CN108304882A

CN108304882A - A kind of image classification method, device and server, user terminal, storage medium

Info

Publication number: CN108304882A
Application number: CN201810124834.1A
Authority: CN
Inventors: 陈承; 冯晓冰; 褚攀; 徐浩
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Shenzhen Yayue Technology Co ltd
Priority date: 2018-02-07
Filing date: 2018-02-07
Publication date: 2018-07-20
Anticipated expiration: 2038-02-07
Also published as: CN108304882B; WO2019154262A1

Abstract

The embodiment of the invention discloses a kind of image classification method, device and server, user terminal, storage mediums, wherein the method includes：Obtain image to be classified, and calling classification model classifies to image to be classified, determine the initial category belonging to image to be classified, Feature Selection Model is called to determine the target feature vector of image to be classified, and then the target feature vector of image to be classified is compared with the candidate classification feature vector configured for initial category, comparing result is obtained, and determines the classification results of image to be classified according to comparing result.The accuracy of image classification can be effectively improved using the embodiment of the present invention.

Description

A kind of image classification method, device and server, user terminal, storage medium

Technical field

This application involves a kind of technical field of image processing more particularly to image classification method, device and server, users Terminal, storage medium.

Background technology

In people's lives, various unknown images are frequently encountered, it, at all can not be with for layman By means of the classification visually identified belonging to these images, it is likely that will appear the embarrassment of " calling a stag a horse ".

With the development of image recognition technology so that people have broken away from the mode of traditional eye recognition image, greatly save It has saved the time of identification image and has improved the efficiency of identification.But how more fast and accurately to various unknown images into point Class is still the hot spot of a research.

Invention content

An embodiment of the present invention provides a kind of image classification method, device and server, user terminal, storage mediums, can Fast accurately classify to image.

On the one hand, an embodiment of the present invention provides a kind of image classification methods, including：

Obtain image to be classified；

Calling classification model classifies to the image to be classified, determines the initial classes belonging to the image to be classified Not；

Feature Selection Model is called to determine the target feature vector of the image to be classified；

By the target feature vector of the image to be classified and the candidate classification feature vector for initial category configuration It is compared, obtains comparing result；

The classification results of the image to be classified are determined according to comparing result.

On the other hand, an embodiment of the present invention provides another image classification methods, including：

When detecting the trigger action for recognition button, camera assembly is called to obtain present preview image；

If detecting that the present preview image is in stable state, the present preview image is determined as waiting for point Class image；

Classification request is sent to server, the image to be classified is carried in the classification request；

The response message that the server returns is received, the response message includes the classification knot of the image to be classified Fruit, the classification results are that the server classifies to the image to be classified according to disaggregated model and Feature Selection Model It is obtained after determination.

In another aspect, an embodiment of the present invention provides a kind of image classification devices, including：

Acquisition module, for obtaining image to be classified；

Calling module classifies to the image to be classified for calling classification model, determines the figure to be sorted As affiliated initial category；

The calling module, be additionally operable to call Feature Selection Model determine the target signature of the image to be classified to Amount；

Contrast module is used for the target feature vector of the image to be classified and the candidate for initial category configuration Characteristic of division vector is compared, and comparing result is obtained；

Determining module, the classification results for determining the image to be classified according to comparing result.

Another aspect, an embodiment of the present invention provides another image classification devices, including：

Detection module, for detecting the trigger action for recognition button；

Acquisition module when detecting the trigger action for recognition button for the detection module, calls camera assembly Obtain present preview image；

The detection module is additionally operable to detect whether the present preview image is in stable state；

Determining module will if detecting that the present preview image is in stable state for the detection module The present preview image is determined as image to be classified；

Sending module carries the image to be classified for sending classification request to server in the classification request；

Receiving module, the response message returned for receiving the server, the response message includes described to be sorted The classification results of image, the classification results be the server according to disaggregated model and Feature Selection Model to described to be sorted Image obtained after classification determines.

Correspondingly, the embodiment of the present invention additionally provides a kind of server, including：Processor and storage device；The storage Device, for storing program instruction；The processor calls described program instruction, for executing：Obtain image to be classified；It adjusts Classified to the image to be classified with disaggregated model, determines the initial category belonging to the image to be classified；It calls special Sign extraction model determines the target feature vector of the image to be classified；By the target feature vector of the image to be classified with Candidate classification feature vector for initial category configuration is compared, and comparing result is obtained；It is determined according to comparing result The classification results of the image to be classified.

Correspondingly, the embodiment of the present invention additionally provides a kind of user terminal, including：Processor and storage device；It is described to deposit Storage device, for storing program instruction；The processor calls described program instruction, for executing：Detect for identification by When the trigger action of button, camera assembly is called to obtain present preview image；If detecting that the present preview image is in steady Determine state, then the present preview image is determined as image to be classified；Classification request, the classification request are sent to server In carry the image to be classified；Receive the response message that the server returns, the response message includes described waits for point The classification results of class image, the classification results, which are the servers, to be waited for point according to disaggregated model and Feature Selection Model to described Class image obtained after classification determines.

Correspondingly, the embodiment of the present invention additionally provides a kind of computer storage media, is stored in the computer storage media There is program instruction, which is performed, for realizing above-mentioned each method.

The embodiment of the present invention can obtain image to be classified, and calling classification model carries out just classification to image to be classified, The target feature vector that Feature Selection Model determines image to be classified is recalled, only by the target feature vector of image to be classified It is only compared with for the candidate classification feature vector that configures in just classifying, and image to be classified is determined according to comparing result Classification results can effectively improve the accuracy of image classification, and need not be with excessively multi-class candidate classification feature vector It is compared, improves operation efficiency, save software and hardware resources.

Description of the drawings

In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with Obtain other attached drawings according to these attached drawings.

Fig. 1 is a kind of block schematic illustration for carrying out image classification of the embodiment of the present invention；

Fig. 2 a are a kind of structural schematic diagrams of preliminary classification model of the embodiment of the present invention；

Fig. 2 b are a kind of structural schematic diagrams of initial characteristics disaggregated model of the embodiment of the present invention；

Fig. 2 c are showing for the correspondence between the feature extraction network of the embodiment of the present invention and character representation optimization module It is intended to；

Fig. 2 d are the schematic diagrames of the training whitening parameters of the embodiment of the present invention；

Fig. 3 is a kind of flow diagram of image classification method of the embodiment of the present invention；

Fig. 4 is the flow diagram of another image classification method of the embodiment of the present invention；

Fig. 5 is a kind of schematic diagram of user interface of the embodiment of the present invention；

Fig. 6 is a kind of structural schematic diagram of image classification device of the embodiment of the present invention；

Fig. 7 is a kind of structural schematic diagram of server of the embodiment of the present invention；

Fig. 8 is the structural schematic diagram of another image classification device of the embodiment of the present invention；

Fig. 9 is a kind of structural schematic diagram of user terminal of the embodiment of the present invention.

