f_i2……f_iEAnd the markup information of sample image determines that fisrt feature extracts network and the The third network losses of i group second feature extraction network.And network and i-th group of second feature extraction net are extracted according to fisrt feature The third network losses of network extract the network parameter of network to adjust fisrt feature extraction network and i-th group of second feature.For example, Each network parameter can be adjusted by gradient descent method.And network and fisrt feature extraction network are extracted in i-th group of second feature When meeting training condition, the training that i-th group of second feature extracts network is completed, then i+1 group second feature is carried out and extracts net The training of network.Further, when adjusting fisrt feature extraction network and i-th group of second feature extracts the network parameter of network, The network parameter that other second feature extract network remains unchanged.

In one possible implementation, network is extracted in each group second feature and fisrt feature extraction network training is complete All second feature, can also be extracted network by Cheng Hou and fisrt feature extracts network and carries out parameter adjustment jointly, in the basis The M the second training characteristics, after station work N group second feature extraction network, the method also includes: according to described the The markup information of one training characteristics, the M the second training characteristics and sample image determines that the fisrt feature is extracted respectively Network and the M second feature extract the third network losses of network；Network and the M are extracted according to the fisrt feature Second feature extracts the third network losses of network, adjusts the fisrt feature and extracts network and M second feature extraction The network parameter of network.

In one possible implementation, each group second feature can not be repartitioned and extract network, and according to fisrt feature It extracts network and all second feature extracts the output of network (that is, M second feature extracts network) and the mark of sample image It infuses information and determines third network losses jointly, and network and all second feature are extracted to fisrt feature according to third network losses The network parameter for extracting network is adjusted.It is less than preset threshold in adjustment preset times or the network losses or converges on pre- If when section, the fisrt feature after can get training extracts network and M second feature extracts network.The adjustment amplitude of the adjustment The adjustment amplitude of the network parameter of network and fisrt feature extraction network is extracted when being smaller than station work to each group second feature, That is, the whole fine tuning of all feature extraction networks can be carried out.And the adjustment can make fisrt feature extract the second spy of network and each group The characteristic extraction procedure that sign extracts network is more coordinated.Fisrt feature after training extracts network and second feature is extracted network and be can be used During carrying out classification processing to image.

Neural network training method according to an embodiment of the present disclosure can extract network by multiple second feature and mention respectively Take the various features in sample image, improve the accuracy of neural network, and can the multiple second feature of station work extract networks, Make to train difficulty decline, improves the training effectiveness of neural network.And each second feature can be made to extract network and extract sample graph respectively The different features of picture reduce multiple second feature and extract the possibility that network extracts the same or similar second training characteristics Property so that multiple second feature extract network extract feature it is richer, be conducive to target object appearance shooting it is unintelligible, Posture is complicated, proportion is smaller in the picture, background is more complex or target object it is imperfect in the picture when, obtain Various features improve the accuracy of classification results.Further, compared to only fisrt feature being trained to extract network, by multiple Second feature is extracted network and is trained jointly with fisrt feature extraction network, and the performance that fisrt feature extracts network can be promoted, Fisrt feature can be used only after training and extract network to image progress classification processing, promote fisrt feature and extract Web vector graphic Flexibility, and promote the classification accuracy that fisrt feature extracts network.

In one possible implementation, the fisrt feature after training can be used extracts network and second feature extracts net Network carries out classification processing to image to be processed.In this example, network can be extracted without using second feature, is mentioned using only fisrt feature It takes network to carry out classification processing, network can also be extracted without using whole second feature, fisrt feature is used only and extracts network and portion Divide second feature to extract network and carries out classification processing.The method also includes: at least through the fisrt feature after the training It extracts network handles processing image and carries out feature extraction processing, obtain the fourth feature of the image to be processed；According to described Four features obtain the classification results to the target object in the image to be processed.

It in one possible implementation, may include one or more target objects in the image to be processed, it is described Target object can be people, vehicle, article etc..Fisrt feature after training is extracted network and can be mentioned to image to be processed progress feature It takes.

In this example, network handles only can be extracted by fisrt feature and handles image progress feature extraction, fisrt feature mentions The fourth feature can be confirmed as by taking the output result of network.Network can also be extracted by fisrt feature and second feature mentions Network handles processing image is taken to carry out feature extraction.For example, some or all of can be extracted by M second feature in network with And fisrt feature extracts network handles processing image and carries out feature extraction, obtains multiple fourth feature.At least through the training Fisrt feature afterwards extracts network handles processing image and carries out feature extraction processing, and obtain the image to be processed the 5th is special Sign, comprising: at least one second feature after extracting network and the training by the fisrt feature after the training extracts net Network carries out feature extraction processing to image to be processed, obtains at least one fourth feature of the image to be processed.

