CN110222728A

CN110222728A - The training method of article discrimination model, system and article discrimination method, equipment

Info

Publication number: CN110222728A
Application number: CN201910404189.3A
Authority: CN
Inventors: 唐平中; 吴澄杰; 陆依鸣; 李俊
Original assignee: Turing Deep View (nanjing) Technology Co Ltd
Current assignee: Turing Deep View (nanjing) Technology Co Ltd
Priority date: 2019-05-15
Filing date: 2019-05-15
Publication date: 2019-09-10
Anticipated expiration: 2039-05-15
Also published as: CN110222728B

Abstract

The application discloses the training method, system and article discrimination method, equipment of a kind of article discrimination model, which comprises obtains the training sample image of several articles to be identified, training sample image includes true images of items and false images of items；True images of items fake at network using a pair of of antibiosis and obtains fraud image；And false images of items make really obtaining using confrontation generation network and makes true image；Fraud image is input to one and obscures differentiation e-learning to export fraud image as genuine result；And true image will be made and be input to obscure and differentiate e-learning to export and make the result that true image is vacation.The training method of article discrimination model provided by the present application, differentiate that network combines by that will fight to generate network and obscure, simultaneously using confrontation generate network and obscure differentiation network alternating iteration train by the way of, confrontation can be made to generate network and obscure and differentiate network training, common study mutually, to improve the distinguishing ability and robustness of article discrimination model.

Description

The training method of article discrimination model, system and article discrimination method, equipment

Technical field

This application involves field of image recognition more particularly to a kind of training method, system and the articles of article discrimination model Discrimination method, equipment.

Background technique

Depth learning technology is to allow machine that can handle and understand the skill of the information such as image, voice by artificial neural network Art.Deep learning be academia research popular domain, also the numerous areas such as face study, natural language processing have at The application of function.Deep learning can be accurately located and be learnt to export different types of article, and done according to its nuance Classification.In terms of image generation, confrontation, which generates network, can synthesize the figure of human eye different articles hard to tell whether it is true or false according to demand Picture.

Luxury goods skin label are the important identification points for identifying that luxury goods are true and false.Font, the imprint process of skin label are production fakements It is difficult to imitate completely.The mainly artificial identification of the mode of identification luxury goods both at home and abroad at present, there are many problems.First, None unified professional standard, industry standard of the country at present, lacks occupation qualification attestation system, it is difficult to guarantee artificial identification teacher's Identification capacity and quality；Second, artificial appraisal cost is higher and time-consuming, and high cost and time overhead directly limit luxurious The fixed application range of wasteful tasting, the Identification Service of convenient and efficient can be obtained in daily shopping by directly limiting ordinary populace；The Three, the pattern and technique of luxury goods genuine piece constantly change, and the imitated mode of fakement is also constantly changing, and luxury goods style is numerous More, the artificial identification accuracy for identifying teacher is difficult to be guaranteed.

Be therefore, how the artificial intelligence technology based on deep learning is automatic and high-accuracy identify the true of luxury goods skin label Vacation is already practitioner in the art's urgent problem to be solved.

Summary of the invention

The shortcomings that in view of the relevant technologies described above, the training side for being designed to provide article discrimination model of the application Method, system and article discrimination method, equipment, client and computer storage medium.

In order to achieve the above objects and other related objects, the first aspect of the application provides a kind of instruction of article discrimination model Practice method, comprising the following steps: obtain the training sample image of several articles to be identified, the training sample image includes true object Product image and false images of items；The true images of items fake at network using a pair of of antibiosis and obtains fraud image； And the false images of items make really obtaining using confrontation generation network and makes true image；The fraud image is defeated Enter to one to obscure and differentiates e-learning to export the fraud image as genuine result；And it makes true image by described and is input to institute It states to obscure and differentiates e-learning described to make true image be false result to export.

It further include the positioning true images of items and false images of items in the certain embodiments of the first aspect The step of middle mirror fixed position is with extraction property information.

In the certain embodiments of the first aspect, in the positioning true images of items and false images of items Reflect fixed position with the step of extraction property information for using Fast Recursive convolutional neural networks position the true images of items with And the fixed position that reflects in false images of items, and convolutional network extraction property information is separated using depth.

In the certain embodiments of the first aspect, it includes making true generator and life of faking that the confrontation, which generates network, It grows up to be a useful person, in which: for the true images of items is faked using the fraud generator the step of acquisition fraud image Obtain fraud image；The step of true image is made in the acquisition is made to make true generator described in for the false images of items True obtain makes true image.

In the certain embodiments of the first aspect, the fraud generator or to make true generator include depth residual error Network.

It further include that the step that network is trained is generated to the confrontation in the certain embodiments of the first aspect It is rapid: to make true generator using described and made the true images of items very to obtain as genuine authentication image；It is made using described True generator, which is made the fraud image, very really resets image to obtain；Calculate the true images of items and described true Authentication image and/or the image similarity for really resetting image so that it is described confrontation generate network loss function value It is maintained in desired extent.

It further include that the step that network is trained is generated to the confrontation in the certain embodiments of the first aspect Rapid: it is false authentication image that the false images of items, which is faked to obtain, using the fraud generator；It is made using described To make true image and fake to obtain be false resetting image to false generator by described；Calculate the false images of items and the vacation Authentication image and/or the false resetting image image similarity so that the confrontation generates the loss function value of network It is maintained in desired extent.

In the certain embodiments of the first aspect, it is described obscure differentiate that network includes false article arbiter and true object Product arbiter, wherein it is true that the fraud image, which is input to the false article arbiter study to export the fraud image, Result；And it makes true image by described and is input to the true article arbiter study described to make true image be false knot to export Fruit.

It further include to the step obscured and differentiate that network is trained in the certain embodiments of the first aspect It is rapid: the fraud image is input to the false article arbiter study to export the fraud image as genuine result；By institute It states true images of items and is input to the false article arbiter study to export the true images of items as the genuine result of prediction；By institute Stating false images of items and being input to the false article arbiter study to export the false images of items is the false result of prediction.

It further include to the step obscured and differentiate that network is trained in the certain embodiments of the first aspect It is rapid: the true images of items being input to the true article arbiter study and is really tied with exporting the true images of items as prediction Fruit；It is the false knot of prediction that the false images of items, which is input to the true article arbiter study to export the false images of items, Fruit；To make true image to be input to the true article arbiter study to export the fraud image be false result by described.

It further include to the step obscured and differentiate that network is trained in the certain embodiments of the first aspect It is rapid: by the true images of items be input to it is described obscure differentiate e-learning to export the true images of items as prediction and really tie Fruit；Alternatively, by the false images of items be input to it is described obscure differentiate e-learning to export the false images of items as prediction False result；By a unknown true and false image be input to it is described obscure differentiation e-learning, with export the unknown true and false image with The true images of items or the false images of items be with attribute prediction result, or the output unknown true and false image with it is described True images of items or the false images of items are different attribute prediction result；According to the same attribute prediction result or different attribute Prediction result, and the genuine result of prediction or the result of prediction vacation obtain the qualification result of the unknown true and false image.

In the certain embodiments of the first aspect, it is described obscure differentiate network include Siamese network.

In the certain embodiments of the first aspect, the article includes having the luxury goods packet of skin label or with book Write the article of person's handwriting.

The second aspect of the application provides a kind of training system of article discrimination model, comprising: sample input module is used for The training sample image of several articles to be identified is obtained, the training sample image includes true images of items and false article figure Picture；Generation module is fought, fake by the true images of items at network using a pair of of antibiosis obtains fraud image；And benefit The false images of items make really obtaining with confrontation generation network and makes true image；Obscure discrimination module, is used for institute State fraud image be input to one obscure differentiate e-learning with export the fraud image be genuine result；And it is made described very Image be input to it is described obscure differentiate e-learning described to make true image be false result to export.

It further include feature extraction unit in the certain embodiments of the second aspect, for positioning the true article Fixed position is reflected in image and false images of items with extraction property information.

In the certain embodiments of the second aspect, the feature extraction unit utilizes Fast Recursive convolutional Neural net Network positions the fixed position that reflects in the true images of items and false images of items, and special using the separable convolutional network extraction of depth Property information.

In the certain embodiments of the second aspect, the confrontation generation module includes: to make true generator, and being used for will The true images of items, which fake, obtains fraud image；Fraud generator is really obtained for make the false images of items True image must be made.

In the certain embodiments of the second aspect, the fraud generator or to make true generator include depth residual error Network.

In the certain embodiments of the second aspect, it is described make true generator be also used to by the true images of items into Row is made very to obtain as genuine authentication image；The fraud image is made and very really resets image to obtain；And meter The true images of items and the genuine authentication image and/or the image similarity for really resetting image are calculated, so that described The loss function value that confrontation generates network is maintained in desired extent.

In the certain embodiments of the second aspect, the fraud generator be also used to by the false images of items into Row fake with for；The fraud image is made very to obtain is false resetting image；And calculate the false article The image similarity of the authentication image and/or the false resetting image of image and the vacation, so that the confrontation generates network Loss function value be maintained in desired extent.

In the certain embodiments of the second aspect, described to obscure discrimination module include: false article arbiter, is used for The fraud image of input is learnt to export the fraud image as genuine result；True article arbiter, for that will input True image of making learnt to make true image described in exporting to be false result.

In the certain embodiments of the second aspect, the vacation article arbiter is also used to the true object to input Product image is learnt to predict genuine result to export the true images of items；And the false images of items to input Practising to export the false images of items is the false result of prediction.

In the certain embodiments of the second aspect, the true article arbiter is also used to the true article figure to input Genuine result is predicted as being learnt to export the true images of items；And to the false images of items of input learnt with Exporting the false images of items is the false result of prediction.

It is described to obscure discrimination module and be also used to the true images of items in the certain embodiments of the second aspect Obscure described in being input to and differentiates e-learning to export the true images of items as the genuine result of prediction；Alternatively, by the false object Product image obscures differentiation e-learning described in being input to export the result that the false images of items is prediction vacation；It is unknown true by one Fault image be input to it is described obscure differentiation e-learning, to export the unknown true and false image and the true images of items or described False images of items is with attribute prediction result, or the output unknown true and false image and the true images of items or the false object Product image is different attribute prediction result；And according to the same attribute prediction result or different attribute prediction result, and it is pre- It surveys genuine result or predicts that false result obtains the qualification result of the unknown true and false image.

In the certain embodiments of the second aspect, described to obscure discrimination module include Siamese network.

In the certain embodiments of the second aspect, the article includes having the luxury goods packet of skin label or with book Write the article of person's handwriting.

The third aspect of the application provides a kind of article discrimination method, comprising the following steps: obtains the object to be identified of shooting The image of product；Described image includes an at least objective identification point；Utilize the training of article discrimination model as described in relation to the first aspect The article discrimination model that method obtains identifies the image comprising the objective identification point；Exporting the article to be identified is Genuine result or the article to be identified are false result.

In the certain embodiments of the third aspect, the article includes having the luxury goods packet of skin label or with book The article of person's handwriting is write, the objective identification point is the position of the skin label or written handwriting.

The fourth aspect of the application provides a kind of article discrimination equipment, comprising: filming apparatus, for obtain shooting wait reflect The image of earnest product；Described image includes an at least objective identification point；Memory, for storing program code；It is one or more Processor executes the article discrimination method as described in the third aspect for calling the program code stored in the memory.

In the certain embodiments of the fourth aspect, the article includes having the luxury goods packet of skin label or with book The article of person's handwriting is write, the objective identification point is the position of the skin label or written handwriting.