Specific implementation mode

In embodiments of the present invention, on the one hand, by completing the disaggregated model of classification based training optimization to figure to be sorted Picture is classified, and determines initial category, which can be trained optimization by dedicated classification based training mode, and And can multi-class classification preferably be carried out to image in the equipment with classification feature such as be pre-configured in server.One side More fine accurate target is extracted in face by completing the Feature Selection Model of feature extraction optimization from image to be classified Feature vector, the candidate classification feature based on one or more candidate image under the target feature vector and initial category to Amount is compared, and is finally determined the classification of image to be classified, is obtained classification results.There is disaggregated model initially to be divided in this way Class, then be finely divided in preliminary classification by finer feature vector so that confirm that the classification come is more acurrate, and special The data processing amount that sign compares also efficiently reduces, and saves software and hardware resources, improves treatment effeciency.

In one embodiment, for above-mentioned disaggregated model, universal classification model may be used, pass through universal classification mould Type can carry out more broad classification to image, such as can the content of some image be divided into the classes such as cat or dog Not.The disaggregated model can also use fine granularity image classification model based on strong supervision message or based on Weakly supervised information Fine granularity image classification model.Based on fine granularity image classification model, the classification that can be more refined, such as can segment To kind, the kind of dog etc. of cat.

Fine granularity image classification model based on strong supervision message refers to：It is more preferable in order to obtain when model training optimizes Nicety of grading also use the additional artificial mark such as object callout box and position mark point other than the class label of image Information is noted, higher nicety of grading may be implemented in such disaggregated model.

Fine granularity image classification model based on Weakly supervised information refers to：In the case where not marking point by position, It can accomplish that preferable local message captures.There is gap compared to strong supervised classification precision, but cost of implementation is low, more suitable for Engineer application.

The embodiment of the present invention includes mainly the configuration phase to disaggregated model and Feature Selection Model, and based on classification mould The image classification stage of type and Feature Selection Model.In configuration phase, initial point can be established come offline based on configuration server Class model and initial characteristics extraction model, and by a large amount of training image to preliminary classification model and initial characteristics extraction model It is trained optimization, obtaining preferably can be to disaggregated model and Feature Selection Model that image is classified.It will be trained Disaggregated model and Feature Selection Model after optimization are loaded into (or some other for carrying out for the server of image classification The equipment of image classification) on, in order to provide the classified service to unknown images to the user online.In the image classification stage, use Image to be classified need to be only shot by user terminal for family or (such as the modes such as download) obtain figure to be sorted by other means Picture, and the image to be classified is carried to the server being sent in classification is asked for carrying out image classification, by the server Calling classification model and Feature Selection Model handle the image to be classified, determine the classification knot of image to be classified Fruit, and the response message asked the classification is returned to user terminal, it is carried in the response message and image to be classified is divided Class result.

In one embodiment, Fig. 1 shows a kind of frame signal for carrying out image classification of the embodiment of the present invention Figure, the frame include user side and service side, wherein the user terminal of user side can be the bands such as smart mobile phone, tablet computer The intelligent terminal of camera function and network function, naturally it is also possible to for terminals such as PCs.The online service for servicing side can be with It is asked by a classification that can receive intended application, is exclusively used in carrying out server or the server group of image classification to carry For the processed offline service for servicing side can be by certain services for being exclusively used in design and train classification models and Feature Selection Model Device or server group provide.In other embodiments, the online service and offline service for servicing side can also be by one Dedicated server or server group provide.In one embodiment, which can be with scanning recognition mould The application of block.

As shown in Figure 2 a, a kind of preliminary classification model is shown.The preliminary classification model includes：Input layers, Stem layers, Inception-resnet layers, Reduction layers, pooling layers, Dropout layers, Softmax layers of Average.Wherein, Input layers are input layer, can be used for input picture；Inception-resnet layers are hidden layer, which gathers around there are three hidden Containing layer, respectively：The first hidden layers of Inception-resnet-A, the second hidden layers of Inception-resnet-B and Inception-resnet-C third hidden layers；Stem layers are pretreatment layer, be can be used for inputting Inception-resnet- The data of A are pre-processed, which may include that data carry out multiple convolution and pond；Pooling layers of Average is Average pond layer can be used for carrying out dimension-reduction treatment to the data that Inception-resnet-C is exported；It Dropout layers, can be with For preventing preliminary classification model from the case where over-fitting occur, preliminary classification model is effectively avoided to can be good at training image Classify, but need the image classified to actual after deployment, then the poor situation of classifying quality；Softmax layers are point Class computation layer, its output result can be the probability that the image inputted by Input layers belongs to each classification.

In one embodiment, when configuring above-mentioned disaggregated model, model Core Generator can be utilized in advance by multiple figures As the training image of classification is input to preliminary classification model, preliminary classification model is trained.In one embodiment, when making When model Core Generator is a kind of Tensorflow (deep learning frame for fields such as image recognitions) tool, due to Reversed gradient on Tensorflow tools calculates automatically, therefore can be in training optimization process, rapidly according to defeated The training image entered adjusts and generates the corresponding parameter of each node in the preliminary classification model, to realize to initially dividing The training optimization of class model, obtains disaggregated model.Wherein, the classification of the training image can be carried out according to different design objects Adjustment.

In one embodiment, design object is to establish to can be used for first category (such as cat class) and second category The disaggregated model that (such as dog class) image distinguishes.It, can be with when being trained to preliminary classification model for such case Choosing M in advance, (M is positive integer, such as 10000) a dog class for having been identified as the other image of dog class as preliminary classification model Training image and P (P is positive integer, such as 10000) has been identified as cat of the other image of cat class as preliminary classification model Class training image.In one embodiment, after a certain dog class training image is input to preliminary classification model, preliminary classification mould Type can extract the image feature data of the dog class training image, and according to these image feature datas to dog class training image into Row classification shows sorter network model pair if the classification results of output indicate that the classification of the dog class training image is also dog The classification of the dog class training image is successful.Further, the other training image of dog class is had been labeled as M to carry out After classification, if success rate, which is more than, is preset to power threshold (such as 90%), it is determined that the preliminary classification model can be right well Otherwise the other image of dog class, which carries out Classification and Identification, can then adjust the corresponding parameter of each node in the preliminary classification model, And the disaggregated model by adjusting after again classifies to M dog class training image.Similarly, same mode profit may be used Preliminary classification model is trained and is optimized with P cat class training image, if finally to dog class training image and cat class The classification success rate of training image is satisfied by preset success rate threshold value, then is completed to the training of preliminary classification model, and will instruction Practice the preliminary classification model completed as the disaggregated model in the embodiment of the present invention.In other embodiments, can also be arranged more More is different classes of, is made to be trained optimization to preliminary classification model by obtaining a large amount of different classes of training image It obtains the success rate that finally obtained disaggregated model can classify to the image of each type and is above a certain success rate threshold value. The classifications such as above-mentioned first category and second category can be the classifications such as cat or dog, can also be specific cat and/or The classification that the kind etc. of dog more refines, such as can be shepherd dog, Shiba Inu etc. subdivision classification.