In one possible implementation, network only can be extracted by fisrt feature, feature is carried out to image to be processed Extraction process obtains fourth feature.

In one possible implementation, image to be processed can be inputted fisrt feature and extracts network, fisrt feature mentions Image to be processed can be handled by taking each network level of network, and fisrt feature can be extracted network default level the One training characteristics carry out deep copy, and the copy input second feature for the first training characteristics that copy is obtained extracts network.And Network can be extracted by M second feature and fisrt feature extracts network handles and handles image progress feature extraction.It can be by first The fisrt feature of the default level of feature extraction network carries out deep copy, obtains M fisrt feature, and input M second respectively Feature extraction network.Fisrt feature extracts network and M second feature extraction network can be respectively to fisrt feature or fisrt feature Copy carry out feature extraction processing, and obtain fourth feature.For example, can get M+1 fourth feature.

In one possible implementation, a part and fisrt feature in network can be extracted by M second feature It extracts network handles processing image and carries out feature extraction, for example, one or more second feature can be used to extract network and first Feature extraction network handles handle image and carry out feature extraction.Fisrt feature extracts network and one or more second feature are extracted Network can carry out feature extraction processing by the copy to fisrt feature or fisrt feature respectively, and obtain multiple fourth feature.

In one possible implementation, the target object in image to be processed is determined using the fourth feature Classification results.For example, it may be determined that the characteristic similarity between fourth feature and the feature of reference picture is (for example, cosine similarity Deng), if the characteristic similarity is greater than or equal to similarity threshold, one can will be classified as with reference picture in image to be processed Class.Alternatively, image to be processed is compared with multiple reference pictures using fourth feature, for example, determine fourth feature with The characteristic similarity of the fixed reference feature of each reference picture, with determining special with the fourth feature highest object reference of characteristic similarity Sign, and image to be processed reference picture corresponding with object reference feature is classified as same class.Or by multiple figures to be processed The fourth feature of picture is compared with one or more reference pictures respectively, to classify to multiple images to be processed, for example, Can determine respectively each video frame in one section of video classification (for example, can by include target object video frame be classified as one kind, By do not include target object video frame be classified as it is another kind of etc.).

Fig. 3 shows the application schematic diagram of the neural network training method according to the embodiment of the present disclosure, as shown in figure 3, in mind In training process through network, the sample image input fisrt feature including one or more target objects can be extracted network, The multiple network layers grade that fisrt feature extracts network can successively carry out feature extraction processing, and obtain the in first of network level One training characteristics h^l。

It in one possible implementation, can be by the first training characteristics h^lDeep copy is carried out, and by the M of acquisition first The copy of training characteristics inputs M second feature respectively and extracts network G₁,G₂..G_M, M the second training characteristics f are obtained respectively₁、 f₂……f_M.Also, the subsequent network level that fisrt feature extracts first of network level of network can continue special to the first training Sign is handled, and the output feature h that fisrt feature extracts network is obtained^l+n.Further, using the second training characteristics f₁、 f₂……f_MAnd fisrt feature extracts the output feature h of network^l+nStation work second feature extracts network G₁,G₂..G_MWith first Feature extraction network.

In one possible implementation, second feature can be extracted network G₁,G₂..G_MIt is divided into N group, for example, first Group is G₁₁、G₁₂……G_1Q, second group is G₂₁、G₂₂……G_2W, i-th group is G_i1、G_i2……G_iE, N group is G_N1、G_N2…… G_NR.Every group of second feature can be respectively trained and extract network and fisrt feature extraction network.

In one possible implementation, by taking i-th of trained batch as an example, i-th of trained batch can train i-th group Two feature extraction networks, when i-th group of second feature of training extracts network, other group of second feature extracts the network parameter of network It remains unchanged.When i-th group of second feature of training extracts network, the output feature h of network can be extracted according to fisrt feature^l+n, I group second feature extracts the second training characteristics f of network output_i1、f_i2……f_iEAnd the markup information of sample image determines I-th group of second feature extracts network and fisrt feature extracts the network losses of network, and adjusts fisrt feature according to network losses It extracts network and i-th group of second feature extracts the network parameter of network, for example, being adjusted according to the direction for minimizing network losses Network parameter.And when adjusting network parameter, distance restraint can be added, that is, can obtain according to making i-th group of second feature extract network The direction that characteristic distance between the second training characteristics taken increases, carrys out regulating networks parameter, each second feature extract network it Between parameter differences it is increasing, the difference that each second feature is extracted between the second training characteristics that network extracts is increasing, That is, each second feature is made to extract the different features that network extracts sample image respectively, be conducive to the diversification for the feature extracted. In the training condition of i-th of satisfaction trained batch, the training of i-th of batch is completed.