The 5th aspect of the application provides a kind of article discrimination client, and the article discrimination client is loaded into an intelligence In equipment, comprising: input module calls the shooting of the smart machine to fill when for receiving the identification instruction of user's input Set the image for obtaining the article to be identified of shooting；Described image includes an at least objective identification point；Processing module calls the intelligence The program code that can store in equipment executes the article discrimination method as described in the third aspect；Display module, display output institute State that article to be identified is genuine result or the article to be identified is false result.

In the certain embodiments of the 5th aspect, the input module further includes a selecting unit, for receiving When the identification instruction inputted to user, the image of the image library of smart machine article to be identified for selection by the user is called； Described image includes an at least objective identification point.

In the certain embodiments of the 5th aspect, the article includes having the luxury goods packet of skin label or with book The article of person's handwriting is write, the objective identification point is the position of the skin label or written handwriting.

The 6th aspect of the application provides a kind of computer readable storage medium, is stored with the computer for article discrimination Program, the computer program are performed the training method for realizing article discrimination model as described in relation to the first aspect.

The 7th aspect of the application provides a kind of computer readable storage medium, is stored with the computer for article discrimination Program, the computer program are performed the article discrimination method realized as described in the third aspect.

In conclusion training system, the object of the training method of article discrimination model provided by the present application, article discrimination model Product discrimination method, article discrimination equipment, article discrimination client and computer readable storage medium, have the following beneficial effects: It will fight to generate network and obscure and differentiate that network combines, make true generator and fraud generator pair using confrontation generation network Training sample image make true or is faked, and will be made true or after faking image and is input to and obscure differentiation network, thus to obscuring The true article arbiter and false article arbiter for differentiating network are trained, and can be improved the distinguishing ability obscured and differentiate network. Meanwhile by the way of generating network using confrontation and obscuring differentiation network alternating iteration training, confrontation can be made to generate network and mix Confuse and differentiate network training, common study mutually, while improving the robustness and accuracy of the two, and then effectively improve article mirror The distinguishing ability and robustness of other model.It is obtained using the training method training by article discrimination model provided by the present application Article discrimination model identifies article, and accuracy rate is high, and stability is strong.

Detailed description of the invention

Fig. 1 is shown as the flow diagram of the training method of article discrimination model in the application in one embodiment.

Fig. 2 is shown as the schematic illustration of the training method of article discrimination model in the application in another embodiment.

Fig. 3 is shown as confrontation in the application and generates the schematic illustration of the training process of network in one embodiment.

Fig. 4 is shown as confrontation in the application and generates the schematic illustration of the training process of network in another embodiment.

Fig. 5 is shown as the schematic illustration of the training method of article discrimination model in the application in another embodiment.

Fig. 6 is shown as obscuring the schematic illustration of the training process for differentiating network in one embodiment in the application.

Fig. 7 is shown as obscuring the schematic illustration of the training process for differentiating network in another embodiment in the application.

Fig. 8 is shown as obscuring the schematic illustration of the training process for differentiating network in another embodiment in the application.

Fig. 9 is shown as obscuring schematic illustration of the training process for differentiating network in another embodiment in the application.

Figure 10 is shown as the structural schematic diagram of the training system of article discrimination model in the application in one embodiment.

Figure 11 is shown as the flow diagram of article discrimination method in one embodiment in the application.

Figure 12 is shown as the structural schematic diagram of article discrimination equipment in one embodiment in the application.

Figure 13 is shown as the structural schematic diagram of the application article discrimination client in one embodiment.

Specific embodiment

Presently filed embodiment is illustrated by particular specific embodiment below, those skilled in the art can be by this explanation Content disclosed by book understands other advantages and effect of the application easily.

Although term first, second etc. are used to describe various elements or parameter herein in some instances, this A little elements or parameter should not be limited by these terms.These terms be only used to by one or parameter part and another or parameter into Row is distinguished.For example, the first Euclidean distance can be referred to as the second Euclidean distance, and similarly, the second Euclidean distance can be by Referred to as the first Euclidean distance, without departing from the range of various described embodiments.First Euclidean distance and the second Euclidean distance It is the value of Euclidean distance being calculated in description two images, but unless context otherwise explicitly points out, it is no Then they are not the same Euclidean distances, do not have identical value yet.Similar situation further includes first distance and second distance Device or the first arbiter and the second arbiter.

Furthermore as used in herein, singular " one ", "one" and "the" are intended to also include plural number shape Formula, unless there is opposite instruction in context.It will be further understood that term "comprising", " comprising " show that there are the spies Sign, step, operation, element, component, project, type, and/or group, but it is not excluded for one or more other features, step, behaviour Presence, appearance or the addition of work, element, component, project, type, and/or group.Term "or" used herein and "and/or" quilt It is construed to inclusive, or means any one or any combination.Therefore, " A, B or C " or " A, B and/or C " mean " with Descend any one: A；B；C；A and B；A and C；B and C；A, B and C ".Only when element, function, step or the combination of operation are in certain sides When inherently mutually exclusive under formula, it just will appear the exception of this definition.

The increasingly developed Sales Channel for not only enriching article of electronic trade platform, also provides to forge or copying article Convenient and fast sales platform, user must not be not concerned about the true or false of bought article.However, article discrimination ability and non-generic disappearing What expense person was easy to have, similarly, as items sold platform, the true and false of article is not that electronic trade platform is easy to audit yet 's.As luxury goods of some fashion, such as packet, table luxury goods etc., article discrimination people needs to lay in a large amount of relative article knowledge, And be familiar with the used material of article, workmanship process, theme feature etc., therefore, it has been unable to satisfy otherwise using artificial mirror The user of high-end article is bought to demands such as article discrimination accuracy, timeliness.

Deep learning is the algorithm that a kind of pair of data carry out representative learning in machine learning.Derived from grinding for artificial neural network To study carefully, motivation is to establish the neural network that simulation human brain carries out analytic learning, by imitating the mechanism of human brain come interpretation of images, Sound and text data.Deep learning, which forms more abstract high level by combination low-level feature, indicates attribute classification or feature, To find that the distributed nature of data indicates.Deep learning has become a kind of representative machine learning techniques, is widely used In fields such as image, sound, texts, however in existing research patent and document, it there is no and identified using depth learning technology The precedent of luxury goods skin label.The application creatively uses the artificial intelligence technology based on deep learning, in conjunction with depth convolution net Network and confrontation generate network article discrimination model is trained, and using trained article discrimination model to images of items into Row identifies, and can identify the true and false of luxury goods skin label and other items automatically.

Referring to Fig. 1, it is shown as the flow diagram of the training method of the application article discrimination model in one embodiment, As shown, the training method of the article discrimination model the following steps are included:

Step S11, obtains the training sample image of several articles to be identified, and the training sample image includes true article figure Picture and false images of items.

The article to be identified includes the luxury goods packet with skin label or the article with written handwriting.For example, identifying certain The bag with skin label such as the both shoulders packet of brand, shoulder bag, handbag, oblique satchel, traveling bag；Such as bracelet, ring, wrist-watch, Pocket-watch, contract etc. have the article of written handwriting.Wherein, in some embodiments, the luxury goods packet is, for example, Ai Mashi (Hermes) packet, Louis Vuitton (Louis Vuitton) packet, Armani (Giorgio Armani), Bao Geli (BVLGARI), Findi (FenDi), Balenciaga (BALENCIAGA), it is fragrant how (CHANEL) packet, Prada (PRADA) packet, Preserve butterfly man (BottegaVeneta) packet, Gucci (Gucci) packet, Kou Chi (COACH) packet and Christian Dior (Dior) packet etc..At this Apply in the embodiment provided, is temporarily illustrated so that the article to be identified is luxury goods packet as an example.

It is readily appreciated that, training sample image is made of the multiple image obtained, includes true article figure in the multiple image Picture and false images of items, as training sample, the true images of items for article to be identified is accredited as genuine piece and in advance to institute The image of genuine piece shooting is stated, the vacation images of items is that article to be identified is accredited as fakement and to fakement shooting in advance Image.In embodiment, the described prior identification can manually identify or technical appraisement by way of.

The true images of items and false images of items, which can be, carries out shooting by the picture pick-up device of intelligent terminal to directly It obtains, is also possible to obtain by network transmission.In some embodiments, true images of items and false article figure are obtained The mode of picture includes but is not limited to: being treated by using the picture pick-up device of the intelligent terminals such as mobile phone, tablet computer, laptop Identify article to be shot, by the image of shooting directly as true images of items or false images of items；In some embodiments, By having identified true and false images of items on downloading webpage or on server as true images of items or false images of items；? In some embodiments, by the received image of the modes such as wire communication or wireless communication as true images of items or false article figure Picture.The format of described image stores the format of image, for example, bmp, jpg, png, tiff, gif, pcx, tga for computer, The format of the storages such as exif, fpx, svg, psd, cdr, pcd, dxf, ufo, eps, ai, raw, WMF.

In practical applications, on the one hand, image may be different with the difference of shooting condition.For example, in the sun The images of items to be identified shot under the images of items to be identified of shooting and light indoors is different, furthermore, the face of light Color can also influence image.On the other hand, according to article to be identified using the length of time or the difference of usage degree, The image of shooting is also different.For example, the article that uses of long term frequent may than the degree of wear weight for the article just purchased, or Person experienced slacking and make color burn, or be stained with spot during use etc..

Therefore, in some embodiments, expand the image of acquisition the quantity of image using data enhancement methods, with Increase the robustness of training.The data enhancing includes that such as mirror symmetry, random cropping, rotation, stretching, diminution, distortion become Shape, local buckling, color conversion, the covering of random color lump, addition noise etc..For example, respectively to R, G and B in piece image Different distortion values is added on three channels, to obtain the image after color conversion, and will be after original image and color conversion Image is all used as training sample image.The channel refers to RGB channel or the channel CMYK in image, such as a width 64 × 64 image of RGB can be indicated with one 64 × 64 × 3 vector；Wherein, " 3 " are used to indicate channel, respectively Red (Red), green (Green), blue three channels (Blue).For another example, according to different noise types, respectively by noise with Machine number adds on multiple image, and all regard the multiple image after original image and addition noise as training sample image.

In practical applications, there are the size of a small number of images of acquisition and resolution ratio be very big or very small situation, If it not being screened or being pre-processed and be directly inputted into neural network and be trained, article discrimination may be will affect The accuracy of model.Therefore, in some embodiments, further include the steps that screening the image of acquisition, rejecting and object The image that the input picture of other model is not adapted to is judged, to ensure the quality of training sample image, increases article discrimination model instruction Experienced stability.

Since the image that every width obtains is not necessarily identical in the parameters such as clarity, image size, resolution ratio, in some realities It applies in mode, acquired all images or the image retained after screening is identified for the ease of subsequent step, institute The training method for stating article discrimination model further includes carrying out pretreated step to several described training sample images.Wherein, institute Pre-treatment step is stated for acquired image to be adapted to neural network model input picture to be transferred, to be instructed Practice sample image.The pretreated step includes carrying out modification of dimension, bi-directional scaling to described image plus making an uproar, invert, revolve Turn, translation, scale transformation, shearing, contrast variation, random channel offset, filter one of or a variety of processing.For example, institute Stating pretreatment includes the image rejected picture size and be less than pre-set dimension threshold value.For another example, the pretreatment includes to acquired Image carries out clarity analysis and clarity is selected to meet the image of default clarity condition as training sample image.For another example, It is described pretreatment include by received 550 × 300 pixel image rotation at 300 × 550.For another example, described pre-process includes By received 1000 × 800 pixel image down or shearing after, then rotate to 300 × 550.For another example, the pretreatment packet The processing such as be sharpened plus make an uproar to the acquired image by U.S. face is included to restore true picture.For another example, described pre-process includes It identifies the article profile in image, image is rotated, translated and is sheared in the position of entire image according to images of items, with It reduces the background in images of items to the greatest extent and obtains meeting the image that subsequent confrontation generates the received size of network.For another example, described Pretreatment includes carrying out gray scale overturning to acquired each image, in order to subsequent inhibition background or to prominent article characteristics.