In one embodiment, in the training optimization process of preliminary classification model, successfully instruction will can each be trained Practice image and be determined as candidate figure, and is the class label that the candidate schemes one its generic of setting according to the classification described in it, By in candidate figure and corresponding class label associated storage to database, waited in order to subsequently be determined based on such distinguishing label Classification belonging to choosing figure.

In embodiments of the present invention, features described above extraction model can be optimized according to feature extraction network and character representation Module composition, this feature extracts the initial characteristics vector that network is used to extract image to be classified based on neural network, this feature Optimization module is indicated for optimizing to obtain the target feature vector of N-dimensional to initial characteristics vector, N is positive integer, this feature The target feature vector about image of extraction model final output N-dimensional, such as 2048 dimension target feature vector of output.

In one embodiment, in order to generate features described above extraction model, initial characteristics disaggregated model and instruction can be obtained Practice image, which includes that the initial characteristics constituted based on neural network extract network.Further, may be used To be trained optimization to initial characteristics disaggregated model according to the training image got, tagsort model, this feature are obtained Disaggregated model includes being trained obtained feature extraction network after optimization to initial characteristics extraction network, and then can be from this Feature extraction network is obtained in tagsort model, generates character representation optimization module based on whitening parameters, and according to this feature It extracts network and this feature indicates that optimization module generates Feature Selection Model.In one embodiment, this feature indicates optimization mould Block can be realized based on R-MAC (a kind of image characteristic extracting method).

In one embodiment, this feature indicates that optimization module can be trained based on training image whitening parameter It is obtained after optimization, which can be made of matrix w and bias matrix b, and this feature indicates that optimization module can be used for The feature vector of feature extraction network output is optimized, and exports the feature vector after optimization.This feature indicates optimization mould The processing procedure of block can be the process of a whitening processing.In one embodiment, the training optimization process packet of whitening parameters It includes：The feature extraction network can be called to extract feature vector respectively from multiple training images first, and according to getting Feature vector is calculated to optimize to obtain matrix w and bias matrix b, and then obtains character representation optimization module, this feature table Show that optimization module is based on matrix w and bias matrix b is generated.For example, there is 1,000,000 training sample, that is, there are 1,000,000 training and scheme Picture can obtain the character representation training matrix x of 2048 dimensions for each training image, need to calculate a 2048*2048 Matrix w and 2048 bias matrix b come albefaction this 2048 dimension character representation training matrix x so that 2048 dimension features Indicate in matrixes of the training matrix x after albefaction, it is uncorrelated between dimension data, that is, know some dimension data value it The value for guessing or calculating other dimension is not helped afterwards.Specifically, as shown in Figure 2 d, the image in left side is in Fig. 2 d The bivector of 1000000 * 2048 dimensions, it can be seen that in the image in left side, than comparatively dense, the value of dimension data between dimension data Between there are certain correlations, by way of mathematical computations, constantly calculate matrix w and bias matrix b, it is final so that a left side The bivector of side passes through x_wAfter the calculation formula of=x*w+b calculates, the image on right side is obtained.As can be seen that the image on right side In, the correlation between each dimension data is low.It can directly calculate w and b in theory, but often due to sample Quantity is too big, but removes one w and b of approach by trained mode so that the bivector of input is after w and b, Neng Goucong The image in left side is converted to the image on right side.Wherein, 2048 dimensions refer to that character representation optimization module needs the dimension amount exported, such as There are the demands of other dimension amounts for fruit, such as need the data of the 4096 even more dimension amounts of dimension of output, then raw in training w and b During at character representation optimization module, above-mentioned 2048 need to adjust to the numerical value of the 4096 even more dimension amounts of dimension.

In one embodiment, this feature indicates that optimization module is generated according to a conversion formula, the conversion formula For：x_w=x*w+b.Namely after initial matrix x input feature vector indicates optimization module, this feature indicates that optimization module can be with According to the matrix x after transformational relation output optimization_w=x*w+b.Likewise, when feature vector is input to by feature extraction network After character representation optimization module, this feature indicate optimization module can also adopt in a similar manner, output optimization after feature to Amount.Using such optimization processing mode, it can make in the feature vector after optimization processing, between the feature represented by characteristic value Correlation is low, for example, need be less than certain relevance threshold, each feature have it is same or similar as variance.

In one embodiment, features described above disaggregated model and character representation optimization module may each be and generated using model Tool is trained initial characteristics disaggregated model and whitening parameters and generates to obtain.Specifically, used model generates work Tool or Tensorflow tools can be more fast since the reversed gradient on Tensorflow calculates automatically Tagsort model and character representation optimization module are obtained promptly.

As shown in Figure 2 b, a kind of initial characteristics disaggregated model is shown, which includes based on nerve The initial characteristics that network is constituted extract network, which may be used convolutional neural networks.It can be seen from the figure that this is first Beginning feature extraction network includes three convolutional layers：First convolutional layer cfgl [0] block, second convolutional layer cfgl [1] block, Three convolutional layer cfgl [2] block can also include in other embodiments more convolutional layers, such as in cfgl [2] block Later, it can also include Volume Four lamination cfgl [3] block etc..During training optimizes, initial characteristics extract in network Each convolutional layer process of convolution is carried out to the training image of input, and export the feature vector of the training image about input (i.e. the preliminary classification feature vector of training image).Wherein, initial characteristics extraction network includes the convolutional layer and the of the first kind The convolutional layer of two types, in Fig. 2 b, cfgl [0] block, cfgl [1] block is the convolutional layer of the first kind, cfgl [0] Data after process of convolution are sent to cfgl [1] block by block, and cfgl [1] block is then by the data after process of convolution It is sent to cfgl [2] block.Cfgl [2] block shown in Fig. 2 b is the convolutional layer of Second Type, and cfgl [2] block will be rolled up Product treated data (i.e. the preliminary classification feature vector of training image) export, in order to which subsequent network layer is to preliminary classification Feature vector is calculated, and determines the classification of image.Tagsort mould is being obtained to initial characteristics disaggregated model training end After type, the network layer for calculating characteristic of division vector is extracted from the tagsort model that training obtains, by these networks Layer is used as feature extraction network, such as by cfgl [0] block, cfgl [1] block and cfgl [2] block, cfgl [3] Block is as feature extraction network.

