CN110321447A

CN110321447A - Determination method, apparatus, electronic equipment and the storage medium of multiimage

Info

Publication number: CN110321447A
Application number: CN201910612100.2A
Authority: CN
Inventors: 王鑫宇
Original assignee: Beijing ByteDance Network Technology Co Ltd
Current assignee: Beijing ByteDance Network Technology Co Ltd
Priority date: 2019-07-08
Filing date: 2019-07-08
Publication date: 2019-10-11

Abstract

The embodiment of the present disclosure provides determination method, apparatus, electronic equipment and the storage medium of a kind of multiimage, comprising: obtains the first image to be processed and the second image to be processed；Determine the first Hash codes of the first image to be processed and the second Hash codes of the second image to be processed；Based on the first Hash codes and the second Hash codes, determine whether the first image to be processed and the second image to be processed repeat.In the embodiments of the present disclosure, due to the Hash codes got be based on training after neural network model the Hash codes of different images can be made to be different, the Hash codes of similar image are identical, and then when carrying out duplicate removal to image, multiimage can be determined according to the Hash codes of image to be processed, to solve the problems, such as to be supplied in the image of user there may be the duplicate image of content, both can quickly remove multiimage, the accuracy rate of duplicate removal can be improved again, and then promoted user experience.

Description

Determination method, apparatus, electronic equipment and the storage medium of multiimage

Technical field

This disclosure relates to technical field of image processing, specifically, this disclosure relates to a kind of determination method of multiimage, Device, electronic equipment and storage medium.

Background technique

In the prior art, for original image, user may will do it some changes to image, for example, image is carried out Image on the image plus several texts, is zoomed in or out etc. simple map function by rotation, at this time transformed image and former Beginning image is essentially identical in terms of content.But when user is to want to check some images, in prior art be by Transformed image and original image are both provided to user, that is, there may be the duplicate images of content in the image provided, at this time The user experience of user may be reduced.Therefore, a kind of scheme for eliminating multiimage is needed at present, has solved the prior art In the problem of being supplied to image duplicate there may be content in the image of user, and then promote user experience.

Summary of the invention

The purpose of the disclosure is intended at least can solve above-mentioned one of technological deficiency, promotes the usage experience of user.This public affairs Open the technical solution adopted is as follows::

In a first aspect, the embodiment of the present disclosure provides a kind of training method of neural network model, this method comprises:

Training sample data are obtained, training sample data include the first image set and the second image set, wherein the first image Concentrating includes different original image, and the second image set includes the changing image of each original image；

Neural network model is trained based on training sample data, until meeting preset trained termination condition；Its In, the input of neural network model is image, exports the Hash codes for image, and training termination condition includes:

First similarity and the second similarity meet preset condition, wherein the first similarity be different original images it Between similarity, similarity of second similarity between each original image and the changing image of the original image, the first phase Like the Hash codes determination that degree and the second similarity are based on image.

In first aspect optional embodiment, which includes:

First prediction accuracy is not less than first threshold, and the second prediction accuracy is not less than second threshold, wherein first Less than the first similarity of the first setting value in calculated all first similarities when prediction accuracy is each training The accounting of quantity is greater than the second setting value in calculated all second similarities when the second prediction similarity is each training The second similarity quantity accounting；

Alternatively,

Third prediction accuracy is greater than third threshold value, wherein third prediction accuracy all similarities when being training every time The accounting of the quantity of the similarity pair met the requirements in, each similarity is to similar for corresponding to an original image first Minimum value in second similarity corresponding to maximum value and an original image in degree, it is desirable that be less than minimum for maximum value Value.

Second aspect, the embodiment of the present disclosure provide a kind of determination method of multiimage, this method comprises:

Obtain the first image to be processed and the second image to be processed；

Determine the first Hash codes of the first image to be processed and the second Hash codes of the second image to be processed, wherein first Hash codes and the second Hash codes are determined by neural network model, and neural network model is by the method in first aspect What training obtained；

Based on the first Hash codes and the second Hash codes, determine whether the first image to be processed and the second image to be processed weigh It is multiple.

In second aspect optional embodiment, the first image to be processed and the second image to be processed are obtained, comprising:

Obtain picture search keyword；

Initial image search result, the first image to be processed and the second figure to be processed are obtained based on picture search keyword As being the image in initial image search result；

Determine whether the first image to be processed and the second image to be processed repeat, comprising:

Determine whether the image in initial image search result repeats；

This method further include:

Based on the multiimage determined in initial image search result, the image search result after duplicate removal is supplied to User.