In an actual application scenarios, luxury goods are really wrapped and the overall appearance of false packet is quite similar, from overall appearance On carry out identify difficulty it is higher, therefore the point of the identification such as the font of its skin label, imprint process be the key that identify it is true and false.Cause This obtains the training sample image comprising article to be identified in some embodiments, further includes the positioning true images of items And the fixed position that reflects in false images of items with extraction property information the step of, the characteristic information includes font, stroke, line Reason, relative position and size etc..It is to be understood that the mirror fixed position of article to be identified also can be different, example due to the difference of article Ms's packet such as luxury goods, then its identification point generally comprises skin label, zipper and the label including article trade mark；Again Such as wrist-watch, then the fixed position that reflects is generally placed upon watchband, dial plate etc.；The printed matters such as contract for another example, then identify point It sets and is generally placed upon written handwriting or seal etc..

In actual implementation state, the mode for positioning the mirror fixed position includes but is not limited to: 1) artificial calibration.Example Such as, a width is shot using mobile phone and include the image of luxury goods packet (corresponding article to be identified), and draw take a certain size on the image Range as mirror fixed position, include the skin label (corresponding identification point) of luxury goods packet in the range.For another example, more in acquisition After width training sample image, artificially whole training sample images are identified and demarcated one by one.2) pass through recursive convolution mind Position through network positions identification point.For example, using Fast Recursive convolutional neural networks, first in the true images of items of every width and vacation The various dimensions feature for obtaining image there may be the candidate region of mirror fixed position is found in images of items, then utilizes target The sub-networks such as detection differentiates, search box position returns obtain the accurate location of mirror fixed position.

It is postponed in positioning identification point, the mode of extraction property information is believed including the use of depth convolutional network extraction property Breath.The depth convolutional network include Xception (depth separates convolutional network), ResNet (depth residual error network), MobileNet etc..For example, can be divided setting in the image of the luxury goods packet of skin label position using depth for width identification point From convolutional network from characteristic informations such as skin label extracting section font, stroke, texture, relative position and sizes in image.Depth Separable convolutional network carries out convolution to each channel of image respectively, and by the spatial position of characteristic information in image and respectively The convolution of interchannel is decoupling, to improve the ability of extraction property information, and is more advantageous to network and is learnt and optimized.

Referring to Fig. 2, being shown as the principle signal of the training method of article discrimination model in the application in one embodiment Figure, as shown, for example, the training sample image that a width includes luxury goods packet is input to a recursive convolution neural network In, the recursive convolution neural network find in the picture there may be mirror fixed position the candidate region (candidate region The region of rectangle frame delimitation is shown as in Fig. 2).After determining multiple candidate regions, the recursive convolution neural network will be generated The image comprising candidate region be input in a depth convolutional network, the depth convolutional network using target detection differentiate, The sub-networks such as search box position recurrence analyze the candidate region in image, to obtain the accurate position of mirror fixed position It sets.

Step S12, fake by the true images of items at network using a pair of of antibiosis obtains fraud image；And benefit The false images of items make really obtaining with confrontation generation network and makes true image.

In order to further enhance the distinguishing ability of the article discrimination model, neural network is allowed really to learn to obtain luxury goods Nuance between packet genuine piece and the skin label of fakement, the embodiment of the present application is using a pair of of antibiosis at network to the training sample Image is trained.In the training process, the true images of items and false images of items are input to the confrontation and generate network, The confrontation generates network and carries out acquisition fraud image of faking to the true images of items, and carries out to the false images of items It makes true obtain and makes true image.

It should be strongly noted that the fraud refers to carrying out image procossing to the true images of items to be faked The process of image.In some embodiments, it is described make it is true or fake include to the false images of items or true images of items into Row is at the image such as mirror symmetry, rotation, torsional deformation, local buckling, color conversion, the covering of random color lump, addition noise Reason.For example, the processing on any image made for the image comprising a genuine piece luxury goods packet (such as stretches, cuts, is colored Transformation etc.), it can not influence the true and false of luxury goods packet material object.Therefore, no matter how to fake to the true images of items, The corresponding article of fraud image of acquisition remains as genuine piece, that is to say, that the fraud image substantially remains as the true object of a width Product image.If identifying to the fraud image, the result identified is remained as very.Similarly, described make really refers to pair The vacation images of items carries out image procossing to obtain the process for making true image.Likewise, no matter to the false images of items such as What make true, and the corresponding article of true image of making obtained remains as fakement, described to make true image and substantially remain as a width False images of items；If making true image to described and identifying, the result identified remains as vacation.

In some embodiments, it includes making true generator and fraud generator that the confrontation, which generates network, described to make very Generator and fraud generator can generate according to the true images of items or false images of items of input and be different from the true article Image or false images of items make true image or fraud image.Wherein, the step of acquisition fraud image is described in utilization The true images of items fake by fraud generator obtains fraud image；It is described to obtain the step of making true image to utilize institute It states to make true generator and carry out the false images of items to make true obtain and makes true image.

Usually the number of plies of network more multiple-effect fruit is better in deep learning, i.e., its deeper training of network and test is accurate Rate is higher.But in practical applications due to there is gradient disappearance, when the number of plies of network increases to certain amount, training Decline instead with the accuracy rate of test.In order to solve the problems, such as that gradient disappears while increasing network depth, thus further Ground promotes the accuracy rate of article discrimination model, in some embodiments, the fraud generator or to make true generator include deep Spend residual error network.In some embodiments, the depth residual error network that multiple residual block connection compositions can be used is true as making Generator and fraud generator, the quantity of the residual block can be 8,10,18,34 etc..In the embodiment of the present application The quantity of the residual block is illustrative only, and is not limited its scope.

For neural network model, complicated model tends to over-fitting.Over-fitting refers to model for instruction Practice effect of the sample set as input when far better than using test sample collection as effect when input.Therefore, it is necessary to model Loss function weighed, thus make model have better generalization ability, improve the robustness of model.

Therefore, in some embodiments, further include the steps that generating network to the confrontation is trained.Please refer to figure 3, it is shown as confrontation in the application and generates the schematic illustration of the training process of network in one embodiment, as shown in figure 3, packet It includes: making true generator using described and made the true images of items very to obtain as genuine authentication image, made using described True generator, which is made the fraud image, very really resets image to obtain；Calculate the true images of items and described true Authentication image and/or the image similarity for really resetting image so that it is described confrontation generate network loss function value It is maintained in desired extent.

Described in brought forward, the true images of items in training sample image is input to the fraud generator and is faked, institute It states fraud generator and generates fraud image.By the fraud image be input to it is described make true generator, it is described that make true generator raw Image is really reset at one.Since true images of items remains as very in the image generated after making very, the fraud figure As being that very, the resetting image is really to reset image.True images of items in training sample image is input to described make very Generator make very, described to make true generator generation authentication image.Similarly, true images of items is in the figure generated after making very It is true as remaining as, therefore the authentication image generated is genuine authentication image.

The calculating of described image similarity includes but is not limited to: calculating the true images of items and the genuine authentication image And/or it the Euclidean distance for really resetting image, absolute value distance, histogram intersection, mahalanobis distance, manhatton distance, cuts Than snow husband's distance, Minkowski Distance, standardization Euclidean distance, included angle cosine, central moment etc..In implementation provided by the present application In example, temporarily it is illustrated for calculating Euclidean distance.When calculating the similarity of image using Euclidean distance, the Euclidean distance Refer to the actual range in hyperspace between two points, Euclidean distance is smaller, and the similarity of two images is bigger.

Calculating the true images of items and the genuine authentication image and/or the Euclidean distance for really resetting image has Various ways, including but not limited to: for example, calculate separately the first Euclidean of the true images of items and genuine authentication image away from From and the true images of items with really reset the second Euclidean distance of image, and take first Euclidean distance and second The mean value or root mean square of Euclidean distance, to obtain target Euclidean distance, and judged according to the target Euclidean distance it is described right Whether antibiosis is maintained in desired extent at the loss function value of network；For another example, the true images of items and genuine is calculated separately First Euclidean distance of authentication image and the true images of items and the second Euclidean distance of image is really reset, and takes institute The larger value in the first Euclidean distance and the second Euclidean distance is stated, and using the corresponding Euclidean distance of described the larger value as target Europe Family name's distance, and then whether it is maintained at expected according to the loss function value that the target Euclidean distance judges that the confrontation generates network In range.

Similarly, in some embodiments, referring to Fig. 4, being shown as confrontation in the application generates training for network The schematic illustration of journey in another embodiment is also wrapped as shown in figure 4, generating the step of network is trained to the confrontation Include: it is false authentication image that the false images of items, which is faked to obtain, using the fraud generator, is made using described To make true image and fake to obtain be false resetting image to false generator by described；Calculate the false images of items and the vacation Authentication image and/or the false resetting image image similarity so that the confrontation generates the loss function value of network It is maintained in desired extent.

By the false images of items in training sample image be input to it is described make true generator make it is true, it is described to make true generation True image is made in device generation.It makes true image by described and is input to the fraud generator, the fraud generator generates a false weight Set image.Since false images of items remains as vacation in the image generated after making very, it is described make true image be it is false, it is described Resetting image is false resetting image.False images of items in training sample image is input to the fraud generator to make Vacation, the fraud generator generate authentication image.Similarly, false images of items remains as vacation in the image generated after faking, Therefore the authentication image generated is false authentication image.

The confrontation is generated described in the step of network is trained and above embodiment in the present embodiment Step is similar, and principle and process please refer to above-described embodiment, and details are not described herein again.

It should be understood that if thrown the reins to the training for making true generator and fraud generator, it is described to make true generator It carries out making true at random with fraud generator or fake, it is likely that will appear the case where excessively making true or excessive fraud.Such as it is described What true images of items included is a luxury goods wallet, then its corresponding mirror fixed position includes skin label, brand trademark etc., if gone out It now excessively fakes or excessively makes genuine situation, in the fraud image of generation, it is likely that skin label therein are by image procossing at another A kind of article unrelated with skin label, even all unrelated with wallet.In this case, a kind of and object to be identified is generated by faking The unrelated image of product, the image have lacked required attribute or feature in training sample image, are reflected with the image to article It is meaningless that other model, which is trained,.Therefore, in the training process, the certification figure of the false images of items and the vacation is calculated The image similarity of picture and/or the false resetting image, so that the loss function value that the confrontation generates network is maintained at pre- Within the scope of phase.The loss function is used to estimate the predicted value of model and the inconsistent degree (or departure degree) of true value, The value of loss function is smaller, and the robustness of model is better.The loss function value includes L1 norm loss function, intersects entropy loss Function, log logarithm loss function, quadratic loss function, figure penalties function, Hinge loss function, absolute error loss function Deng.After calculating acquisition target Euclidean distance, the loss letter that the confrontation generates network is obtained according to the target Euclidean distance Numerical value, and the loss function value being calculated is compared with preset value or preset range, it is described to antibiosis to judge Whether it is maintained in desired extent at the loss function value of network, to prevent from described making true generator and fraud generator is excessive It makes true or fakes.