In embodiments of the present invention, optimization is trained to the initial characteristics disaggregated model shown in Fig. 2 b, obtains feature point After class model, network can be extracted from the initial characteristics after obtaining training optimization in tagsort model (i.e. from tagsort mould Feature extraction network is obtained in type), and establish connection relation such as Fig. 2 c institutes of feature extraction network and character representation optimization module Show.It can be seen from the figure that after image input feature vector extracts network, each convolutional layer (cfgl in feature extraction network [0] block, cfgl [1] block and cfgl [2] block) process of convolution can be carried out to the image of input, and then determine Go out the preliminary classification feature vector of input picture, and by the last one convolutional layer cfgl [2] block by the initial of input picture Characteristic of division vector is sent to character representation optimization module, and then using character representation optimization module to initial point of input picture Category feature vector optimizes processing, determines and export the candidate classification feature vector of input picture.Wherein, which can Think that the vectorial that various dimensions are carried out to the preliminary classification feature vector of input picture, such as the vector of 2048 dimensions indicate, Feature vector (i.e. candidate classification feature vector) after optimizing can be the vector of one 2048 dimension, the characteristic of 2048 dimension According to the form of expression for example can be (0.1,0.11,0.15 ..., 0.16).It is understood that based on different mark sheets Show optimization module, the vector of other dimensions can also be obtained, uses the dimension of vector higher, to the classification results of input picture Also more accurate.

In embodiments of the present invention, preliminary classification can be carried out to various training images with calling classification model, and will classification Successful training image stores in the database as candidate figure.It is possible to further regard candidate's figure as feature extraction network Input, call Feature Selection Model that the candidate classification feature vector of all candidate figures is calculated, generated for all candidate figures Corresponding candidate classification feature vector, and candidate's figure and candidate classification feature vector are associated storage, in order to rear Continue in the classification request for receiving user about image, can be directly based upon needed for these candidate classification feature vectors determine The classification results of image of classifying simultaneously return to user terminal, without after receiving about the classification request of image, then adjusting The candidate figure of magnanimity is carried out with Feature Selection Model candidate classification feature vector is calculated, saves a large amount of calculating and inquiry Time.

In one embodiment, when candidate's figure and candidate classification feature vector are associated storage, time can be established The mapping relations list of choosing figure and candidate classification feature vector, row 1 are the storage address of candidate figure, and row 2 are candidate classification feature Vector can quickly find the candidate classification when finding candidate classification feature vector according to the mapping relations list The corresponding candidate figure of feature vector, vice versa.

Disaggregated model and Feature Selection Model are trained and relevant parameter by the completion of a large amount of training image Optimization after, you can will complete training optimization after disaggregated model and Feature Selection Model be configured in corresponding server, And scheme a large amount of candidate with candidate classification feature vector associated storage in corresponding server, provide online figure to the user As classified service.

In one embodiment, a large amount of candidate is schemed with candidate classification feature vector associated storage in corresponding server When, different candidate classification feature vectors can also be configured to different classifications according to the classification belonging to candidate figure.Specifically, Can directly be that the category configures L candidate classification feature when being stored with a candidate figures of L (L is positive integer) under a certain classification Vector that is to say each corresponding candidate classification feature vector of candidate figure.Alternatively, can also be to each candidate figure under the category Candidate classification feature vector carry out similarity calculation, and each candidate classification that similarity is less than to the first similarity threshold is special Sign vector is classified as of a sort feature vector, enables each candidate figure belonged to corresponding to of a sort candidate classification feature vector equal It is characterized using the same candidate classification feature vector namely multiple candidate figures corresponds to the same candidate classification feature vector, And then it is less than a candidate classification feature vectors of L for category configuration.In this manner, it can ensure that image classification is accurate While exactness, it is reduced to the quantity of each classification configuration candidate classification feature vector, and then improves arithmetic speed.In a reality It applies in example, candidate figure can be the various training images of aforementioned use, can also be user by the modes such as downloading or shooting The image of the respective type got, these image known class, and therefrom carried by Feature Selection Model mentioned above Candidate classification feature vector is taken, these images and corresponding candidate classification feature vector can be stored in database, convenient It is follow-up to search.

It is understood that describing in embodiments of the present invention can train by server, configures disaggregated model And Feature Selection Model in other embodiments can also be by the abundant PC of powerful, software and hardware resources come real Existing, the present invention is not especially limited this.

Fig. 3 is referred to again, is a kind of flow diagram of image classification of the embodiment of the present invention, the institute of the embodiment of the present invention The method of stating can be executed by a server or server group.Described method includes following steps for the embodiment of the present invention.

S301：Obtain image to be classified.In the specific implementation, server can receive the classification request of user terminal transmission, And the image to be classified of carrying is obtained from classification request.The image to be classified can be the shooting that user utilizes user terminal What module acquired, or acquired using other modes, the present invention is not especially limited this.

S302：Calling classification model classifies to image to be classified, determines the initial category belonging to image to be classified. In one embodiment, can be development of user be trained optimization to the disaggregated model based on preliminary classification model obtains, should Disaggregated model has the characteristics that TOP5 accuracys rate are high namely the disaggregated model can more accurately determine image to be classified institute The classification (i.e. initial category) of category, for example, when the corresponding object of image to be classified is a Persian cat, then disaggregated model then may be used More accurately to determine the initial category belonging to image to be classified as cat class rather than dog class.

In one embodiment, which is obtained by preliminary classification model by a large amount of training optimization.Initial point Class model can be as shown in Figure 2 a, is trained optimization to the preliminary classification model, obtains the process of disaggregated model, may refer to Above-mentioned associated description, details are not described herein again.

In one embodiment, after image to be classified input disaggregated model, disaggregated model can extract image to be classified Characteristic, and the probability that image to be classified belongs to each candidate categories is calculated according to these characteristics, and then can incite somebody to action Probability K in the top (K is positive integer) a candidate categories, are determined as the initial category belonging to image to be classified.

S303：Feature Selection Model is called to determine the target feature vector of image to be classified.Wherein, this feature extracts mould Type can be constituted according to feature extraction network and character representation optimization module, and this feature extracts network and is used to be based on nerve net Network extracts the initial characteristics vector of image to be classified, and this feature representation module to the initial characteristics vector for optimizing to obtain The target feature vector of N-dimensional, N are positive integer.

In one embodiment, in order to generate features described above extraction model, initial characteristics disaggregated model and instruction can be obtained Practice image, which includes that the initial characteristics constituted based on neural network extract network.Further, may be used To be trained optimization to initial characteristics disaggregated model according to the training image got, tagsort model, this feature are obtained Disaggregated model includes being trained obtained feature extraction network after optimization to initial characteristics extraction network, and then can be from this Feature extraction network is obtained in tagsort model, and according to feature extraction network and the character representation optimization module being generated in advance Generate Feature Selection Model.Wherein, this feature indicates that optimization module can be trained based on training image whitening parameter What optimization obtained, this feature indicates that optimization module can be used for optimizing the feature vector that feature extraction network exports, and Feature vector after output optimization.