In second aspect optional embodiment, determine whether the image in initial image search result repeats, comprising:

According to the image sequence in initial image search result, by first image in initial image search result As first image of the first image group, since second image, following operation is executed:

According to the Hash codes of image, the similarity of first image in every image and each image group, each image group are calculated Including the first image group；

If there is the similarity greater than the 4th threshold value in calculated similarity, add the image to greater than the 4th threshold value Similarity corresponding in image group；

If calculated similarity creates an image group no more than the 4th threshold value, using image as newly-built figure As first image in group；

Wherein, the image in image group including at least two images is multiimage.

In second aspect optional embodiment, the image search result after duplicate removal is supplied to user, comprising:

First image in each image group is supplied to user as the image search result after duplicate removal.

In second aspect optional embodiment, using first image in each image group as the picture search knot after duplicate removal Fruit is supplied to after user, further includes:

Receive continue to check request when, by next figure of the image for being supplied to user last in every group of image group As being supplied to user.

In second aspect optional embodiment, the executing subject of method is server or terminal device, if the execution of method Main body is server, and the image search result after duplicate removal is supplied to user, comprising:

Image search result after duplicate removal is sent to terminal device, so that equipment is by the picture search after duplicate removal in terminal User is given as the result is shown.

In second aspect optional embodiment, the first image to be processed and the second image to be processed are in video to be processed Video frame images, determine whether the first image to be processed and the second image to be processed repeat to refer to the view determined in video to be processed Whether frequency frame image repeats；

Based on the first Hash codes and the second Hash codes, determine whether the first image to be processed and the second image to be processed weigh It is multiple, comprising:

For the current video frame image in video to be processed, according to the Hash codes and current video of current video frame image The Hash codes of the associated video frame image of frame image determine that current video frame image is similar to each associated video frame image Degree；

Similarity based on current video frame image Yu each associated video frame image determines that current video frame image is opposite In the similarity of associated video frame image；

Wherein, the continuous video frame images before associated video frame image includes current video frame image, and/or, when Continuous video frame images after preceding video frame images；

If current video frame image is greater than the 5th threshold value, current video frame relative to the similarity of associated video frame image Image is the multiimage relative to associated video frame image.

In second aspect optional embodiment, if associated video frame image is at least two field pictures, it is based on current video frame The similarity of image and each associated video frame image determines current video frame image relative to the similar of associated video frame image Degree, comprising:

Similarity and preset each associated video based on current video frame image and each associated video frame image The weight of frame image determines similarity of the current video frame image relative to associated video frame image.

In second aspect optional embodiment, the weight of each associated video frame image is based on each associated video frame figure As what is determined with the positional relationship of current video frame image.

In second aspect optional embodiment, positional relationship includes adjacent position relationship, and adjacent position relationship is direct phase Adjacent or indirect neighbor, wherein be greater than with the weight of the associated video frame image of current video frame image direct neighbor and work as forward sight The weight of the associated video frame image of Pin Zheng image indirect neighbor；

When in adjacent position, relationship is identical, the weight of the associated video frame image before current video frame image is greater than The weight of associated video frame image after current video frame image.

The third aspect, the embodiment of the present disclosure provide a kind of training device of neural network model, which includes:

Data acquisition module, for obtaining training sample data, training sample data include the first image set and the second figure Image set, wherein include different original image in the first image set, the second image set includes the changing image of original image；

Model training module, for being trained based on training sample data to neural network model, until meeting default Training termination condition；Wherein, the input of neural network model is image, exports the Hash codes for image, training termination condition Include:

First similarity and the second similarity meet preset condition, wherein the first similarity be different original images it Between similarity, similarity of second similarity between each original image and the changing image of the original image；Wherein, One similarity and the second similarity are the Hash codes determinations based on image.

In the optional embodiment of third face, which includes:

Alternatively,

Fourth aspect, the embodiment of the present disclosure provide a kind of determining device of multiimage, which includes:

Image collection module, for obtaining the first image to be processed and the second image to be processed；

Hash codes determining module, for determine the first image to be processed the first Hash codes and the second image to be processed Two Hash codes, wherein the first Hash codes and the second Hash codes are determined by neural network model, and neural network model is logical What the method training crossed in first aspect obtained；

Multiimage determining module, for be based on the first Hash codes and the second Hash codes, determine the first image to be processed and Whether the second image to be processed repeats.