The cross entropy loss function is used to determine the degree of closeness of actual value and desired value.Cross entropy refers to actual value At a distance from desired value, cross entropy is smaller, and the probability distribution of actual value and desired value is with regard to closer.In some embodiments, It loses using the circulation loss and self loss in L1 norm loss function, and using the confrontation in cross entropy loss function Network is generated to the confrontation to be trained, and can further enhance the distinguishing ability and accuracy of the article discrimination model.

Step S13, it is true that the fraud image, which is input to one to obscure differentiation e-learning to export the fraud image, Result；And by it is described make true image be input to it is described obscure differentiate e-learning described to make true image be false knot to export Fruit.

The fraud image that obtains after the true images of items is faked and the false images of items make true Obtain afterwards make true image be input to one obscure differentiate network in, obscure differentiate network to the fraud image and make true image into Row differentiates study, exports that the fraud image is genuine result and described to make true image be false result.

In some embodiments, it is described obscure differentiate network obscure including the use of the structure of Siamese network described Differentiate that network is trained and optimizes.For example, using one true one false image of two width comprising skin label as input, in addition to exporting two width Outside the qualification result of image, also output is with attribute prediction result or different attribute prediction result, for indicating the two images Whether same class (true or false) or a true vacation are belonged to.Thus, it is possible to further enhance the identification of article discrimination model Ability.

Described in brought forward, no matter how to fake to the true images of items, the corresponding article of fraud image obtained Remain as genuine piece.Therefore, using it is described obscure differentiate that network differentiate that the result of output remains as described make to fraud image Fault image is true.Similarly, no matter how the false images of items is carried out making very, what is obtained makes the corresponding article of true image Remain as fakement.Therefore, using it is described obscure differentiate that network differentiate that the result of output remains as described make to true image is made True image is false.

In some embodiments, it is described obscure differentiate that network includes false article arbiter and true article arbiter, wherein The fraud image is input to the false article arbiter study to export the fraud image as genuine result；And by institute It states and makes true image and be input to the true article arbiter study described to make true image be false result to export.

Referring to Fig. 5, being shown as the principle signal of the training method of the application article discrimination model in another embodiment Figure, wherein when what solid line with the arrow indicated in figure is that true images of items is as training sample image input in one embodiment Training process, when what dotted line with the arrow indicated is that false images of items is inputted as training sample image in one embodiment Training process.As shown in the figure, on the one hand, the true images of items is input to the fraud generator in confrontation generation network In, the fraud generator generates fraud image, then the fraud image is input to the false object of one obscured differentiate in network Product arbiter, the vacation article arbiter differentiate the fraud image.It can be seen that, faked to true images of items by aforementioned Obtained image remains as very, and therefore, it is genuine result that the vacation article arbiter, which exports the fraud image,.On the other hand, One that the fraud image that the fraud generator generates is input in confrontation generation network makes in true generator, and described make really is given birth to Generation one of growing up to be a useful person really resets image；By the true images of items be input to it is described make in true generator, it is described to make true generator Generate a genuine authentication image.Image and/or genuine authentication image are really reset with described by calculating the true images of items Between image similarity, whether be maintained at expected model to determine described in training process that confrontation generates the loss function value of network In enclosing, if loss function value exceeds desired extent, the parameter that confrontation is generated to network is needed to be adjusted, to prevent described It makes true generator and fraud generator is excessively made true or fakes, and then lead to the accuracy for influencing the article discrimination model.

Similarly, the one false images of items is input in confrontation generation network makes in true generator, described to make very Generator generation make true image, then by it is described make true image be input to obscure differentiate network in a true article arbiter, institute True article arbiter is stated to make true image to described and differentiate.It can be seen that, the image really obtained is made to false images of items by aforementioned Vacation is remained as, therefore, making true image described in the true article arbiter output is false result.On the other hand, it is made described very What generator generated makes in the fraud generator that true image is input in confrontation generation network, and the fraud generator generates one False resetting image；The false images of items is input in the fraud generator, the fraud generator generates a vacation Authentication image.By calculating the image between the false images of items and the authentication image of false the resetting image and/or vacation Whether similarity is maintained in desired extent with the loss function value for determining that confrontation generates network described in training process.

In some embodiments, further include the steps that it is described obscure differentiate network be trained: by frauds figure As being input to the false article arbiter study to export the fraud image as genuine result；The true images of items is inputted Genuine result is predicted to the false article arbiter study to export the true images of items；The false images of items is inputted To the false article arbiter study to export the result that the false images of items is prediction vacation.Fig. 6 and Fig. 7 are please referred to, Fig. 6 is aobvious It is shown as the application and obscures the schematic illustration of the training method of differentiation network in one embodiment, Fig. 7, which is shown as the application, to be obscured Differentiate the schematic illustration of the training method of network in another embodiment.As shown in fig. 6, in the training process, will be labeled as Really fraud image is input in the false article arbiter, and the vacation article arbiter exports genuine result；It will be labeled as true True images of items be input in the false article arbiter, genuine result is predicted in the vacation article arbiter output；It will label It is that false false images of items is input in the false article arbiter, the false result of the vacation article arbiter output prediction.It is aobvious So, the genuine result of prediction should be true really, and the false result of the prediction should be false really.

True generator is made and fraud generator is trained similar with to described, is sentenced to false article arbiter and true article It when other device is trained, also needs to use restraint to training process, so that the vacation article arbiter and true article Arbiter improves the accuracy differentiated, while training convergence being kept to stablize.It therefore, in some embodiments, further include by institute It states the genuine result of prediction to be compared with true images of items, by calculating between prediction genuine result and the true images of items Distance, and determine whether the distance is less than preset value.It similarly, in some embodiments, further include that the prediction is false Result be compared with false images of items, by calculating the distance between the false result of the prediction and false images of items, with Determine whether the distance is less than preset value.In some embodiments, further include by the genuine result and fraud image into Row compares, by calculating the distance between the genuine result and fraud image, to determine whether the distance is less than preset value.

In a specific embodiment, it is false that the genuine result, false result, the genuine result of prediction and prediction are set Result be [0,1] range in a real number.Wherein, described " 0 " indicates to correspond to false label, and described " 1 " indicates to correspond to For the label of "true".Described in brought forward, before being input to false article arbiter, the fraud image, true images of items and vacation Images of items is marked by way of handmarking or technical mark, it is clear that the fraud image and true article Image tagged is " 1 ", and representative image is that very, the vacation images of items is labeled as " 0 ", and representative image is false.Under the above conditions, The vacation article arbiter or the result of true article arbiter output are also a real number in [0,1] range.It is arranged described pre- If value, for example, setting 0.5 for preset value.Obviously, if the vacation article arbiter or the result of true article arbiter output are small In 0.5, then illustrate the image of input be it is false, it is on the contrary then illustrate to be true.When fraud image is as inputting, the false article is sentenced The genuine result of other device output is compared with the mark value of fraud image, calculates first distance between the two.Similarly, The second distance between the genuine result of the prediction and the mark value of true images of items is calculated, and calculates the false knot of the prediction Third distance between fruit and the mark value of false images of items.The first distance, second distance and the are obtained by calculating The distance of three distances and, and by the distance and minimized by the optimization algorithms such as Adam, SGD, thus to the false object Product arbiter is trained.

Similarly, as shown in fig. 7, in some embodiments, further include to it is described obscure differentiate that network is trained Step: it is genuine for prediction to export the true images of items that the true images of items is input to the true article arbiter study As a result；It is that prediction is false that the false images of items, which is input to the true article arbiter study to export the false images of items, As a result；To make true image to be input to the true article arbiter study to export the fraud image be false result by described.

Due to Fig. 7 it is corresponding to state obscure differentiate network be trained the step of it is similar to the above, principle and process are asked Referring to above-described embodiment, details are not described herein again.

In some embodiments, further include the steps that it is described obscure differentiate network be trained:

Step S131, by the true images of items be input to it is described obscure differentiate e-learning to export the true article figure As being to predict genuine result；Alternatively, by the false images of items be input to it is described obscure differentiate e-learning to export the vacation Images of items is the false result of prediction.

Step S132, by a unknown true and false image be input to it is described obscure differentiation e-learning, it is described unknown true to export Fault image and the true images of items or the false images of items are with attribute prediction result, or the output unknown true and false figure As being different attribute prediction result with the true images of items or the false images of items.

Step S133, according to the same attribute prediction result or different attribute prediction result, and the genuine result of prediction or The false result of prediction obtains the qualification result of the unknown true and false image.

Please refer to Fig. 8 and Fig. 9, Fig. 8 is shown as obscuring in the application the training process for differentiating network in one embodiment Schematic illustration, Fig. 9 are shown as obscuring the schematic illustration of the training process for differentiating network in another embodiment in the application. As shown in figure 8, by the true images of items be input to it is described obscure differentiate in network and carry out differentiation study, due to the true article It is priori result (the i.e. described true images of items be previously determined to be genuine training sample image) that image, which is genuine result, therefore institute It states to obscure and differentiates that the result that network exports is that prediction is genuine as a result, indicating that the true images of items is true.Then, unknown true by one Fault image be input to it is described obscure differentiate that network carries out differentiation study.Since the unknown true and false image is only included as true or is false Two kinds as a result, when the unknown true and false image is true images of items, the prediction knot of prediction result and the true images of items Fruit attribute is identical, that is, is true；When the unknown true and false image is false images of items, prediction result and the true article figure The prediction result attribute of picture is different, i.e., one true one is false.Therefore, it is described obscure differentiate network differentiation is carried out to unknown true and false image After habit, exports the unknown true and false image and the true images of items is with attribute prediction result or different attribute prediction result. Due to the true and false of the true images of items it is known that being to be predicted with attribute prediction result or different attribute according to output result therefore As a result, can determine that the unknown true and false image is true or is false.

Similarly, as shown in figure 9, by the false images of items be input to it is described obscure differentiate in network and carry out differentiation It practises, the result for obscuring differentiation network output is prediction vacation as a result, indicating that the vacation images of items is false.Then, by one Unknown true and false image be input to it is described obscure differentiate that network carries out differentiation study, similarly, when the unknown true and false image is false object When product image, it is all false that prediction result is identical as the false prediction result attribute of images of items；When described unknown true and false When image is true images of items, prediction result is different from the false prediction result attribute of images of items, i.e., one true one is false.By Similar to the above embodiments in principle, detailed process and step please refer to above-described embodiment, and details are not described herein again.

In the embodiment of the present application, by the confrontation generate network and it is described obscure differentiate network alternately train, the confrontation Generate in network make true generator and how the study of fraud generator is mutually converted true images of items and false images of items, it is raw Differentiation network is obscured at the image mixed the spurious with the genuine to confuse.And the true article arbiter obscured in differentiation network and false article Arbiter then constantly learn input image be very still be it is false, avoid being fought and generate network fascination, to improve identification Ability.Network and the training method for obscuring differentiation network loop iteration are generated by the confrontation, neural network can be made It constantly improve and optimizes in the training process, to improve the distinguishing ability of article discrimination model.

In some embodiments, can also using Adam (Adaptive Moment Estimation) optimization algorithm, SGD (Stochastic Gradient Descent) optimization algorithms such as optimization algorithm or RMSprop optimization algorithm are to nerve net Network optimizes, and neural network weight is iteratively updated based on training data, so that neural network cross entropy loss function is minimum Change, convergence is more stable, while neural network being avoided to fall into local optimum, to enhance the robustness of article discrimination model.