In one possible implementation, every group of second feature can be respectively trained by the above method and extract network and the One feature extraction network.And after the completion of each group second feature extracts network and fisrt feature and extracts network training, by all the Two feature extraction networks and fisrt feature extract network and carry out parameter adjustment jointly, and the fisrt feature after being trained extracts network Network is extracted with second feature.

In one possible implementation, to image to be processed carry out classification processing when, can be at least through training after Fisrt feature extract network and extract the fourth feature of image to be processed.Network and one or more can also be extracted by fisrt feature A second feature extracts the fourth feature that network extracts image to be processed respectively.Further, it can be determined by fourth feature The classification results of target object in image to be processed.For example, according to the feature between fourth feature and the feature of reference picture Similarity determines the classification results of image to be processed.

In one possible implementation, the neural network training method and image classification method can be used for video In the classification processing of frame, for example, inquiring some pedestrian in monitor video, network and one can be extracted by fisrt feature A or multiple second feature extract network to determine the classification of each video frame in monitor video, that is, by the video including the pedestrian Frame is divided into one kind, by do not include the pedestrian video frame be divided into it is another kind of etc..The disclosure to the neural network training method and Image classification method application field is with no restrictions.

It is appreciated that above-mentioned each embodiment of the method that the disclosure refers to, without prejudice to principle logic, To engage one another while the embodiment to be formed after combining, as space is limited, the disclosure is repeated no more.

In addition, the disclosure additionally provides image classification device and feature extraction network training device, electronic equipment, computer Readable storage medium storing program for executing, program, the above-mentioned any method that can be used to realize the disclosure and provide, corresponding technical solution and description and Referring to the corresponding record of method part, repeat no more.

It will be understood by those skilled in the art that each step writes sequence simultaneously in the above method of specific embodiment It does not mean that stringent execution sequence and any restriction is constituted to implementation process, the specific execution sequence of each step should be with its function It can be determined with possible internal logic.

Fig. 4 shows the block diagram of the image classification device according to the embodiment of the present disclosure, as shown in figure 4, described device includes:

First extraction module 11 obtains the figure to be processed for image to be processed input fisrt feature to be extracted network The fisrt feature of picture；

Second extraction module 12 extracts network for the fisrt feature to be inputted M second feature respectively, obtains M Second feature；

Module 13 is obtained, for obtaining the classification results of the image to be processed according to the M second feature.

Fig. 5 shows the block diagram of the feature extraction network training device according to the embodiment of the present disclosure, as shown in figure 5, the dress It sets and includes:

Third extraction module 21 obtains the sample image for sample image input fisrt feature to be extracted network First training characteristics；

4th extraction module 22 extracts network for first training characteristics to be inputted M second feature respectively, obtains M the second training characteristics；

Training module 23, for according to the M the second training characteristics, M second feature of training to extract network.

In one possible implementation, the training module is further configured to:

In one possible implementation, described device further include:

In some embodiments, the embodiment of the present disclosure provides the function that has of device or comprising module can be used for holding The method of row embodiment of the method description above, specific implementation are referred to the description of embodiment of the method above, for sake of simplicity, this In repeat no more

The embodiment of the present disclosure also proposes a kind of computer readable storage medium, is stored thereon with computer program instructions, institute It states when computer program instructions are executed by processor and realizes the above method.Computer readable storage medium can be non-volatile meter Calculation machine readable storage medium storing program for executing.

The embodiment of the present disclosure also proposes a kind of electronic equipment, comprising: processor；For storage processor executable instruction Memory；Wherein, the processor is configured to the above method.

The equipment that electronic equipment may be provided as terminal, server or other forms.

Fig. 6 is the block diagram of a kind of electronic equipment 800 shown according to an exemplary embodiment.For example, electronic equipment 800 can To be mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, Medical Devices are good for Body equipment, the terminals such as personal digital assistant.