During fourth aspect is optionally implemented, which further includes retrieval module, is specifically used for:

Picture search keyword is obtained, initial image search result is obtained based on picture search keyword, first wait locate It manages image and the second image to be processed is the image in initial image search result；

Multiimage determining module is specific to use when determining whether the first image to be processed and the second image to be processed repeat In:

Determine whether the image in initial image search result repeats；

Multiimage determining module is also used to:

During fourth aspect is optionally implemented, image of the multiimage determining module in the initial image search result of determination When whether repeating, it is specifically used for:

According to the image sequence in initial image search result, by first image in initial image search result As first image of the first image group, since second image, following operation is executed, according to the Hash codes of image, is calculated The similarity of first image in every image and each image group, each image group includes the first image group, if calculated similar There is the similarity greater than the 4th threshold value in degree, then adds the image to image group corresponding to the similarity greater than the 4th threshold value In, if calculated similarity creates an image group no more than the 4th threshold value, using image as in newly-built image group First image；

Wherein, the image in image group including at least two images is multiimage.

During fourth aspect is optionally implemented, the image search result after duplicate removal is being supplied to use by multiimage determining module When family, it is specifically used for:

During fourth aspect is optionally implemented, multiimage determining module is also used to:

After first image in each image group is supplied to user as the image search result after duplicate removal, if connecing It receives and continues to check request, next image of the image for being supplied to user last in every group of image group is supplied to user.

During fourth aspect is optionally implemented, the executing subject of method is server or terminal device, if the execution master of method Body is server, and multiimage determining module is specifically used for when the image search result after duplicate removal is supplied to user:

During fourth aspect is optionally implemented, the first image to be processed and the second image to be processed are the view in video to be processed Frequency frame image, determines whether the first image to be processed and the second image to be processed repeat to refer to the video determined in video to be processed Whether frame image repeats；

Multiimage determining module is being based on the first Hash codes and the second Hash codes, determines the first image to be processed and second When whether image to be processed repeats, it is specifically used for:

For the current video frame image in video to be processed, according to the Hash codes and current video of current video frame image The Hash codes of the associated video frame image of frame image determine that current video frame image is similar to each associated video frame image Degree, the similarity based on current video frame image Yu each associated video frame image determine current video frame image relative to pass Join the similarity of video frame images；

During fourth aspect is optionally implemented, if associated video frame image is at least two field pictures, multiimage determining module In the similarity based on current video frame image Yu each associated video frame image, determine current video frame image relative to association When the similarity of video frame images, it is specifically used for:

During fourth aspect is optionally implemented, the weight of each associated video frame image is based on each associated video frame image It is determined with the positional relationship of current video frame image.

During fourth aspect is optionally implemented, positional relationship includes adjacent position relationship, and adjacent position relationship is direct neighbor Or indirect neighbor, wherein be greater than with the weight of the associated video frame image of current video frame image direct neighbor and current video The weight of the associated video frame image of frame image indirect neighbor；

5th aspect, present disclose provides a kind of electronic equipment, which includes processor and memory；

Memory, for storing computer operation instruction；

Processor, for executing the first aspect and second such as the embodiment of the present disclosure by calling computer operation instruction Method shown in any embodiment of aspect.

6th aspect, present disclose provides a kind of computer readable storage medium, the computer-readable recording medium storages Have at least one instruction, at least a Duan Chengxu, code set or instruction set, at least one instruction, an at least Duan Chengxu, code set or Instruction set is loaded by processor and is executed to realize such as the first aspect of the embodiment of the present disclosure and any embodiment of second aspect Shown in method.

The technical solution that the embodiment of the present disclosure provides has the benefit that

In the embodiments of the present disclosure, due to the Hash codes got be based on training after neural network model obtain, And the neural network model can be such that the Hash codes of different images are different, the Hash codes of similar image be it is identical, into And when carrying out duplicate removal to image, it can determine multiimage, according to the Hash codes of image to be processed to solve the prior art In the problem of being supplied to image duplicate there may be content in the image of user, not only can quickly remove multiimage, but also can To improve the accuracy rate of duplicate removal, and then promote user experience.

Detailed description of the invention

In order to illustrate more clearly of technical solution in embodiment of the disclosure, the embodiment of the present disclosure will be described below Needed in attached drawing be briefly described.