The training method of article discrimination model provided by the present application fights the structure building article for generating network based on circulation Identify model, confrontation generation network is combined with differentiation network is obscured, generates making true generator and making for network using confrontation False generator make true or is faked to training sample image, will be made true or after faking image and is input to and obscures differentiation network, To be trained to the true article arbiter and false article arbiter of obscuring differentiation network, it can be improved and obscure differentiation network Distinguishing ability.Meanwhile by the way of generating network using confrontation and obscuring differentiation network alternating iteration training, confrontation can be made to generate Network and obscure the robustness and accuracy for both differentiating network training, common study mutually, while improving, and then effectively mentions The distinguishing ability and robustness of high article discrimination model.

The training method of article discrimination model described herein is executed by the training system of an article discrimination model. Referring to Fig. 10, the structural schematic diagram of the training system of the application article discrimination model in one embodiment is shown as, as schemed institute Show, the training system 100 includes sample input module 101, confrontation generation module 102 and obscures discrimination module 103, wherein The training system of the article discrimination model includes to be realized by the software and hardware in computer equipment.

The computer equipment can be any calculating equipment with mathematics and logical operation, data-handling capacity, Including but not limited to: PC device, single server, server cluster, Distributed Services end, the cloud server terminal Deng.Wherein, the cloud server terminal includes that public cloud (Public Cloud) server-side and private clound (Private Cloud) are serviced End, wherein described public or privately owned cloud server terminal include Software-as-a-Service (software services, abbreviation SaaS), Platform-as-a-Service (platform services, abbreviation PaaS) and (basis Infrastructure-as-a-Service Facility services, abbreviation IaaS) etc..The privately owned cloud server terminal such as Ali's cloud computing service platform, Amazon (Amazon) Cloud computing service platform, Baidu's cloud computing platform, Tencent's cloud computing platform etc..

Wherein, the cloud server terminal provides at least one remote image upload service.The remote image upload service packet Include but be not limited to following at least one: the service etc. that article restocking service, the service of article discrimination and article are complained.Wherein, institute The service such as businessman for stating article restocking uploads item for sale image and related text description etc., the service example of the article discrimination Such as buyer uploads images of items to discern the false from the genuine, and the service that the article is complained is, for example, that buyer can not reach with businessman The service that images of items is reconciled for third party (such as electronic trade platform) intervention is uploaded when consistent.

The computer equipment includes at least: memory, one or more processors, I/O interface, network interface and input Structure etc..Wherein the memory is used to store the multiple image and at least one program of article to be identified.The memory It may include high-speed random access memory, and may also include nonvolatile memory, such as one or more disk storages are set Standby, flash memory device or other non-volatile solid-state memory devices.

Memory is for storing program code.Memory may include volatile memory (Volatile Memory), example Such as random access memory (Random Access Memory, RAM)；Memory also may include nonvolatile memory (Non-Volatile Memory), such as read-only memory (Read-Only Memory, ROM), flash memory (Flash Memory), hard disk (Hard Disk Drive, HDD) or solid state hard disk (Solid-State Drive, SSD)；Memory may be used also With include mentioned kind memory combination.Memory can be used for storing batch processing code, in order to which processor calling is deposited The program code stored in reservoir with realize the sample input module involved in the embodiment of the present application, confrontation generation module with And any one of obscure the functional modules such as discrimination module or multinomial function.

Processor can be made of one or more general processor, such as central processing unit (Central Processing Unit, CPU).Processor can be used for running any one of following or multiple function module in relevant program code Program: sample input module fights generation module and obscures discrimination module etc..That is, processor executes program generation Any one of following or multiple function module function may be implemented in code: sample input module fights generation module and obscures and sentences Other module etc..Wherein, about the sample input module, fight generation module and obscure discrimination module for details, reference can be made to preceding State the related elaboration in embodiment.

In certain embodiments, memory can also include the memory far from one or more processors, such as via The network attached storage of RF circuit or outside port and communication network access, wherein the communication network can be Yin Te Net, one or more intranet, local area network (LAN), wide area network (WLAN), storage area network (SAN) etc. or its is appropriately combined. Access of the other assemblies of such as CPU and Peripheral Interface of Memory Controller controllable device etc to memory.Memory Optionally include high-speed random access memory, and optionally further comprising nonvolatile memory, such as one or more magnetic Disk storage device, flash memory device or other non-volatile solid state memory equipment.By the other assemblies such as CPU and periphery of equipment Interface controls the access of memory alternately through Memory Controller.

One or more of processors are operationally coupled with network interface, will be calculated equipment and be coupled by correspondence To network.For example, network interface, which can will calculate equipment, is connected to local area network (such as LAN), and/or wide area network (such as WAN).Processor It is also operatively coupled to the port I/O and input structure, which aloows calculating equipment to set with various other electronics Standby to interact, which aloows user to interact with equipment is calculated.Therefore, input structure may include by Button, keyboard, mouse, Trackpad etc..In addition, electronic console may include touching component, which is touched by test object Generation and/or the position of its screen are touched to promote user to input.

In concrete application scene, the sample input module fights generation module and obscures discrimination module and can be Software module, these software modules can be deployed on the container on the virtual machine or server on server or server.This Outside, these software modules can dispose according to actual needs on the same server or on different server, and the application does not limit It is fixed.

In another case, the article discrimination system can also be by a kind of application program being loaded on intelligent terminal (APP) Lai Shixian, the intelligent terminal obtain the multiple image of article to be identified by shooting, and are uploaded to via wireless network Cloud server, feedback identifying result after being identified by cloud.

The intelligent terminal be, for example, include but is not limited to smart phone, it is tablet computer, smartwatch, intelligent glasses, a Personal digital assistant (PDA) etc. portable or wearable electronic equipment, it should be understood that the application describes in embodiment Portable electronic device be an application example, the component of the equipment can have more or fewer components than diagram, Or with different component Configurations.Drawing the various assemblies illustrated can be realized with the combination of hardware, software or software and hardware, Including one or more signal processings and/or specific integrated circuit.

The intelligent terminal includes memory, Memory Controller, one or more processors (CPU), Peripheral Interface, RF Circuit, voicefrequency circuit, loudspeaker, microphone, input/output (I/O) subsystem, touch screen, other outputs or control equipment, with And outside port.These components are communicated by one or more communication bus or signal wire.

The intelligent terminal supports various application programs, such as one of the following terms or a variety of: drawing application program, Application program, word-processing application, website creation application program, disk editor application program, spreadsheet application journey is presented Sequence, game application, telephony application, videoconference application, email application, instant message application journey Sequence, body-building support application program, photo management application program, digital camera application program, digital video camera application program, Web page browsing application program, digital music player application program and/or digital video player application.

As shown in Figure 10, the training system 100 includes sample input module 101, fights generation module 102 and obscure and sentence Other module 103, in which: sample input module 101 is used to obtain the training sample image of several articles to be identified, the trained sample This image includes true images of items and false images of items.Fighting generation module 102 will be described true at network using a pair of of antibiosis Images of items, which fake, obtains fraud image；And network is generated using the confrontation and make very by the false images of items True image is made in acquisition.Obscure discrimination module 103 and obscures differentiation e-learning for the fraud image to be input to one to export The fraud image is genuine result；And by it is described make true image and be input to described obscure that differentiate e-learning described to export Making true image is false result.

The article to be identified includes the luxury goods packet with skin label or the article with written handwriting.For example, identifying certain The bag with skin label such as the both shoulders packet of brand, shoulder bag, handbag, oblique satchel, traveling bag；Such as bracelet, ring, wrist-watch, Pocket-watch, contract etc. have the article of written handwriting.It is temporarily luxury with the article to be identified in embodiment provided by the present application It is illustrated for product packet.

It is readily appreciated that, the training sample image is made of the multiple image obtained, includes true object in the multiple image Product image and false images of items, as training sample, the true images of items be article to be identified be accredited as in advance genuine piece and To the image of genuine piece shooting, the vacation images of items is that article to be identified is accredited as fakement in advance and clapped the fakement The image taken the photograph.In embodiment, the described prior identification can manually identify or technical appraisement by way of.

In an actual application scenarios, luxury goods are really wrapped and the overall appearance of false packet is quite similar, from overall appearance On carry out identify difficulty it is higher, therefore the point of the identification such as the font of its skin label, imprint process be the key that identify it is true and false.Cause This, in some embodiments, the training system of the article discrimination model further includes feature extraction unit, described for positioning Fixed position is reflected in true images of items and false images of items with extraction property information.In some embodiments, the feature Extraction unit positions the fixed position that reflects in the true images of items and false images of items using Fast Recursive convolutional neural networks, And convolutional network extraction property information is separated using depth.

It is to be understood that the mirror fixed position of article to be identified also can be different, such as luxury goods due to the difference of article For Ms packet, then its identification point generally comprises skin label, zipper and the label including article trade mark；For another example to wrist-watch and Speech, the then fixed position that reflects are generally placed upon watchband, dial plate etc.；The printed matters such as contract for another example, the then fixed position that reflects are generally placed upon book Write person's handwriting or seal etc..

In some embodiments, the confrontation generation module includes making true generator and fraud generator, described to make very Generator, which is used to carry out the true images of items to fake, obtains fraud image, and the fraud generator is used for the false article Image, which make really obtaining, makes true image.Wherein, the step of acquisition fraud image for using the fraud generator by institute It states true images of items and carries out acquisition fraud image of faking；The step of true image is made in the acquisition will to make true generator described in utilization The vacation images of items, which make really obtaining, makes true image.

Therefore, in some embodiments, it is described make true generator be also used to be made the true images of items very with Obtaining is genuine authentication image；It makes true image by described and is made very to obtain really to reset image；And it calculates described true Images of items and the genuine authentication image and/or the image similarity for really resetting image, so that the confrontation generates The loss function value of network is maintained in desired extent.

Similarly, in some embodiments, the fraud generator is also used to fake the false images of items With for；The fraud image is faked to obtain is false resetting image；And calculate the false images of items with The image similarity of the authentication image of the vacation and/or the false resetting image, so that the confrontation generates the loss of network Functional value is maintained in desired extent.

In some embodiments, described to obscure discrimination module include false article arbiter and true article arbiter, described It is genuine as a result, the true object that false article arbiter, which exports the fraud image for being learnt to the fraud image of input, Make true image of the product arbiter for that will input learns to make true image described in output to be false result.

In some embodiments, described to obscure discrimination module and obscure including the use of the structure of Siamese network described Discrimination module is trained and optimizes.For example, using one true one false image of two width comprising skin label as input, in addition to exporting two width Outside the qualification result of image, also output is with attribute prediction result or different attribute prediction result, for indicating the two images Whether same class (true or false) or a true vacation are belonged to.Thus, it is possible to further enhance the identification of article discrimination model Ability.

Described in brought forward, no matter how to fake to the true images of items, the corresponding article of fraud image obtained Remain as genuine piece.Therefore, obscure discrimination module using described and differentiate that the result of output remains as described make to fraud image Fault image is true.Similarly, no matter how the false images of items is carried out making very, what is obtained makes the corresponding article of true image Remain as fakement.Therefore, obscure discrimination module using described and differentiate that the result of output remains as described make to true image is made True image is false.

In some embodiments, the false article arbiter is also used to learn the true images of items of input Genuine result is predicted to export the true images of items；And the false images of items of input is learnt to export the vacation Images of items is the false result of prediction.