In one embodiment, features described above extraction network can be such as Fig. 2 c with the connection relation of character representation optimization module It is shown.Wherein, this feature extraction network can be that development of user is constituted based on neural network, and this feature extraction network may include The convolutional layer of the convolutional layer of the first kind and the convolutional layer of Second Type, the first kind exports the data of process of convolution to feature Another convolutional layer in network is extracted, the convolutional layer of Second Type exports the data after process of convolution to be optimized to character representation Module.After image to be classified input feature vector extracts network, feature extraction network handles classification image can be called to carry out first Processing can obtain an initial characteristics vector about image to be classified, and will be initial special by the convolutional layer of Second Type The vectorial input feature vector of sign indicates optimization module, and then character representation optimization module is called to optimize initial characteristics vector, obtains To the target feature vector of a N-dimensional about image to be classified, such as the vector of one 2048 dimension, the target feature vector energy Enough to indicate the image to be classified to a certain extent, which is the image of computer understanding, and user is not readily understood, It is not easy to visualize.

S304：The target feature vector of image to be classified is carried out with the candidate classification feature vector for initial category configuration Comparison, obtains comparing result.The mode compared can be to calculate in target feature vector and candidate classification feature vector, per one-dimensional Similarity between vector, such as calculate similarity Hamming distance or Euclidean distance between every one-dimensional vector.It is final to determine Similarity it is higher, then image to be classified belong to the classification belonging to corresponding candidate classification feature vector probability it is bigger.

S305：The classification results of image to be classified are determined according to comparing result.

In one embodiment, it under the initial category may include at least one candidate figure, server or server group The candidate figure for belonging to initial category can be obtained, and is schemed according to candidate, it is that initial category configuration is candidate to call Feature Selection Model Characteristic of division vector.

In one embodiment, the candidate classification feature vector for being initial category configuration can pass through Feature Selection Model In feature extraction network candidate figure is handled, obtain the preliminary classification feature vector under initial category, and pass through feature What the character representation optimization module in extraction model optimized the preliminary classification feature vector under initial category.The time It can be the feature vector of a N-dimensional to select characteristic of division vector, such as can be the vector of 2048 dimensions.

May include Q candidate figure (each candidate corresponding candidate classification feature vector of figure) under initial category.For Such case can directly be initial classes when it is that initial category configures candidate classification feature vector to call Feature Selection Model Not Pei Zhi Q candidate classification feature vector, that is to say each corresponding candidate classification feature vector of candidate figure.Alternatively, also may be used Similarity calculation is carried out with the candidate classification feature vector to each candidate's figure under initial category, and similarity is less than the first phase It is classified as same class like the candidate classification feature vector of degree threshold value, belongs to each corresponding to of a sort candidate classification feature vector Candidate's figure is characterized using the same candidate classification feature vector namely multiple candidate figures correspond to the same candidate classification spy Sign vector, and then be candidate classification feature vector of the initial category configuration less than Q.In this manner, can ensure While image classification accuracy, it is reduced to the quantity of initial category configuration candidate classification feature vector, and then reduces calculation amount, with This improves arithmetic speed.

In one embodiment, which may include first category and second category, and server can be treated point The target feature vector of class image carries out similarity calculation with the candidate classification feature vector for first category configuration, obtains first Similarity；Similarity calculation is carried out with the candidate classification feature vector for second category configuration to target feature vector, obtains the Two similarities, and the first similarity and the second similarity are compared, comparing result is obtained, which indicates first Higher value between similarity and the second similarity.Further, if comparing result indicates that the first similarity is more than the second phase Like degree, then it can determine that image to be classified belongs to first category.Wherein, which can also include third classification, the 4th Classification or other classifications, the present invention are not especially limited this.

For example, the target feature vector of image to be classified be (0.23,0.44 ..., 0.61), initial category is dog class, In, first category that dog class includes is golden hair, and second category is safe enlightening, and the corresponding candidate classification feature vector of golden hair is (0.23,0.44 ..., 0.67), the corresponding candidate classification feature vector of safe enlightening be (0.23,0.31 ..., 0.60), server By to target feature vector (0.23,0.41 ..., 0.61) corresponding with golden hair candidate classification feature vector (0.23, 0.44 ..., 0.67) progress similarity calculation, it is 98% to obtain the first similarity, further, to target feature vector (0.23,0.41 ..., 0.61) corresponding with safe enlightening candidate classification feature vector (0.23,0.31 ..., 0.60) carry out it is similar Degree calculates, and it is 90% to obtain the second similarity, by comparing the size of the first similarity and the second similarity, determines comparing result It is more than the second similarity for the first similarity, then can determines that image to be classified belongs to golden hair.

In one embodiment, it may each comprise and candidate classification feature vector under above-mentioned first category and under second category Associated candidate figure can be according to first if comparing result indicates that above-mentioned first similarity is more than above-mentioned second similarity Similarity determines to obtain the associated images of the image to be classified, which refers to：Candidate classification feature vector and target Similarity between feature vector is the candidate figure of the first similarity.For example, feature vector to be sorted and target feature vector it Between similarity be the candidate figure of the first similarity be image JPG1 and image JPG2, then, then can be by image JPG1 and figure As JPG2 is determined as the associated images of image to be classified.

In one embodiment, the target feature vector of image to be classified with configure for first category or second category When candidate classification feature vector carries out similarity calculation, can be calculated by matching algorithm the target signature of the image to be classified to Euclidean distance or cosine angle under amount and first category or under second category between candidate classification feature vector, and then obtain Similarity degree between target feature vector candidate classification feature vector corresponding with each candidate's figure.Passing through matching algorithm meter It when calculating similarity, can also be matched in such a way that order traversal compares successively, naturally it is also possible to using based on faiss The library of effective similarity searching and dense Vector Clustering (one to increase income be used for) search engine searches for candidate classification feature Vector is simultaneously matched.

In one embodiment, it after server determines the classification results of image to be classified, can be returned to user terminal The response message that back stitching asks above-mentioned classification, the response message may include in above-mentioned classification results and above-mentioned associated images It is at least one.

Fig. 4 is referred to again, is the flow diagram of another image classification method of the embodiment of the present invention, and the present invention is implemented The method of example can be executed by user terminal.Described method includes following steps for the embodiment of the present invention.

S401：When detecting the trigger action for recognition button, camera assembly is called to obtain present preview image.

S402：If detecting that present preview image is in stable state, present preview image is determined as to be sorted Image.

In one embodiment, user terminal is equipped with there are the intended application of image classification functional entrance, the image point Class functional entrance may include scanning recognition module.When user wants to identify some object, the scanning recognition mould can be directed to The recognition button that block is provided inputs trigger action, when user terminal detects the trigger action, you can generates scanning recognition behaviour Make event, camera assembly can be called to obtain present preview image based on the event.Further, if detecting current preview Image is in stable state, then can lead the present preview image in stable state as image to be classified, or directly Enter stored image as image to be classified.Wherein, which for example can be browser, instant messaging application, branch Pay the application with image classification functional entrance such as application.