Fig. 1 is a kind of flow diagram of the training method of neural network model in embodiment of the disclosure；

Fig. 2 is the schematic diagram of original image and changing image in embodiment of the disclosure；

Fig. 3 is the schematic diagram of neural network model training principle in embodiment of the disclosure；

Fig. 4 is a kind of schematic diagram of the process of the determination method of multiimage in embodiment of the disclosure；

Fig. 5 is a kind of schematic diagram that duplicate removal is carried out to initial image search result in embodiment of the disclosure；

Fig. 6 A is a kind of schematic diagram of the video frame images in embodiment of the disclosure in video；

Fig. 6 B is that the video frame images in embodiment of the disclosure in a kind of pair of video carry out the schematic diagram after duplicate removal；

Fig. 7 is a kind of schematic diagram of video frame images in embodiment of the disclosure；

Fig. 8 is a kind of schematic diagram of the training device of neural network model in embodiment of the disclosure；

Fig. 9 is a kind of schematic diagram of the determining device of multiimage in embodiment of the disclosure；

Figure 10 is the structural schematic diagram of a kind of electronic equipment in embodiment of the disclosure.

Specific embodiment

Embodiment of the disclosure is described below in detail, the example of the embodiment is shown in the accompanying drawings, wherein phase from beginning to end Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached drawing The embodiment of description is exemplary, and is only used for explaining the Sense of Technology of the disclosure, and cannot be construed to the limitation to the disclosure.

Those skilled in the art of the present technique are appreciated that unless expressly stated, singular " one " used herein, "one" It may also comprise plural form with "the".It is to be further understood that wording " comprising " used in the specification of the disclosure is Refer to that there are this feature, integer, step, operation, element and/or component, but it is not excluded that in the presence of or addition it is one or more its His feature, integer, step, operation, element, component and/or their combination.It should be understood that when we claim element to be " connected " Or when " coupled " to another element, it can be directly connected or coupled to other elements, or there may also be intermediary elements.This Outside, " connection " or " coupling " used herein may include being wirelessly connected or wirelessly coupling.Wording "and/or" packet used herein Include one or more associated wholes for listing item or any cell and all combination.

How the technical solution of the disclosure and the technical solution of the disclosure are solved with specifically embodiment below above-mentioned Technical problem is described in detail.These specific embodiments can be combined with each other below, for the same or similar concept Or process may repeat no more in certain embodiments.Below in conjunction with attached drawing, embodiment of the disclosure is described.

The embodiment of the present disclosure provides a kind of training method of neural network model, as shown in Figure 1, this method can wrap It includes:

Step S110 obtains training sample data, and training sample data include the first image set and the second image set, In, it include different original image in the first image set, the second image set includes the changing image of original image.

Wherein, used sample data when training sample data are for being trained to initial neural network model, It is referred to as training dataset.It include the first image set and the second data in the training sample data in the embodiment of the present disclosure Collect, may include an at least original image in first image set, and be different image between every original image.It needs It is noted that difference here refers to the difference of the content between every original image, object in image such as can be (such as Foreground object in image) different image, when from the point of view of similarity, two different original images are referred to The setting value of similarity between two original images less than a very little.

In practical applications, the image in the first image set can be from image data and concentrate the image randomly selected, needle When to an original image, removing the image except the original image in the first image set at this time is negative sample.For example, the first figure It include image A, image B and image C in image set, image A can be the image of panda, and image B can be landscape painting picture, image C It can be Architectural drawing.

In addition, further including the second image set in training sample data, the image in second image set is the first image set In original image changing image (negative sample), the main contents of the changing image of original image are substantially the same, original It opens image and its changing image is not belonging to different images.For example, the image of converting of image A can be image for image A The image A1 that A is formed after cutting, the image A2 formed after amplification, or the image A3 formed after reducing.As One optional way, the changing image of original image can be by the various processing modes that the pre-set image that carries out that treated, The processing mode can include but is not limited to scaling, cutting, Fuzzy Processing, addition comment, mirror image processing, rotation and addition text Processing etc..

As an example, the schematic diagram an of original image and the changing image of the original image is shown in Fig. 2, such as Shown in Fig. 2, original image shown in Fig. 2 can be commented on, at mirror image by diminution, cutting, Fuzzy Processing, addition respectively Reason, rotation and addition word processing, obtain the changing image of the original image, the specific figure as in Fig. 2 in addition to base image As shown in.