In the training process, it will be input to labeled as really fraud image in the false article arbiter, the vacation article Arbiter exports genuine result；It will be input in the false article arbiter labeled as genuine true images of items, the vacation article Genuine result is predicted in arbiter output；It will be input in the false article arbiter labeled as false false images of items, the vacation The false result of article arbiter output prediction.Obviously, the genuine result of prediction should be true really, the false result of the prediction It should really be false.

Similarly, in some embodiments, the true article arbiter is also used to carry out the true images of items of input Study predicts genuine result to export the true images of items；And the false images of items of input is learnt to export Stating false images of items is the false result of prediction.Its step is similar to the above, and principle and process please refer to above-described embodiment, herein It repeats no more.It is described obscure discrimination module be also used to for the true images of items being input to it is described obscure differentiate e-learning with defeated The true images of items is to predict genuine result out；Described differentiation network science is obscured alternatively, the false images of items is input to Practising to export the false images of items is the false result of prediction；One unknown true and false image is input to and described obscures differentiation network science Practise, with export the unknown true and false image and the true images of items or the false images of items be with attribute prediction result, or It is different attribute prediction result that person, which exports the unknown true and false image and the true images of items or the false images of items,；And Institute is obtained according to the same attribute prediction result or different attribute prediction result, and the genuine result of prediction or the result of prediction vacation State the qualification result of unknown true and false image.

For example, by the true images of items be input to it is described obscure differentiation study is carried out in discrimination module, due to described true Images of items is that genuine result is priori result (the i.e. described true images of items is to be previously determined to be genuine training sample image), because The result that discrimination module output is obscured described in this is that prediction is genuine as a result, indicating that the true images of items is true.Then, not by one Know that true and false image is input to described to obscure discrimination module and carry out differentiation study.Due to the unknown true and false image be only included as it is true or It is false two kinds as a result, prediction result is pre- with the true images of items when the unknown true and false image is true images of items It is identical to survey result attribute, that is, is true；When the unknown true and false image is false images of items, prediction result and the true object The prediction result attribute of product image is different, i.e., one true one is false.Therefore, described to obscure discrimination module unknown true and false image is sentenced Not Xue Xi after, export the unknown true and false image and the true images of items be to tie with attribute prediction result or different attribute prediction Fruit.Due to the true and false of the true images of items it is known that being pre- with attribute prediction result or different attribute according to output result therefore It surveys as a result, can determine that the unknown true and false image is true or is false.

Similarly, by the false images of items be input to it is described obscure differentiation study is carried out in discrimination module, it is described to obscure The result of discrimination module output is prediction vacation as a result, indicating that the false images of items is false.Then, by a unknown true and false image Discrimination module, which is obscured, described in being input to carries out differentiation study, it is similarly, pre- when the unknown true and false image is false images of items It is identical as the false prediction result attribute of images of items to survey result, i.e., is all false；When the unknown true and false image is true article When image, prediction result is different from the false prediction result attribute of images of items, i.e., one true one is false.Due to principle with it is above-mentioned Embodiment is similar, and detailed process and step please refer to above-described embodiment, and details are not described herein again.

In the embodiment of the present application, by the confrontation generation module and it is described obscure discrimination module alternately train, the confrontation In generation module make true generator and how the study of fraud generator is mutually converted true images of items and false images of items, it is raw Discrimination module is obscured at the image mixed the spurious with the genuine to confuse.And the true article arbiter obscured in discrimination module and false article Arbiter then constantly learn input image be very still be it is false, avoid being fought generation module fascination, to improve identification Ability.By the confrontation generation module and the training method for obscuring discrimination module loop iteration, enable to antibiosis At module and obscure discrimination module and constantly improve and optimize in the training process, to improve the identification energy of article discrimination model Power.

In some embodiments, the training system of the article discrimination model can also include optimization module, described excellent Changing module to utilize includes Adam (Adaptive Moment Estimation) optimization algorithm, SGD (Stochastic Gradient Descent) optimization algorithms such as optimization algorithm or RMSprop optimization algorithm optimize neural network, are based on Training data iteratively updates neural network weight, so that neural network cross entropy loss function minimizes, convergence is more stable, Neural network is avoided to fall into local optimum simultaneously, to enhance the robustness of article discrimination model.

It should be noted that Figure 10 is only a kind of possible implementation of the embodiment of the present application, and in practical application, article Authentication equipment can also include more or fewer components, here with no restriction.About being not shown in the embodiment of the present application or not The content of description, reference can be made to the related elaboration in previous embodiment, which is not described herein again.

The training system of article discrimination model provided by the present application fights the structure building article for generating network based on circulation Identify model, confrontation generation network is combined with differentiation network is obscured, generates making true generator and making for network using confrontation False generator make true or is faked to training sample image, will be made true or after faking image and is input to and obscures differentiation network, To be trained to the true article arbiter and false article arbiter of obscuring differentiation network, it can be improved and obscure differentiation network Distinguishing ability.Meanwhile by the way of generating network using confrontation and obscuring differentiation network alternating iteration training, confrontation can be made to generate Network and obscure the robustness and accuracy for both differentiating network training, common study mutually, while improving, and then effectively mentions The distinguishing ability and robustness of high article discrimination model.

Figure 11 is please referred to, the flow diagram of the application article discrimination method in one embodiment is shown as, as shown, The article discrimination method the following steps are included:

Step S21 obtains the image of the article to be identified of shooting；Described image includes an at least objective identification point.

The article to be identified includes the luxury goods packet with skin label or the article with written handwriting.For example, identifying certain The bag with skin label such as the both shoulders packet of brand, shoulder bag, handbag, oblique satchel, traveling bag；Such as bracelet, ring, wrist-watch, Pocket-watch, contract etc. have the article of written handwriting.It is temporarily luxury with the article to be identified in embodiment provided by the present application It is illustrated for product packet.It is to be understood that the objective identification point position of article to be identified also can be different due to the difference of article, Such as Ms's packet for luxury goods, then its objective identification point generally comprises skin label, zipper and including article trade mark Label；For another example for wrist-watch, then objective identification point position is generally placed upon watchband, dial plate etc..

In some embodiments, the image for obtaining the article to be identified of shooting includes but is not limited to: by using mobile phone, The camera of the intelligent terminals such as tablet computer, laptop shoots article to be identified, using the image of shooting as to Identify the image of article.The mode for positioning objective identification point position includes but is not limited to: 1) artificial calibration.For example, using Mobile phone shoots a width and includes the image of luxury goods packet (corresponding article to be identified), and draws a certain size range is taken to make on the image It include the skin label (corresponding objective identification point) of luxury goods packet for objective identification point position, in the range.For another example, more in acquisition After width training sample image, artificially all images are identified and demarcated one by one.2) fixed by recursive convolution neural network The position of position identification point.For example, there may be identifications for searching first in each image using Fast Recursive convolutional neural networks The candidate region of position is put to obtain the various dimensions feature of image, then returned using target detection differentiation, search box position etc. The accurate location of sub-network acquisition objective identification point position.

In some embodiments, the objective identification point is positioned to postpone, it can also be including the use of depth convolutional network Extraction property information.The depth convolutional network includes Xception (depth separates convolutional network), ResNet (depth residual error Network), MobileNet etc..For example, using depth setting in the image of the luxury goods packet of skin label position for width identification point Separable convolutional network is spent from characteristics letters such as skin label extracting section font, stroke, texture, relative position and sizes in image Breath.Depth separates convolutional network and carries out convolution respectively to each channel of image, and by the space of characteristic information in image The convolution of position and each interchannel is decoupling, to improve the ability of extraction property information, and is more advantageous to network It practises and optimizes.

Step S23, the article discrimination model obtained using the training method of the article discrimination model is to comprising described The image of objective identification point is identified.

Image comprising objective identification point is input in article discrimination model, the article discrimination model carries out image Identify.The article discrimination model is obtained by the training method training of above-mentioned article discrimination model.

Step S23, the output article to be identified is genuine result or the article to be identified is false result.

The article discrimination model identifies the image comprising objective identification point of input, if identifying described image It is that very, then it is genuine result that the article discrimination model, which exports the article to be identified,；If identifying described image is false, institute Stating article discrimination model and exporting the article to be identified is false result.

Article discrimination method provided by the present application, by a preparatory trained article discrimination model to shooting comprising to The image of identification article is identified, and is avoided the identification teacher's identification capacity difference of difference present in artificial identification method and is caused Identification result difference, the stability and accuracy of identification are high.Simultaneously, on the one hand, by using article provided by the present application Discrimination method, user, which only needs to shoot photo, can receive the identification result of return, significantly reduce what artificial identification was spent Economic cost and time cost really meet user and identify the needs of article is true and false immediately in daily life.On the other hand, originally The article discrimination method that application provides can effectively check the circulation of fakement, promote user's shopping experience.

Figure 12 is please referred to, the structural schematic diagram of the application article discrimination equipment in one embodiment is shown as, as shown, The article discrimination equipment 110 includes filming apparatus 111, memory 112 and one or more processors 113, filming apparatus 111, memory 112 can be connected with processor 113 by bus or other way, and the embodiment of the present application by bus to be connected For.Wherein, filming apparatus 111 is used to obtain the image of the article to be identified of shooting；Described image is reflected including an at least target Fixed point.

The filming apparatus 111 includes intelligent terminal and other devices or equipment with camera, for obtaining shooting The image of article to be identified；Described image includes an at least objective identification point.The intelligent terminal is, for example, to include but is not limited to Smart phone, tablet computer, smartwatch, intelligent glasses, personal digital assistant (PDA) etc. portable or wearable electricity Sub- equipment, it should be understood that the application portable electronic device described in embodiment is an application example, the equipment Component can have more or fewer components than diagram, or with different component Configurations.Various groups for drawing diagram Part can be realized with the combination of hardware, software or software and hardware, including one or more signal processings and/or dedicated integrated electricity Road.

The article to be identified includes the luxury goods packet with skin label or the article with written handwriting.For example, identifying certain The bag with skin label such as the both shoulders packet of brand, shoulder bag, handbag, oblique satchel, traveling bag；Such as bracelet, ring, wrist-watch, Pocket-watch, contract etc. have the article of written handwriting.Due to the difference of article, the mirror fixed position of article to be identified also can be different, Such as Ms's packet for luxury goods, then its identification point generally comprises skin label, zipper and the label including article trade mark； For another example for wrist-watch, then the fixed position that reflects is generally placed upon watchband, dial plate etc..In some embodiments, shooting is obtained The image of article to be identified includes but is not limited to: by using the camera shooting of the intelligent terminals such as mobile phone, tablet computer, laptop Head shoots article to be identified, using the image of shooting as the image of article to be identified.The format of described image is to calculate Machine stores the format of image, for example, bmp, jpg, png, tiff, gif, pcx, tga, exif, fpx, svg, psd, cdr, pcd, The format of the storages such as dxf, ufo, eps, ai, raw, WMF.

The memory may include random access memory (Random Access Memory, RAM), read-only memory (Read Only Memory, ROM), programmable read only memory (Programmable Read-Only Memory, PROM), Erasable programmable read only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable are compiled Journey read-only memory Electric Erasable Programmable Read-Only Memory, EEPROM) etc..Memory For storing program, processor executes the program after receiving and executing instruction.