Referring to Fig. 6, electronic equipment 800 may include following one or more components: processing component 802, memory 804, Power supply module 806, multimedia component 808, audio component 810, the interface 812 of input/output (I/O), sensor module 814, And communication component 816.

The integrated operation of the usual controlling electronic devices 800 of processing component 802, such as with display, call, data are logical Letter, camera operation and record operate associated operation.Processing component 802 may include one or more processors 820 to hold Row instruction, to perform all or part of the steps of the methods described above.In addition, processing component 802 may include one or more moulds Block, convenient for the interaction between processing component 802 and other assemblies.For example, processing component 802 may include multi-media module, with Facilitate the interaction between multimedia component 808 and processing component 802.

Memory 804 is configured as storing various types of data to support the operation in electronic equipment 800.These data Example include any application or method for being operated on electronic equipment 800 instruction, contact data, telephone directory Data, message, picture, video etc..Memory 804 can by any kind of volatibility or non-volatile memory device or it Combination realize, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM) is erasable Except programmable read only memory (EPROM), programmable read only memory (PROM), read-only memory (ROM), magnetic memory, fastly Flash memory, disk or CD.

Power supply module 806 provides electric power for the various assemblies of electronic equipment 800.Power supply module 806 may include power supply pipe Reason system, one or more power supplys and other with for electronic equipment 800 generate, manage, and distribute the associated component of electric power.

Multimedia component 808 includes the screen of one output interface of offer between the electronic equipment 800 and user. In some embodiments, screen may include liquid crystal display (LCD) and touch panel (TP).If screen includes touch surface Plate, screen may be implemented as touch screen, to receive input signal from the user.Touch panel includes one or more touches Sensor is to sense the gesture on touch, slide, and touch panel.The touch sensor can not only sense touch or sliding The boundary of movement, but also detect duration and pressure associated with the touch or slide operation.In some embodiments, Multimedia component 808 includes a front camera and/or rear camera.When electronic equipment 800 is in operation mode, as clapped When taking the photograph mode or video mode, front camera and/or rear camera can receive external multi-medium data.It is each preposition Camera and rear camera can be a fixed optical lens system or have focusing and optical zoom capabilities.

Audio component 810 is configured as output and/or input audio signal.For example, audio component 810 includes a Mike Wind (MIC), when electronic equipment 800 is in operation mode, when such as call mode, recording mode, and voice recognition mode, microphone It is configured as receiving external audio signal.The received audio signal can be further stored in memory 804 or via logical Believe that component 816 is sent.In some embodiments, audio component 810 further includes a loudspeaker, is used for output audio signal.

I/O interface 812 provides interface between processing component 802 and peripheral interface module, and above-mentioned peripheral interface module can To be keyboard, click wheel, button etc..These buttons may include, but are not limited to: home button, volume button, start button and lock Determine button.

Sensor module 814 includes one or more sensors, for providing the state of various aspects for electronic equipment 800 Assessment.For example, sensor module 814 can detecte the state that opens/closes of electronic equipment 800, the relative positioning of component, example As the component be electronic equipment 800 display and keypad, sensor module 814 can also detect electronic equipment 800 or The position change of 800 1 components of electronic equipment, the existence or non-existence that user contacts with electronic equipment 800, electronic equipment 800 The temperature change of orientation or acceleration/deceleration and electronic equipment 800.Sensor module 814 may include proximity sensor, be configured For detecting the presence of nearby objects without any physical contact.Sensor module 814 can also include optical sensor, Such as CMOS or ccd image sensor, for being used in imaging applications.In some embodiments, which may be used also To include acceleration transducer, gyro sensor, Magnetic Sensor, pressure sensor or temperature sensor.

Communication component 816 is configured to facilitate the communication of wired or wireless way between electronic equipment 800 and other equipment. Electronic equipment 800 can access the wireless network based on communication standard, such as WiFi, 2G or 3G or their combination.Show at one In example property embodiment, communication component 816 receives broadcast singal or broadcast from external broadcasting management system via broadcast channel Relevant information.In one exemplary embodiment, the communication component 816 further includes near-field communication (NFC) module, short to promote Cheng Tongxin.For example, radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra wide band can be based in NFC module (UWB) technology, bluetooth (BT) technology and other technologies are realized.

In the exemplary embodiment, electronic equipment 800 can be by one or more application specific integrated circuit (ASIC), number Word signal processor (DSP), digital signal processing appts (DSPD), programmable logic device (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic components are realized, for executing the above method.