Step S120 is trained neural network model based on training sample data, until meeting preset training knot Beam condition；Wherein, the input of neural network model is image, exports the Hash codes for image, and training termination condition includes:

First similarity and the second similarity meet preset condition, wherein the first similarity be different original images it Between similarity, similarity of second similarity between each original image and the changing image of the original image；First phase Like the Hash codes determination that degree and the second similarity are based on image.

Wherein, similarity of first similarity between different original images, the second similarity are each original image Similarity between the changing image of the original image, for example, original image has image a and image b, the Transformation Graphs of image a As being a1, the changing image of image b is b1；The similarity of image a and image b are the first similarity, image a and image at this time The similarity and image b of a1 and the similarity of image b1 are the second similarity.In addition, judging the first similarity and second It is the first similarity and the second phase for judging to be calculated based on each training result when whether similarity meets preset condition Whether meet preset condition like degree.

It in practical applications, can be by image included in training sample data after getting training sample data It is separately input into initial neural network model, obtains the Hash codes of each image；Hash based on different original images The first similarity is calculated in code, the Hash codes of the changing image of the Hash codes based on each original image and each original image The second similarity is calculated, then judges whether the first similarity and the second similarity meet preset condition, if not meeting Preset condition then illustrates that requirement has not been reached yet in the precision of current neural network model；Adjustable neural network mould at this time Then image included in training sample data is separately input into initial neural network model by the parameter of type again, The Hash codes of each image are obtained, and determine the first similarity and the second similarity, then first obtained is this time trained in judgement Whether similarity and the second similarity meet preset condition, if not meeting preset condition, adjust neural network model again Parameter, until obtaining the first similarity and the second similarity meets preset condition.

Wherein, the similarity between different images can be determined by Hamming distance, and Hamming distance is referred to two A Hash codes are compared, if Hash codes are 64 01 character strings, the Hamming distance between two Hash codes, i.e., by two 64 0 and 1 character string of position is compared, and the different digit of value is Hamming distance in two Hash codes, if two images Hash codes between Hamming distance it is smaller, then illustrate two images it is more similar.

The training method that the embodiment of the present disclosure provides, can be by the setting of training termination condition, so that passing through instruction When the neural network model obtained after white silk obtains the Hash codes of image, it can guarantee that the Hash codes of original image are converted with it Hamming distance between the Hash codes of image be it is as small as possible, between the Hash codes of original image and the Hash codes of different images Hamming distance be it is as big as possible, that is, ensure that the Hash codes of different images are different, similar image be it is identical, And then when carrying out duplicate removal to image, multiimage can be determined according to the Hash codes of different images, to solve the prior art In the problem of being supplied to image duplicate there may be content in the image of user, and then promote user experience.

Based on this, show as shown in figure 3, the embodiment of the present disclosure provides the principle that a kind of pair of neural network model is trained It is intended to.Wherein, original image is I in Fig. 3, and the negative example diagram picture of original image is N, and the changing image of original image is I ', right The target that initial neural network model is trained is enabled between the Hash codes of original image and the Hash codes of its changing image Hash codes it is as identical as possible, the Hash codes of original image and negative example diagram picture be not as identical as possible, i.e. the Hash codes of original image (i.e. Hanming (H (I), H (I ')) is as small as possible, original image to Hamming distance between the Hash codes of its changing image Hash codes and negative example diagram picture Hash codes between Hamming distance (i.e. Hanming (H (I), H (N)) is as big as possible.

In disclosure optional embodiment, preset condition includes:

Alternatively,

It is understood that when the similarity between image using Hamming distance to characterize when, when similarity be greater than it is a certain When threshold value, that is, Hamming distance is characterized less than a certain threshold value, is illustrated by taking the first prediction accuracy among the above as an example, such as every time When training in calculated all first similarities less than the accounting of the quantity of the first similarity of the first setting value, that is, characterize It is greater than the accounting of the quantity of the Hamming distance of a certain setting value in all Hamming distances.

In practical applications, preset condition can be configured according to actual needs.It such as can be the first prediction accuracy Not less than first threshold, and the second prediction accuracy is greater than third threshold not less than second threshold or third prediction accuracy Value, is below described in detail both different situations respectively.