The processor includes IC chip, has signal handling capacity；Or general processor, for example, it may be Digital signal processor (DSP), specific integrated circuit (ASIC), discrete gate or transistor logic, discrete hardware components, can To realize or execute disclosed each method, step and the logic diagram in the embodiment of the present application.The general processor can be with It is microprocessor or any conventional processors etc., such as central processing unit (Central Processing Unit, CPU).Institute Processor 113 is stated for calling the program code stored in the memory 112 to execute article discrimination method above-mentioned: being obtained The training sample image of several articles to be identified, the training sample image include true images of items and false images of items；Benefit The true images of items fake at network with a pair of of antibiosis and obtains fraud image；And network is generated using the confrontation The false images of items make really obtaining and makes true image；By the fraud image be input to one obscure differentiate e-learning with Exporting the fraud image is genuine result；And by it is described make true image be input to it is described obscure differentiate e-learning to export It is described to make the result that true image is vacation.

For executing the article discrimination method, principle and detailed process please refer to herein described article discrimination equipment Above-described embodiment, details are not described herein again.It should be noted that Figure 12 is only the possible realization side of one kind of the embodiment of the present application Formula, in practical application, article discrimination equipment can also include more or fewer components, here with no restriction.About the application The content for being not shown in embodiment or not describing, reference can be made to the related elaboration in previous embodiment, which is not described herein again.

Article discrimination equipment provided by the present application, user, which only needs to shoot photo, can receive the identification result of return, pole The earth reduces the artificial economic cost for identifying cost and time cost, really meets user and identifies object immediately in daily life The true and false demand of product, while also can effectively check the circulation of fakement, promote user's shopping experience.

Article discrimination method provided by the present application can also be executed by an article discrimination client.The article discrimination visitor Family end is loaded into a smart machine, is realized by the software and hardware in smart machine.The article discrimination client exists Loading form on smart machine includes APP application program or small routine (such as small routine in wechat etc.).

Figure 13 is please referred to, the structural schematic diagram of the application article discrimination client in one embodiment is shown as, as schemed institute Show, the article discrimination client 120 being loaded into a smart machine includes input module 121, processing module 122 and display Module 123.Wherein: when the input module 121 is used to receive the identification instruction of user's input, calling the smart machine Filming apparatus obtains the image of the article to be identified of shooting；Described image includes an at least objective identification point；The processing module 122 call the program code stored in the smart machine to execute article discrimination method above-mentioned；The display module 123 is aobvious Show that the output article to be identified is genuine result or the article to be identified is false result.

The smart machine be, for example, include but is not limited to smart phone, it is tablet computer, smartwatch, intelligent glasses, a Personal digital assistant (PDA) etc. portable or wearable electronic equipment, it should be understood that the application describes in embodiment Portable or wearable electronic equipment be an application example, the component of the equipment can have than diagram it is more or Less component, or with different component Configurations.The various assemblies for drawing diagram can use hardware, software or software and hardware Combination is to realize, including one or more signal processings and/or specific integrated circuit.

The article includes the luxury goods packet with skin label or the article with written handwriting, and the objective identification point is institute State the position of skin label or written handwriting.For example, identifying the both shoulders packet of certain brand, shoulder bag, handbag, oblique satchel, traveling bag etc. Bag with skin label；Such as bracelet, ring, wrist-watch, pocket-watch, contract etc. have the article of written handwriting.Not due to article Together, the objective identification point position of article also can be different, such as Ms's packet for luxury goods, then its objective identification point is general Including skin label, zipper and label including article trade mark；For another example for wrist-watch, then objective identification point position is generally placed upon Watchband, dial plate etc.；The printed matters such as contract for another example, then objective identification point position is generally placed upon written handwriting or seal etc..? In embodiment provided by the present application, temporarily it is illustrated so that the article to be identified is luxury goods packet as an example.

The article discrimination client 120 obtains the identification instruction of user's input, the identification by its input module 121 The filming apparatus that instruction is used to indicate smart machine described in the article discrimination client call obtains the article to be identified of shooting Image.Wherein, the mode for obtaining the identification instruction includes but is not limited to: the text data by monitoring the smart machine Input, touch data input, voice data input etc., to obtain the identification instruction.

For example, user inputs specific character or character string, the character on the display interface of the smart machine Perhaps character string is associated with identification instruction when the input module monitors the character or character string inputs in advance, defeated Enter module and get the character or character string, is instructed to obtain corresponding identification according to preset incidence relation. For another example, user clicks on the display interface of the smart machine, slides, dragging the touch-controls such as virtual slider, gesture operation When operation, the touch control operation is associated with identification instruction in advance；The input module monitors the touch control operation, thus according to Preset incidence relation obtains corresponding identification instruction.For another example, the voice data of the input module monitoring users, according to The incidence relation of the preset voice data and identification instruction obtains corresponding identification instruction.Obtaining identification instruction Afterwards, the image for the article to be identified that the input module calls user to shoot by the filming apparatus of smart machine, and sent out Give processing module.

In some embodiments, the input module further includes a selecting unit, for receiving the mirror of user's input When instructing surely, the image of the image library of smart machine article to be identified for selection by the user is called；Described image includes extremely A few objective identification point.After the input module obtains identification instruction, the selecting unit calls described image library, by image library In all or part of image to user show, so that user selects.The image that the input module selects user is made For the image of article to be identified, and send it to processing module.Described image library is stored in a storage medium, the storage Medium can be the storage device of a server-side, be also possible to the storage device of the smart machine, or such as SD card, Other storage mediums such as flash.

The objective identification point can be before input module calls image, first pass through handmarking or technical mark in advance Etc. modes determine, be also possible to the input module call image after carry out the modes such as handmarking or technical mark determine 's.The principle for carrying out the modes such as handmarking or technical mark in advance before input module calls image please refers to Fig. 1 and corresponds to Embodiment, details are not described herein again.In some embodiments, the input module can also be in smart machine after calling image Display interface on prompt user delimit one or more candidate regions on the image, include the target in the candidate region Point is identified, to obtain the image of the article to be identified comprising mirror fixed position.

The processing module 122 is connected with the input module 121, for receiving the transmission of input module 121 Article to be identified image after, call the program code stored in the smart machine to execute article discrimination side above-mentioned Method.The described method includes: obtaining the image of the article to be identified of shooting；Described image includes an at least objective identification point；It utilizes The article discrimination model that the training method of the article discrimination model obtains reflects to the image comprising the objective identification point Not；Export that the article to be identified is genuine result or the article to be identified is false result.

In some embodiments, user can also directly using the smart machine filming apparatus directly to it is described to Identification article mirror fixed position (such as skin label, brand trademark, hardware etc.) shot, obtain respectively such as skin label image, The images such as trademark image, hardware image execute article discrimination method for the processing unit.

The display module 123 is connected with the processing module 122, for receive the processing module send described in Identify that article is genuine result or the article to be identified is false as a result, and it will be shown in displays circle of the smart machine On face.

It in some embodiments, can also be the display module with percentage to the identification result of the article to be identified The form of ratio shows the image similarity or images match degree of article and genuine piece to be identified.For example, the processing module is to image Identification result is sent to display module after being identified, the display module shows image on the display interface of smart machine Matching degree is 95%, then it represents that it may be genuine piece that the article to be identified is very big；Alternatively, the display module is in smart machine Show that images match degree is 23% on display interface, then it represents that it may be fakement that the article to be identified is very big.It is to be understood that institute It states 95% and 23% specific value to be illustrative only, and is not intended to limit output described image similarity or images match degree The range of numerical value.

In conclusion the article discrimination client obtains what user carried out on the smart machine by input module Input operation obtains the input and operates corresponding mirror according to the incidence relation of preset input operation and identification instruction Fixed instruction, then identification instruction is sent to processing module by the input module, and the processing module passes through based on described Trained article identification model executes article discrimination method to the training method of article identification model in advance, thus to described wait reflect The image of earnest product is identified, and identifies the true and false of described image, and then is identified the article to be identified and be true or be false.? In actual application scenarios, user can open an APP on mobile phone (corresponding smart machine) or open the little Cheng of wechat Sequence, to start article discrimination client to identify to article.

Article discrimination client provided by the present application, by being loaded into the article discrimination client of a smart machine, user It only needs to shoot photo by the filming apparatus of a smart machine, and is identified by the article discrimination client, immediately The identification result for receiving return significantly reduces the artificial economic cost for identifying cost and time cost, really meets user Identify the true and false demand of article immediately in daily life, user experience is good.Meanwhile article discrimination client provided by the present application The image comprising article to be identified of shooting is identified by a preparatory trained article discrimination model, is avoided artificial Difference present in identification method identify teacher's identification capacity it is different and caused by identification result difference, the stability and standard of identification True property is high.

The application also provide it is a kind of it is computer-readable write storage medium, store the computer journey of active route selection method Sequence, the computer program of the source route selection method, which is performed, realizes above-described embodiment object described in Fig. 1 to Fig. 9 Judge the training method of other model.

The application also provide it is a kind of it is computer-readable write storage medium, store the computer journey of active route selection method Sequence, the computer program of the source route selection method, which is performed, realizes that above-described embodiment reflects about article described in Figure 11 Other method.

It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product It is stored in a computer readable storage medium.Based on this understanding, the technical solution of the application is substantially in other words The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a People's computer, server or network equipment etc.) execute each embodiment the method for the application all or part of the steps.

In embodiment provided by the present application, it is described it is computer-readable write storage medium may include read-only memory, with Machine accesses memory, EEPROM, CD-ROM or other optical disk storage apparatus, disk storage device or other magnetic storage apparatus, sudden strain of a muscle Deposit, USB flash disk, mobile hard disk or can be used in storage have instruction or data structure form desired program code and can Any other medium accessed by computer.In addition, any connection can be properly termed as computer-readable medium.Example Such as, if instruction is using coaxial cable, optical fiber cable, twisted pair, digital subscriber line (DSL) or such as infrared ray, wirelessly The wireless technology of electricity and microwave etc is sent, then the coaxial cable, optical fiber light from website, server or other remote sources The wireless technology of cable, twisted pair, DSL or such as infrared ray, radio and microwave etc includes in the definition of the medium. It is to be understood, however, that it is computer-readable write storage medium and data storage medium do not include connection, carrier wave, signal or Other fugitive mediums, and be intended to be directed to non-transitory, tangible storage medium.Disk and light as used in application Disk includes compact disk (CD), laser-optical disk, CD, digital versatile disc (DVD), floppy disk and Blu-ray Disc, wherein disk Usual magnetically replicate data, and CD is then with laser come optically replicate data.

In one or more illustrative aspects, the training method and article discrimination method of herein described article discrimination model Computer program described function can be realized with the mode of hardware, software, firmware or any combination thereof.When with soft When part is realized, these functions can be stored as one or more instruction or code or be transmitted on computer-readable medium. The step of method disclosed in the present application or algorithm, software module can be performed with processor to embody, and wherein processor can be held Row software module can be located at that tangible, non-transitory is computer-readable writes on storage medium.It is tangible, non-transitory is computer-readable Writing storage medium can be any usable medium that computer can access.

Flow chart and block diagram in the above-mentioned attached drawing of the application illustrate the system according to the various embodiments of the application, side The architecture, function and operation in the cards of method and computer program product.In this regard, every in flowchart or block diagram A box can represent a part of a module, program segment or code, and a part of the module, program segment or code includes One or more executable instructions for implementing the specified logical function.It should also be noted that in some realizations as replacement In, function marked in the box can also occur in a different order than that indicated in the drawings.For example, two succeedingly indicate Box can actually be basically executed in parallel, they can also be executed in the opposite order sometimes, this is according to involved in Depending on function.It is also noted that each box in block diagram and or flow chart and the box in block diagram and or flow chart Combination, can be realized by the dedicated hardware based system of functions or operations as defined in executing, or can pass through The combination of specialized hardware and computer instruction is realized.