In one embodiment, present preview image can be obtained according to prefixed time interval, is determining current preview figure Seem it is no in stable state when, can by present preview image with first T time acquisition preview image progress similarity comparison, such as Fruit similarity is more than or equal to similarity threshold, then can determine that present preview image is in stable state.Wherein T be more than The occurrence of 0 positive integer, T can be adjusted accordingly according to different designs demand, and the present invention is not especially limited this.

In one embodiment, as shown in figure 5, intended application is browser, the scanning recognition on clicking browser After button 501, user terminal enters image and obtains interface, and it includes image display area 502 and prompt to be obtained on interface in image The "+" button 503 of image is imported, user, which clicks button 503, can enter image selection and confirm interface.When detecting that image is aobvious When showing that the preview image shown by region 502 is in stable state, then the preview image in stable state is determined as waiting for point Class image.

S403：Classification request is sent to server, image to be classified is carried in classification request.Obtained it is to be sorted After image, user terminal can generate the classification request for carrying the image to be classified, be then sent to classification request and provide The server of line classified service.

S404：Receive the response message that server returns.The response message includes the classification results of image to be classified, this point Class result can be that server to image to be classified obtain after classification determines according to disaggregated model and Feature Selection Model. In one embodiment, after server receives classification request, classification request can be responded, determines image to be classified Classification results, and classification results carrying is back to user terminal in response message.Further, user terminal receives Server return the response message for carrying classification results when, which can be shown on a user interface, so as to It is checked in user.Wherein, which can also include the association of image to be classified other than it can carry classification results Image or other information, the present invention are not especially limited this.

In one embodiment, user terminal is receiving after the response message that server returns, and can show user Interface, the user interface are shown：In the classification results of image to be classified, associated images and description information any one or It is multiple, what the classification results and associated images can be in response to carry in information, which can be according to classification results What inquiry obtained.Wherein, which can be text message associated with classification results and image information.For example, point Class result shows that the object that image to be classified includes is X systems offroad vehicle 01, then the description information may include then that X systems are cross-country The basic introduction of vehicle 01, such as capabilities list, Time To Market, price, the shops 4S etc. for nearby selling the X systems offroad vehicle 01, also May include the platform network address etc. for selling the X systems offroad vehicle 01, the present invention does not introduce this specifically.

In one embodiment, user terminal can be directly based upon in response message and carry after receiving response message Classification results on-line search, and then obtain the description information of image to be classified, and include in user interface by the description information On.Alternatively, can be when receiving the trigger action for user interface, then on-line search or local obtain the classification results Description information, and the description information got is shown on a user interface.

It should be noted that the description information that above-mentioned user interface is shown is in addition to can inquire to obtain by classification results In addition, can also be that response message carries description information, what user terminal was acquired from the response message, the present invention couple This is not especially limited.

The embodiment of the present invention additionally provides a kind of computer storage media, and have program stored therein finger in the computer storage media It enables, which is performed, for realizing the correlation method described in above-described embodiment.

Fig. 6 is referred to again, is a kind of structural schematic diagram of image classification device of the embodiment of the present invention, image classification dress Setting can be arranged in the server, or some software and hardware resources can also be arranged compared in the intelligent terminal of horn of plenty, such as one In a little PCs.

In one realization method of the described device of the embodiment of the present invention, described device includes such as lower structure.

Acquisition module 601, for obtaining image to be classified；

Calling module 602 is classified to the image to be classified for calling classification model, is determined described to be sorted Initial category belonging to image；

The calling module 602 is additionally operable to the target signature for calling Feature Selection Model to determine the image to be classified Vector；

Contrast module 603, for what is configured by the target feature vector of the image to be classified and for the initial category Candidate classification feature vector is compared, and comparing result is obtained；

Determining module 604, the classification results for determining the image to be classified according to comparing result.

In one embodiment, the Feature Selection Model is according to feature extraction network and character representation optimization module structure At, the feature extraction network is used to extract the initial characteristics vector of image to be classified, the mark sheet based on neural network Show optimization module for optimizing to obtain the target feature vector of N-dimensional to the initial characteristics vector, N is positive integer.

In one embodiment, described device can also include：Training module 605, generation module 606, wherein：Obtain mould Block 601 is additionally operable to obtain initial characteristics disaggregated model and training image, and the initial characteristics disaggregated model includes based on nerve The initial characteristics that network is constituted extract network；Training module 605, the training for being got according to acquisition module 601 are schemed As being trained optimization to the initial characteristics disaggregated model, tagsort model is obtained, wherein in the tagsort model After being trained optimization to initial characteristics extraction network, obtained feature extraction network；Acquisition module 601, is also used In obtaining the feature extraction network from the tagsort model；Generation module 606, it is special for being generated based on whitening parameters Sign indicates optimization module；Generation module 606 is additionally operable to according to the feature extraction network and the character representation optimization module, Generate Feature Selection Model, wherein the feature that the character representation optimization module is used to export the feature extraction network to Amount optimizes, and for exporting the feature vector after optimizing.

In one embodiment, described device can also include：Configuration module 607, wherein：Acquisition module 601, is additionally operable to Obtain the candidate figure for belonging to the initial category；Configuration module 607, the candidate for being got according to acquisition module 601 Figure, it is that the initial category configures candidate classification feature vector to call the Feature Selection Model.

In one embodiment, configuration module 607 can be specifically used for carrying by the feature in the Feature Selection Model It takes network to handle the candidate figure, obtains the preliminary classification feature vector under the initial category, and pass through the spy Character representation optimization module in sign extraction model optimizes the preliminary classification feature vector under the initial category, obtains The candidate classification feature vector of N-dimensional under the initial category.

In one embodiment, the initial category includes first category and second category, and the contrast module 603 can To include：Computing unit 6031 is used to configure to the target feature vector of the image to be classified and for the first category Candidate classification feature vector carries out similarity calculation, obtains the first similarity；To target feature vector and it is the second category The candidate classification feature vector of configuration carries out similarity calculation, obtains the second similarity；Comparison unit 6032, for by described the One similarity and second similarity are compared, and obtain comparing result, it is similar that the comparing result indicates described first Higher value between degree and second similarity.

In one embodiment, if determining module 604 can be specifically used for the comparing result and indicate first phase It is more than second similarity like degree, it is determined that the image to be classified belongs to the first category.

In one embodiment, include and candidate classification feature vector under the first category and under the second category Associated candidate figure, if determining module 604 can be also used for the comparing result and indicate first similarity more than described Second similarity then determines to obtain the associated images of the image to be classified, the associated images according to first similarity Refer to：Similarity between the candidate classification feature vector and the target feature vector is the candidate figure of the first similarity.

In embodiments of the present invention, the specific implementation of above-mentioned modules can refer to the embodiment corresponding to aforementioned figures 3 The description of middle related content.