1, the first prediction accuracy is not less than first threshold, and the second prediction accuracy is not less than second threshold.

Wherein, less than the first setting value in calculated all first similarities when the first prediction accuracy is training every time The first similarity quantity accounting, that is, be directed to variant original image, obtaining the of each original image between any two After one similarity, judge whether the first similarity of original image between any two is less than first threshold, then determines all first Less than the ratio of the quantity of the first setting value and the quantity of all first similarities in similarity.

In one example, if the original image in the first image set includes image A, image B and image C, based on nerve Network model obtains the Hash codes of the Hash codes of image A, the Hash codes of image B and image C, is then based on the Hash codes of image A With the Hash codes of image B, it is determined that the first similarity AB of image A and image B, the Kazakhstan of Hash codes and image C based on image A Uncommon code, it is determined that the first similarity AC of image A and image C, and the Hash codes of the Hash codes based on image B and image C, really The first similarity BC of image C and image B is determined, and the first similarity AB is less than the first setting value, the first similarity AC and One similarity BC is all larger than the first setting value, and the first prediction accuracy is 1/3 at this time.

It is greater than the of the second setting value when second prediction similarity is training every time in calculated all second similarities The accounting of the quantity of two similarities is directed to variant original image, is obtaining each original image and its respective Transformation Graphs After the second similarity between any two, whether the second similarity judged then determines second less than the second setting value Ratio of the similarity less than the quantity of the second setting value and the quantity of all second similarities.

In one example, if including the changing image that original image includes image A and image A in training sample data The changing image B1 and image C of A1, image B and image B and the changing image C1 of image C, and obtained by neural network model To the Hash codes of image A, the Hash codes of image A1, the Hash codes of image B, the Hash codes of image B1 and image C Hash codes, The Hash codes of image C1 are then based on the Hash codes of image A and the Hash codes of image A1, determine the second of image A and image A1 Similarity AA1, the Hash codes of Hash codes and image B1 based on image B determine the second similarity BB1 of image B and image B1, The Hash codes of Hash codes and image C1 based on image C, determining and image C and image C1 the second similarity CC1, and determine Second similarity AA1 is all larger than the second setting value less than the second setting value, the second similarity BB1 and the second similarity CC1, then this When the second prediction accuracy be 2/3.

In practical applications, however, it is determined that the first prediction accuracy be not less than first threshold, and the second prediction accuracy is not Less than second threshold, then illustrate the precision of neural network model to reach requirement, otherwise need to continue to neural network model into Row training, until the first prediction accuracy is not less than first threshold, and the second prediction accuracy is not less than second threshold.

Wherein, the first setting value, the second setting value, first threshold and the specific numerical value of second threshold can be according to practical need It is pre-configured with, the embodiment of the present disclosure is without limitation.

2, third prediction accuracy is greater than third threshold value.

Wherein, the quantity of the similarity pair met the requirements in all similarities pair when third prediction accuracy is training every time Accounting, and each similarity refers in an original image the first similarity corresponding from other different original images most Minimum value in second similarity corresponding to big value and this original image and its each changing image.

It in one example, include the changing image A1 of image A, image B and image C and image A in training sample data With changing image A2, the changing image B1 and changing image B2 of image B, the changing image C1 and changing image C2 of image C；For Image A, it is determined that the first similarity AC of the first similarity AB, image A and the image C of image A and image B, and AB > AC, with And the second similarity AA2 of the second similarity AA1, image A and the image A2 of image A and image A1, and AA1 > AA2, scheme at this time As the corresponding similarity of A is to being then the first similarity AB and the second similarity AA2, and AB < AA2, then it is assumed that image A is corresponding Similarity to the similarity pair to meet the requirements；For image B, it is determined that the first similarity BA of image B and image A, figure As the first similarity BC of B and image C, and the second similarity BB1, image B and the image B2 of BA > BC, image B and image B1 The second similarity BB2, and BB2 > BB1, the corresponding similarity of image C is to being then that the first similarity BA and second is similar at this time BB1, and BA > BB1 are spent, then recognizes the corresponding similarity of image B to being unsatisfactory for requiring；For image C, it is determined that image C and figure As the first similarity BC of the first similarity AC, image B and the image C of A, and the second of AC > BC, image C and image C1 is similar The second similarity CC2 of CC1, image C and image C2, and CC2 > CC1 are spent, the corresponding similarity of image C is to being then first at this time Similarity AC and the second similarity CC1, and AC < CC1 then recognize the corresponding similarity of image C to meeting the requirements；That is, It altogether include 3 pairs of similarities pair in this training, and the quantity of the similarity pair met the requirements in all similarities pair is 2 pairs, then Third prediction accuracy at this time is 2/3.