The training method of article discrimination model provided by the present application, the training system of article discrimination model, article discrimination side Method, article discrimination equipment, article discrimination client and computer readable storage medium are sentenced by that will fight to generate network and obscure Other network combines, using confrontation generate network make true generator and fraud generator to training sample image make it is true or It fakes, true or after faking image will be made and be input to and obscure differentiation network, thus to the true article arbiter for differentiating network is obscured It is trained with false article arbiter, can be improved the distinguishing ability obscured and differentiate network.Meanwhile using confrontation generate network and Obscure the mode for differentiating that network alternating iteration is trained, confrontation can be made to generate network and obscure and differentiate that network trains mutually, is common Study, while the robustness and accuracy of the two are improved, and then effectively improve the distinguishing ability and robust of article discrimination model Property.Article is carried out using the article discrimination model that the training method training by article discrimination model provided by the present application obtains Identify, accuracy rate is high, and stability is strong.

The principles and effects of the application are only illustrated in above-described embodiment, not for limitation the application.It is any ripe Know the personage of this technology all can without prejudice to spirit herein and under the scope of, carry out modifications and changes to above-described embodiment.Cause This, those of ordinary skill in the art is complete without departing from spirit disclosed herein and institute under technical idea such as At all equivalent modifications or change, should be covered by claims hereof.

Claims

1. a kind of training method of article discrimination model, which comprises the following steps:

The training sample image of several articles to be identified is obtained, the training sample image includes true images of items and false article Image；

The true images of items fake at network using a pair of of antibiosis and obtains fraud image；And it utilizes described to antibiosis The false images of items make really obtaining at network and makes true image；

The fraud image is input to one and obscures differentiation e-learning to export the fraud image as genuine result；And it will It is described make true image be input to it is described obscure differentiate e-learning described to make true image be false result to export.

2. the training method of article discrimination model according to claim 1, which is characterized in that further include the positioning true object The step of reflecting fixed position in product image and false images of items with extraction property information.

3. the training method of article discrimination model according to claim 2, which is characterized in that the positioning true article The step of reflecting fixed position in image and false images of items with extraction property information are as follows: utilize Fast Recursive convolutional neural networks The fixed position that reflects in the true images of items and false images of items is positioned, and separates convolutional network extraction property using depth Information.

4. the training method of article discrimination model according to claim 1, which is characterized in that the confrontation generates network packet Include and make true generator and fraud generator, in which: the step of the acquisition fraud image for using the fraud generator by institute It states true images of items and carries out acquisition fraud image of faking；The step of true image is made in the acquisition will to make true generator described in utilization The vacation images of items, which make really obtaining, makes true image.

5. the training method of article discrimination model according to claim 4, which is characterized in that the fraud generator is made True generator includes depth residual error network.

6. the training method of article discrimination model according to claim 4, which is characterized in that further include to described to antibiosis The step of being trained at network:

It makes true generator using described and is made the true images of items very to obtain as genuine authentication image；

It makes true generator using described and is made the fraud image very to obtain really to reset image；

The true images of items and the genuine authentication image and/or the image similarity for really resetting image are calculated, with The loss function value for making the confrontation generate network is maintained in desired extent.

7. the training method of article discrimination model according to claim 4, which is characterized in that further include to described to antibiosis The step of being trained at network:

It is false authentication image that the false images of items, which is faked to obtain, using the fraud generator；

Using the fraud generator, to make true image and fake to obtain be false resetting image by described；

The authentication image of the false images of items and the vacation and/or the image similarity of the false resetting image are calculated, with The loss function value for making the confrontation generate network is maintained in desired extent.

8. the training method of article discrimination model according to claim 1, which is characterized in that described to obscure differentiation network packet Include false article arbiter and true article arbiter, wherein by the fraud image be input to the false article arbiter study with Exporting the fraud image is genuine result；And it makes true image by described and is input to the true article arbiter study to export It is described to make the result that true image is vacation.

9. the training method of article discrimination model according to claim 8, which is characterized in that further include to it is described obscure sentence The step of other network is trained:

The fraud image is input to the false article arbiter study to export the fraud image as genuine result；

It is genuine for prediction to export the true images of items that the true images of items is input to the false article arbiter study As a result；

It is that prediction is false that the false images of items, which is input to the false article arbiter study to export the false images of items, As a result.

10. the training method of article discrimination model according to claim 8, which is characterized in that further include obscuring described Differentiate the step of network is trained:

It makes true image by described and is input to the true article arbiter study described to make true image be false result to export；

It is genuine for prediction to export the true images of items that the true images of items is input to the true article arbiter study As a result；

It is that prediction is false that the false images of items, which is input to the true article arbiter study to export the false images of items, As a result.

11. the training method of article discrimination model according to claim 8, which is characterized in that further include obscuring described Differentiate the step of network is trained:

The true images of items is input to and described obscures that differentiate e-learning genuine for prediction to export the true images of items As a result；Described obscure that differentiate e-learning to export the false images of items be pre- alternatively, the false images of items is input to Survey false result；

By a unknown true and false image be input to it is described obscure differentiation e-learning, to export the unknown true and false image and described true Images of items or the false images of items are with attribute prediction result, or the output unknown true and false image and the true article Image or the false images of items are different attribute prediction result；

It is obtained according to the same attribute prediction result or different attribute prediction result, and the genuine result of prediction or the result of prediction vacation Obtain the qualification result of the unknown true and false image.

12. the training method of article discrimination model according to claim 8, which is characterized in that described to obscure differentiation network Including Siamese network.

13. the training method of article discrimination model according to claim 1, which is characterized in that the article includes having The luxury goods packet of skin label or article with written handwriting.

14. a kind of training system of article discrimination model characterized by comprising

Sample input module, for obtaining the training sample image of several articles to be identified, the training sample image includes true Images of items and false images of items；

Generation module is fought, fake by the true images of items at network using a pair of of antibiosis obtains fraud image；And The false images of items make really obtaining using confrontation generation network and makes true image；

Obscure discrimination module, obscures differentiation e-learning for the fraud image to be input to one to export the fraud image For genuine result；And by it is described make true image be input to it is described obscure differentiate e-learning described to make true image be false to export Result.

15. the training system of article discrimination model according to claim 14, which is characterized in that further include feature extraction list Member reflects fixed position for positioning in the true images of items and false images of items with extraction property information.

16. the training system of article discrimination model according to claim 15, which is characterized in that the feature extraction unit The fixed position that reflects in the true images of items and false images of items is positioned using Fast Recursive convolutional neural networks, and using deeply Spend separable convolutional network extraction property information.

17. the training system of article discrimination model according to claim 14, which is characterized in that the confrontation generation module Include:

True generator is made, obtains fraud image for fake the true images of items；

Fraud generator makes true image for make really obtaining the false images of items.

18. the training system of article discrimination model according to claim 17, which is characterized in that the fraud generator or Making true generator includes depth residual error network.

19. the training system of article discrimination model according to claim 17, which is characterized in that described to make true generator also For the true images of items to be made very to obtain as genuine authentication image；By it is described make true image made very with obtain Really to reset image；And it calculates the true images of items and the genuine authentication image and/or described really resets image Image similarity so that it is described confrontation generate network loss function value be maintained in desired extent.

20. the training system of article discrimination model according to claim 17, which is characterized in that the fraud generator is also For by the false images of items fake with for；The fraud image is faked to obtain is false resetting figure Picture；And the authentication image of the false images of items and the vacation and/or the image similarity of the false resetting image are calculated, So that the loss function value that the confrontation generates network is maintained in desired extent.

21. the training system of article discrimination model according to claim 14, which is characterized in that described to obscure discrimination module Include:

False article arbiter, for being learnt to the fraud image of input to export the fraud image as genuine result；

True article arbiter, the true image of making for that will input learn to make true image described in output to be false result.

22. the training system of article discrimination model according to claim 21, which is characterized in that the vacation article arbiter It is also used to learn the true images of items of input to predict genuine result to export the true images of items；And it is right The false images of items of input is learnt to export the false images of items to be the false result of prediction.

23. the training system of article discrimination model according to claim 21, which is characterized in that the true article arbiter It is also used to learn the true images of items of input to predict genuine result to export the true images of items；And to input False images of items learnt to export the result that the false images of items be prediction vacation.

24. the training system of article discrimination model according to claim 14, which is characterized in that described to obscure discrimination module Obscuring described in being also used to for the true images of items being input to and differentiating e-learning to export the true images of items is that prediction is true Result；Alternatively, by the false images of items be input to it is described obscure differentiate e-learning to export the false images of items and be Predict false result；By a unknown true and false image be input to it is described obscure differentiation e-learning, to export the unknown true and false figure As with the true images of items or the false images of items be with attribute prediction result, or the output unknown true and false image with The true images of items or the false images of items are different attribute prediction result；And according to the same attribute prediction result or Different attribute prediction result, and the genuine result of prediction or the result of prediction vacation obtain the identification knot of the unknown true and false image Fruit.

25. the training system of article discrimination model according to claim 24, which is characterized in that described to obscure discrimination module Including Siamese network.

26. the training system of article discrimination model according to claim 14, which is characterized in that the article includes having The luxury goods packet of skin label or article with written handwriting.

27. a kind of article discrimination method, which comprises the following steps:

Obtain the image of the article to be identified of shooting；Described image includes an at least objective identification point；

The article discrimination model obtained using the training method of the described in any item article discrimination models of the claims 1-13 Image comprising the objective identification point is identified；

Export that the article to be identified is genuine result or the article to be identified is false result.

28. article discrimination method according to claim 27, which is characterized in that the article includes the luxury with skin label Product packet or article with written handwriting, the objective identification point are the position of the skin label or written handwriting.

29. a kind of article discrimination equipment characterized by comprising

Filming apparatus, the image of the article to be identified for obtaining shooting；Described image includes an at least objective identification point；

Memory, for storing program code；

One or more processors require in 27-28 for calling the program code stored in the memory to carry out perform claim Described in any item article discrimination methods.

30. article discrimination equipment according to claim 29, which is characterized in that the article includes the luxury with skin label Product packet or article with written handwriting, the objective identification point are the position of the skin label or written handwriting.

31. a kind of article discrimination client, is loaded into a smart machine characterized by comprising

Input module when for receiving the identification instruction of user's input, calling the filming apparatus of the smart machine to obtain and clapping The image for the article to be identified taken the photograph；Described image includes an at least objective identification point；

Processing module is called the program code stored in the smart machine to carry out perform claim and is required described in any one of 27-28 Article discrimination method；

Display module, the display output article to be identified is genuine result or the article to be identified is false result.

32. article discrimination client according to claim 31, which is characterized in that the input module further includes a selection Unit calls the image library of the smart machine for selection by the user wait reflect when for receiving the identification instruction of user's input The image of earnest product；Described image includes an at least objective identification point.

33. article discrimination client according to claim 31, which is characterized in that the article includes have skin label luxurious Wasteful product packet or the article with written handwriting, the objective identification point are the position of the skin label or written handwriting.

34. a kind of computer readable storage medium is stored with the computer program for article discrimination, which is characterized in that described Computer program is performed the training method for realizing the described in any item article discrimination models of claim 1-13.

35. a kind of computer readable storage medium is stored with the computer program for article discrimination, which is characterized in that described Computer program, which is performed, realizes the described in any item article discrimination methods of claim 27-28.