In the exemplary embodiment, a kind of non-volatile computer readable storage medium storing program for executing is additionally provided, for example including calculating The memory 804 of machine program instruction, above-mentioned computer program instructions can be executed by the processor 820 of electronic equipment 800 to complete The above method.

Fig. 7 is the block diagram of a kind of electronic equipment 1900 shown according to an exemplary embodiment.For example, electronic equipment 1900 It may be provided as a server.Referring to Fig. 7, electronic equipment 1900 includes processing component 1922, further comprise one or Multiple processors and memory resource represented by a memory 1932, can be by the execution of processing component 1922 for storing Instruction, such as application program.The application program stored in memory 1932 may include it is one or more each Module corresponding to one group of instruction.In addition, processing component 1922 is configured as executing instruction, to execute the above method.

Electronic equipment 1900 can also include that a power supply module 1926 is configured as executing the power supply of electronic equipment 1900 Management, a wired or wireless network interface 1950 is configured as electronic equipment 1900 being connected to network and an input is defeated (I/O) interface 1958 out.Electronic equipment 1900 can be operated based on the operating system for being stored in memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or similar.

In the exemplary embodiment, a kind of non-volatile computer readable storage medium storing program for executing is additionally provided, for example including calculating The memory 1932 of machine program instruction, above-mentioned computer program instructions can by the processing component 1922 of electronic equipment 1900 execute with Complete the above method.

The disclosure can be system, method and/or computer program product.Computer program product may include computer Readable storage medium storing program for executing, containing for making processor realize the computer-readable program instructions of various aspects of the disclosure.

Computer readable storage medium, which can be, can keep and store the tangible of the instruction used by instruction execution equipment Equipment.Computer readable storage medium for example can be-- but it is not limited to-- storage device electric, magnetic storage apparatus, optical storage Equipment, electric magnetic storage apparatus, semiconductor memory apparatus or above-mentioned any appropriate combination.Computer readable storage medium More specific example (non exhaustive list) includes: portable computer diskette, hard disk, random access memory (RAM), read-only deposits It is reservoir (ROM), erasable programmable read only memory (EPROM or flash memory), static random access memory (SRAM), portable Compact disk read-only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical coding equipment, for example thereon It is stored with punch card or groove internal projection structure and the above-mentioned any appropriate combination of instruction.Calculating used herein above Machine readable storage medium storing program for executing is not interpreted that instantaneous signal itself, the electromagnetic wave of such as radio wave or other Free propagations lead to It crosses the electromagnetic wave (for example, the light pulse for passing through fiber optic cables) of waveguide or the propagation of other transmission mediums or is transmitted by electric wire Electric signal.

Computer-readable program instructions as described herein can be downloaded to from computer readable storage medium it is each calculate/ Processing equipment, or outer computer or outer is downloaded to by network, such as internet, local area network, wide area network and/or wireless network Portion stores equipment.Network may include copper transmission cable, optical fiber transmission, wireless transmission, router, firewall, interchanger, gateway Computer and/or Edge Server.Adapter or network interface in each calculating/processing equipment are received from network to be counted Calculation machine readable program instructions, and the computer-readable program instructions are forwarded, for the meter being stored in each calculating/processing equipment In calculation machine readable storage medium storing program for executing.

Computer program instructions for executing disclosure operation can be assembly instruction, instruction set architecture (ISA) instructs, Machine instruction, machine-dependent instructions, microcode, firmware instructions, condition setup data or with one or more programming languages The source code or object code that any combination is write, the programming language include the programming language-of object-oriented such as Smalltalk, C++ etc., and conventional procedural programming languages-such as " C " language or similar programming language.Computer Readable program instructions can be executed fully on the user computer, partly execute on the user computer, be only as one Vertical software package executes, part executes on the remote computer or completely in remote computer on the user computer for part Or it is executed on server.In situations involving remote computers, remote computer can pass through network-packet of any kind It includes local area network (LAN) or wide area network (WAN)-is connected to subscriber computer, or, it may be connected to outer computer (such as benefit It is connected with ISP by internet).In some embodiments, by utilizing computer-readable program instructions Status information carry out personalized customization electronic circuit, such as programmable logic circuit, field programmable gate array (FPGA) or can Programmed logic array (PLA) (PLA), the electronic circuit can execute computer-readable program instructions, to realize each side of the disclosure Face.

Referring herein to according to the flow chart of the method, apparatus (system) of the embodiment of the present disclosure and computer program product and/ Or block diagram describes various aspects of the disclosure.It should be appreciated that flowchart and or block diagram each box and flow chart and/ Or in block diagram each box combination, can be realized by computer-readable program instructions.