Wherein, third threshold value can according to need is pre-configured with, and the embodiment of the present disclosure without limitation, and needs to illustrate It is that the first prediction accuracy, the second prediction accuracy described in the embodiment of the present disclosure and third prediction accuracy are bases What the result obtained after secondary training each time determined, if trained next time, the first prediction accuracy, the second prediction Accuracy and third prediction accuracy are determined based on the result obtained train next time after.

Alternatively, in practical applications, it can use existing image data set, image data set includes The image of each style and classification, the object that each image is included are different.The each image concentrated for image data (be used as original image) (such as zooms in and out image, cuts, mirror image, watermark and scheming by copying identical special efficacy Increase comment etc. as in), the changing image of the image is obtained, the positive example as the image.For every original image, instructing Other original images that certain data volume can be randomly selected when practicing from data set, as negative example.For every original graph Picture, may pass through the training with its positive example and negative example of certain number, to finally make original image and its changing image It can produce very much like Hash codes, and there will be relatively different Hash codes for different images.

In the embodiments of the present disclosure, since the condition that the training terminates is different similarity between original image and every Similarity preset condition between a original image and the changing image of each original image passes through the nerve net that is to say, being illustrated Hamming distance between the Hash codes of the Hash codes that network model obtains, the Hash codes of original image and its changing image is as far as possible Small, the Hamming distance between the Hash codes of original image and the Hash codes of different images is as big as possible, i.e. different images Hash codes be different, similar image is identical, so to image carry out duplicate removal when, can be according to the Hash of image Code determines multiimage, asking there may be the duplicate image of content in the image to solve to be supplied to user in the prior art Topic, and then promote user experience.

In practical applications, when user is to want to search for some images, to user's provided in prior art Image, it will usually comprising original image and the original changing image, and for users, original image and its changing image institute The content for the main substance for including is identical, that is, the duplicate image of content, therefore can reduce the user's body of user It tests.Based on this, the embodiment of the present disclosure provides a kind of determination method of multiimage, as shown in figure 4, this method may include:

Step 210, the first image to be processed and the second image to be processed are obtained.

Step 220, the first Hash codes of the first image to be processed and the second Hash codes of the second image to be processed are determined, In, the first Hash codes and the second Hash codes are determined by neural network model, and neural network model is based on refreshing among the above What the training method training through network model obtained.

Wherein, the first image to be processed may include an image or multiple images, and the second image to be processed also can wrap Include an image or multiple images.That is, image to be processed for first and second to be processed in the embodiment of the present application The quantity of image does not limit.

In practical applications, after the first image to be processed and the second image to be processed being separately input into training in advance Neural network model, obtain the Hash codes of the first image to be processed and the Hash codes of the second image to be processed, then will The Hash codes storage arrived in the database, can be obtained directly from database in the Hash codes for needing image.Certainly, in reality In the application of border, the Hash codes of the first image to be processed and the second image to be processed can not also be stored in advance, it is to be processed in needs When the Hash codes of image, the first image to be processed and the second image to be processed are separately input into the neural network mould after training Type, and then obtain the Hash codes of the first image to be processed and the second image to be processed.

Wherein, the specific implementation of training neural network model, as soon as it may refer to the description in embodiment, herein It repeats no more.

Step 230, the first Hash codes and the second Hash codes are based on, determine the first image to be processed and the second image to be processed Whether repeat.

Wherein, the first image to be processed is being determined according to the first Hash codes and the second Hash codes and the second image to be processed is When no repetition, the similarity between image can be characterized by the Hamming distance between the first Hash codes and the second Hash codes, so Determine whether the first image to be processed and the second image to be processed repeat based on the similarity between image afterwards.

In disclosure optional embodiment, the first image to be processed and the second image to be processed are obtained, comprising:

Obtain picture search keyword；

Determine whether the image in initial image search result repeats；

This method further include:

Wherein, for picture search keyword for indicating that user wants which type of image got, which characterizes need to search The feature of the image of rope.For example, when wanting user to want to search for image relevant to Doraemon, it can be by " Doraemon " as figure As search key, it can scanned for based on " Doraemon ", and get the related initial image with " Doraemon " and search Hitch fruit.