Refer to Fig. 7 again, be a kind of structural schematic diagram of server of the embodiment of the present invention, the embodiment of the present invention it is described Server includes the structures such as power supply module, and includes processor 701, storage device 702 and network interface 703.The processing Corresponding image point can be realized by processor 701 with interaction data between device 701, storage device 702 and network interface 703 Class function.

The storage device 702 may include volatile memory (volatile memory), such as random access memory Device (random-access memory, RAM)；Storage device 702 can also include nonvolatile memory (non-volatile Memory), such as flash memory (flash memory), solid state disk (solid-state drive, SSD) etc.；It is described to deposit Storage device 702 can also include the combination of the memory of mentioned kind.

The network interface 703 can interaction data, user terminal can be between other servers, various user terminals The classification for carrying image to be classified request is sent to the network interface 703, is exported to server by the network interface 703 Processor 701 handled.

The processor 701 can be central processing unit (central processing unit, CPU).Implement at one In example, the processor 701 can also be graphics processor 701 (Graphics Processing Unit, GPU).The place Reason device 701 can also be the combination by CPU and GPU.In the server, can as needed include multiple CPU and GPU into The corresponding image procossing of row.In one embodiment, the storage device 702 is for storing program instruction.The processor 701 Described program can be called to instruct, realized such as the above-mentioned various methods being related in the embodiment of the present invention.

In first possible embodiment, the processor 701 of the server calls the storage device The program instruction stored in 702, for obtaining image to be classified；Calling classification model classifies to the image to be classified, Determine the initial category belonging to the image to be classified；Feature Selection Model is called to determine the target of the image to be classified Feature vector；By the target feature vector of the image to be classified and the candidate classification feature vector for initial category configuration It is compared, obtains comparing result；The classification results of the image to be classified are determined according to comparing result.

In one embodiment, the processor 701 is additionally operable to obtain initial characteristics disaggregated model, the initial characteristics Disaggregated model includes that the initial characteristics constituted based on neural network extract network；Training image is obtained, and according to the training Image is trained optimization to the initial characteristics disaggregated model, obtains tagsort model, wherein the tagsort model It include obtained feature extraction network after being trained optimization to initial characteristics extraction network；From the tagsort The feature extraction network is obtained in model；Character representation optimization module is generated based on whitening parameters；According to the feature extraction Network and the character representation optimization module generate Feature Selection Model, wherein the character representation optimization module is used for institute The feature vector for stating the output of feature extraction network optimizes, and for exporting the feature vector after optimizing.

In one embodiment, the processor 701 is additionally operable to obtain the candidate figure for belonging to the initial category；According to The candidate figure, it is that the initial category configures candidate classification feature vector to call the Feature Selection Model.

In one embodiment, the processor 701 is additionally operable to through the feature extraction net in the Feature Selection Model Network handles the candidate figure, obtains the preliminary classification feature vector under the initial category, and carry by the feature Character representation optimization module in modulus type optimizes the preliminary classification feature vector under the initial category, obtains described The candidate classification feature vector of N-dimensional under initial category.

In one embodiment, the initial category includes first category and second category, and the processor 701 is also used In the target feature vector to the image to be classified phase is carried out with the candidate classification feature vector configured for the first category It is calculated like degree, obtains the first similarity；To target feature vector and the candidate classification feature vector for second category configuration Similarity calculation is carried out, the second similarity is obtained；First similarity and second similarity are compared, obtained pair Than as a result, the comparing result indicates the higher value between first similarity and second similarity.

In one embodiment, the processor 701, if being additionally operable to the comparing result indicates first similarity More than second similarity, it is determined that the image to be classified belongs to the first category.

In one embodiment, include and candidate classification feature vector under the first category and under the second category Associated candidate figure, the processor 701, if being additionally operable to the comparing result indicates that first similarity is more than described the Two similarities then determine to obtain the associated images of the image to be classified according to first similarity, and the associated images are Refer to：Similarity between the candidate classification feature vector and the target feature vector is the candidate figure of the first similarity.

In embodiments of the present invention, the specific implementation of the processor 701 can refer to the embodiment corresponding to aforementioned figures 3 The description of middle related content.

Fig. 8 is referred to again, is the structural schematic diagram of another image classification device of the embodiment of the present invention, the image classification Device can be arranged in the user terminal.

Detection module 801, for detecting the trigger action for recognition button；

Acquisition module 802 when detecting the trigger action for recognition button for the detection module, calls camera shooting group Part obtains present preview image；

The detection module 801 is additionally operable to detect whether the present preview image is in stable state；

Determining module 803, if detecting that the present preview image is in stable state for the detection module, The present preview image is determined as image to be classified；

Sending module 804 carries the figure to be sorted for sending classification request to server in the classification request Picture；

Receiving module 805, the response message returned for receiving the server, the response message include described wait for point The classification results of class image, the classification results, which are the servers, to be waited for point according to disaggregated model and Feature Selection Model to described Class image obtained after classification determines.

In one embodiment, which can also include：Display module 806, for receiving from server return Response message after, show user interface, the user interface shows：Classification results, the associated diagram of the image to be classified In picture and description information any one or it is multiple, the classification results and the associated images are in the response message It carries, the description information inquires to obtain according to the classification results.

In embodiments of the present invention, the specific implementation of above-mentioned modules can refer to the embodiment corresponding to aforementioned figures 4 The description of middle related content.

Fig. 9 is referred to again, is a kind of structural schematic diagram of user terminal of the embodiment of the present invention, the institute of the embodiment of the present invention It may include the structures such as power supply module to state user terminal, and includes processor 901, storage device 902 and transceiver 903.Institute Corresponding figure can be realized with interaction data between processor 901, storage device 902 and transceiver 903 by processor 901 by stating As classification feature.

The storage device 902 may include volatile memory (volatile memory), such as random access memory Device (random-access memory, RAM)；Storage device 902 can also include nonvolatile memory (non-volatile Memory), such as flash memory (flash memory), solid state disk (solid-state drive, SSD) etc.；It is described to deposit Storage device 902 can also include the combination of the memory of mentioned kind.

The transceiver 903 interaction data, server can will carry between server, various user terminals The response message of classification results is sent to the transceiver 903, is exported to the processor of user terminal by the transceiver 903 901 are handled.

The processor 901 can be central processing unit 901 (central processing unit, CPU).At one In embodiment, the processor 901 can also be graphics processor 901 (Graphics Processing Unit, GPU).Institute It can also be the combination by CPU and GPU to state processor 901.Can include multiple CPU as needed in the user terminal Corresponding image procossing is carried out with GPU.In one embodiment, the storage device 902 is for storing program instruction.The place Reason device 901 can call described program to instruct, and realize such as the above-mentioned various methods being related in the embodiment of the present invention.

In first possible embodiment, the processor 901 of the user terminal calls the storage device The program instruction stored in 902 when for detecting the trigger action for being directed to recognition button, calls camera assembly to obtain current pre- Look at image；If detecting that the present preview image is in stable state, the present preview image is determined as waiting for point Class image；Classification request is sent to server, the image to be classified is carried in the classification request；Receive the server The response message of return, the response message include the classification results of the image to be classified, and the classification results are the clothes Business device to the image to be classified obtain after classification determines according to disaggregated model and Feature Selection Model.