These computer-readable program instructions can be supplied to general purpose computer, special purpose computer or other programmable datas The processor of processing unit, so that a kind of machine is produced, so that these instructions are passing through computer or other programmable datas When the processor of processing unit executes, function specified in one or more boxes in implementation flow chart and/or block diagram is produced The device of energy/movement.These computer-readable program instructions can also be stored in a computer-readable storage medium, these refer to It enables so that computer, programmable data processing unit and/or other equipment work in a specific way, thus, it is stored with instruction Computer-readable medium then includes a manufacture comprising in one or more boxes in implementation flow chart and/or block diagram The instruction of the various aspects of defined function action.

Computer-readable program instructions can also be loaded into computer, other programmable data processing units or other In equipment, so that series of operation steps are executed in computer, other programmable data processing units or other equipment, to produce Raw computer implemented process, so that executed in computer, other programmable data processing units or other equipment Instruct function action specified in one or more boxes in implementation flow chart and/or block diagram.

The flow chart and block diagram in the drawings show system, method and the computer journeys according to multiple embodiments of the disclosure The architecture, function and operation in the cards of sequence product.In this regard, each box in flowchart or block diagram can generation One module of table, program segment or a part of instruction, the module, program segment or a part of instruction include one or more use The executable instruction of the logic function as defined in realizing.In some implementations as replacements, function marked in the box It can occur in a different order than that indicated in the drawings.For example, two continuous boxes can actually be held substantially in parallel Row, they can also be executed in the opposite order sometimes, and this depends on the function involved.It is also noted that block diagram and/or The combination of each box in flow chart and the box in block diagram and or flow chart, can the function as defined in executing or dynamic The dedicated hardware based system made is realized, or can be realized using a combination of dedicated hardware and computer instructions.

The presently disclosed embodiments is described above, above description is exemplary, and non-exclusive, and It is not limited to disclosed each embodiment.Without departing from the scope and spirit of illustrated each embodiment, for this skill Many modifications and changes are obvious for the those of ordinary skill in art field.The selection of term used herein, purport In the principle, practical application or technological improvement to the technology in market for best explaining each embodiment, or lead this technology Other those of ordinary skill in domain can understand each embodiment disclosed herein.

Claims

1. a kind of image classification method characterized by comprising

2. the method according to claim 1, wherein the fisrt feature is that the fisrt feature is extracted in network Default network level output feature, it is described according to the M second feature, obtain the classification knot of the image to be processed Fruit, comprising:

The subsequent level that network level is preset described in network is extracted by the fisrt feature to fisrt feature progress spy Extraction process is levied, third feature is obtained；

3. a kind of feature extraction network training method characterized by comprising

4. according to the method described in claim 3, it is characterized in that, training M second according to the M the second training characteristics Feature extraction network, comprising:

According to the M the second training characteristics, station work N group second feature extracts network, and every group of second feature extracts network At least one of network is extracted including the M second feature, M, N are integer and 1 < N≤M.

5. according to the method described in claim 4, it is characterized in that, according to the M the second training characteristics, station work N group Second feature extracts network, comprising:

The markup information that corresponding second training characteristics of network and sample image are extracted according to i-th group of second feature, determines i-th group Second feature extracts the first network loss of network, and i is integer and 1≤i≤N；

The first network loss of network is extracted according to i-th group of second feature, training i-th group of second feature extracts network.

6. according to the method described in claim 4, it is characterized in that, according to the M the second training characteristics, station work N group Feature extraction network, comprising:

According to first training characteristics and the M the second training characteristics, N group second feature described in station work extracts network And the fisrt feature extracts network.

7. a kind of image classification device characterized by comprising

First extraction module obtains the of the image to be processed for image to be processed input fisrt feature to be extracted network One feature；

Second extraction module extracts network for the fisrt feature to be inputted M second feature respectively, obtains M second spy Sign；

8. a kind of feature extraction network training device characterized by comprising

Third extraction module obtains the first instruction of the sample image for sample image input fisrt feature to be extracted network Practice feature；

4th extraction module extracts network for first training characteristics to be inputted M second feature respectively, obtains M the Two training characteristics；

9. a kind of electronic equipment characterized by comprising

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to: perform claim require any one of 1 to 6 described in method.

10. a kind of computer readable storage medium, is stored thereon with computer program instructions, which is characterized in that the computer Method described in any one of claim 1 to 6 is realized when program instruction is executed by processor.