In one embodiment, the processor 901 is additionally operable to receiving after the response message that server returns, Show that user interface, the user interface are shown：In the classification results of the image to be classified, associated images and description information Any one or it is multiple, carried in the classification results and the described response message of the associated images, it is described to retouch Information is stated to inquire to obtain according to the classification results.

In embodiments of the present invention, the specific implementation of the processor 901 can refer to the embodiment corresponding to aforementioned figures 4 The description of middle related content.

One of ordinary skill in the art will appreciate that realizing all or part of flow in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the program can be stored in a computer read/write memory medium In, the program is when being executed, it may include such as the flow of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access Memory, RAM) etc..

Above disclosed is only the section Example of the present invention, cannot limit the right of the present invention with this certainly Range, those skilled in the art can understand all or part of the processes for realizing the above embodiment, and according to right of the present invention Equivalent variations made by it is required that, still belong to the scope covered by the invention.

Claims

1. a kind of image classification method, which is characterized in that the method includes：

Obtain image to be classified；

Calling classification model classifies to the image to be classified, determines the initial category belonging to the image to be classified；

The target feature vector of the image to be classified is carried out with the candidate classification feature vector for initial category configuration Comparison, obtains comparing result；

The classification results of the image to be classified are determined according to the comparing result.

2. according to the method described in claim 1, it is characterized in that, the Feature Selection Model be according to feature extraction network and What character representation optimization module was constituted, the feature extraction network is used to extract the initial spy of image to be classified based on neural network Sign vector, the character representation optimization module be used for the initial characteristics vector optimize to obtain the target signature of N-dimensional to Amount, N is positive integer.

3. method according to claim 1 or 2, which is characterized in that the method further includes：

Initial characteristics disaggregated model is obtained, the initial characteristics disaggregated model includes the initial characteristics constituted based on neural network Extract network；

Training image is obtained, and optimization is trained to the initial characteristics disaggregated model according to the training image, obtains spy Levy disaggregated model, wherein the tagsort model includes after being trained optimization to initial characteristics extraction network, obtaining The feature extraction network arrived；

The feature extraction network is obtained from the tagsort model；

Character representation optimization module is generated based on whitening parameters, the character representation optimization module is used for the feature extraction net The feature vector of network output optimizes, and for exporting the feature vector after optimizing；

According to the feature extraction network and the character representation optimization module, Feature Selection Model is generated.

4. according to the method described in claim 3, it is characterized in that, the method further includes：

Obtain the candidate figure for belonging to the initial category；

According to the candidate figure, it is that the initial category configures candidate classification feature vector to call the Feature Selection Model.

5. according to the method described in claim 4, it is characterized in that, it is described according to it is described it is candidate scheme, call the feature extraction Model is that initial category configuration candidate classification feature vector includes：

The candidate figure is handled by the feature extraction network in the Feature Selection Model, obtains the initial category Under preliminary classification feature vector；And

By the character representation optimization module in the Feature Selection Model to the preliminary classification feature under the initial category to Amount optimizes, and obtains the candidate classification feature vector of the N-dimensional under the initial category.

6. method according to claim 1 or 2, which is characterized in that the initial category includes first category and the second class Not, the target feature vector by the image to be classified with for the initial category configuration candidate classification feature vector into Row comparison, obtains comparing result, including：

The target feature vector and the candidate classification feature vector for first category configuration of the image to be classified are carried out Similarity calculation obtains the first similarity；

Similarity calculation is carried out with the candidate classification feature vector for second category configuration to target feature vector, obtains the Two similarities；

First similarity and second similarity are compared, comparing result is obtained, the comparing result indicates Higher value between first similarity and second similarity.

7. according to the method described in claim 6, it is characterized in that, it is described determined according to the comparing result it is described to be sorted The classification results of image, including：

If the comparing result indicates that first similarity is more than second similarity, it is determined that the image to be classified Belong to the first category.

8. according to the method described in claim 6, it is characterized in that, under the first category and including under the second category With the associated candidate figure of candidate classification feature vector, the method further includes：

If the comparing result indicates that first similarity is more than second similarity, according to first similarity Determination obtains the associated images of the image to be classified, and the associated images refer to：The candidate classification feature vector with it is described Similarity between target feature vector is the candidate figure of the first similarity.

9. a kind of image classification method, which is characterized in that the method includes：

If detecting that the present preview image is in stable state, the present preview image is determined as figure to be sorted Picture；

The response message that the server returns is received, the response message includes the classification results of the image to be classified, institute It is that the server carries out classification determination according to disaggregated model and Feature Selection Model to the image to be classified to state classification results It obtains afterwards.

10. according to the method described in claim 9, it is characterized in that, the method further includes：

It is receiving after the response message that server returns, is showing that user interface, the user interface are shown：It is described to wait for point Any one in the classification results of class image, associated images and description information or multiple, the classification results and the pass It is carried in the connection described response message of image, the description information inquires to obtain according to the classification results.

11. a kind of image classification device, which is characterized in that including：

Acquisition module, for obtaining image to be classified；

Calling module classifies to the image to be classified for calling classification model, determines the image to be classified institute The initial category of category；

The calling module is additionally operable to the target feature vector for calling Feature Selection Model to determine the image to be classified；

Contrast module is used for the target feature vector of the image to be classified and the candidate classification for initial category configuration Feature vector is compared, and comparing result is obtained；

Determining module, the classification results for determining the image to be classified according to the comparing result.

12. a kind of image classification device, which is characterized in that including：

Detection module, for detecting the trigger action for recognition button；

Acquisition module when detecting the trigger action for recognition button for the detection module, calls camera assembly to obtain Present preview image；

Determining module will be described if detecting that the present preview image is in stable state for the detection module Present preview image is determined as image to be classified；

Receiving module, the response message returned for receiving the server, the response message includes the image to be classified Classification results, the classification results be the server according to disaggregated model and Feature Selection Model to the image to be classified Obtained after classification determines.

13. a kind of server, which is characterized in that including processor and storage device, the processor is mutually interconnected with storage device It connects, wherein the storage device is for storing computer program, and the computer program includes program instruction, the processor It is configured for calling described program instruction, executes such as claim 1-8 any one of them methods.

14. a kind of user terminal, which is characterized in that including processor and storage device, the processor and storage device are mutual Connection, wherein the storage device is for storing computer program, and the computer program includes program instruction, the processing Device is configured for calling described program instruction, executes such as claim 9-10 any one of them methods.

15. a kind of computer storage media, which is characterized in that have program stored therein instruction in the computer storage media, the program Instruction is performed, and for realizing such as claim 1-8 any one of them method or is realized as claim 9-10 is any Method described in.