Correspondingly, in practical applications, since the first image to be processed and the second image to be processed are that initial image is searched Image in hitch fruit, and then when determining whether the first image to be processed and the second image to be processed repeat, it as determines just Whether the image in the image search result of beginning repeats.Further, initial image is being obtained based on picture search keyword After search result, the initial image search result can be determined according to the Hash codes of each initial image search result whether There are duplicate images, if there are duplicate images for the initial image search result, in order to guarantee the image for being supplied to user Multiimage is not present in search result, and then improves the usage experience of user, it can be in initial image search result Image carries out duplicate removal (multiimage in the first image to be processed and the second image to be processed is carried out duplicate removal), and by duplicate removal Image search result afterwards is supplied to user.

In one example, include image A, image B and image C in initial image search result, and determine image B and figure Picture C is multiimage, remaining image and image A can be provided to the wherein removal in image B and image C at this time To user.

In disclosure optional embodiment, determine whether the image in initial image search result repeats, comprising:

If calculated similarity no more than the 4th threshold value, creates an image group, using image as creating First image in image group；

Wherein, the image in image group including at least two images is multiimage.

In practical applications, when obtaining initial image search result according to picture search keyword, usual image is searched Rope is the result is that foundation and the sequence of picture search Keywords matching degree successively obtained from small to large, that is to say, that initial figure As the typically matching for searching for image and picture search keyword of serial number more forward of the image sequence in search result It spends higher.

It in practical applications, can be using first image as the first figure after getting initial image search result As first image of group, then since second image in initial image search result, according to the Hash codes of image, meter The similarity of first image in every image and each image group is calculated, and judges first image in every image and each image group Similarity whether be greater than the 4th threshold value, wherein each image group includes the first image group.

Correspondingly, if the similarity of first image Y in a certain image X and a certain group is greater than the 4th threshold value, explanation The image X and first image Y is duplicate image, and image X can be divided to figure corresponding to the first image Y at this time As in group；Certainly, if the similarity of first image in a certain image Z and all groups is no more than the 4th threshold value, explanation Image Z is not mutually duplicate image with first figure in all groups, an image group can be created at this time, by image Z As first image in newly-built image group, and in subsequent determining multiimage, calculate in every image and each image group First image similarity when, including with calculate every image and image Z similarity.

Further, after the completion of the image in initial image search result is handled, if including at least in image group Two images illustrate that each image in the image group is multiimage at this time.

Certainly, the similarity in image search result between each image can also be determined in practical application, but it is opposite When only calculating the similarity of first image in each image and each group image, calculation amount may be increased, and it also requires institute Some image search results are loaded and are come out, and may will increase time delay at this time.

In disclosure optional embodiment, the image search result after duplicate removal is supplied to user, comprising:

In practical applications, due in each image group each image be multiimage, i.e., first image in each image group Between be not multiimage.It therefore, can be only by each image when the image search result after duplicate removal is supplied to user First image in group is supplied to user, and can be avoided in the image for be supplied to user has the duplicate image of content.

It in one example, include image A, image B, image C and the figure successively searched in initial image search result As D, and being determined as image B and image C is multiimage, and image A and image D are multiimage, at this time can be by image B and figure As C is divided into an image group, image A and image D are divided into an image, image A and image B are then supplied to use Family.

In disclosure optional embodiment, using first image in each image group as the picture search knot after duplicate removal Fruit is supplied to after user, further includes:

In practical applications, after first image in image group is supplied to user, if user also wants to continue to check The image searched can trigger continue to check request at this time；Correspondingly, after receiving and continuing to check request, due to determination It is successively to be determined according to the image sequence in initial image search result when picture group belonging to each image, therefore, Image in each image group is naturally also to be arranged successively according to image sequence, it can is sorted according to the image in every group of image Which image next image for determining that the last time is supplied to the image of user is, and then continues to check request receiving When, next image of the image for being supplied to user last in every group of image group can be supplied to user.

In one example, the image in initial image search result is ordered as image A, image B, image C and image D, Image B and image C is multiimage, i.e. image B and image C are an image group, and image A and image D are multiimage, that is, are schemed Picture A and image D is an image, and image A and image B are supplied to user.After receiving and continuing to check request, determine The next image of image A is image D out, and the next image of image B is image C, can be mentioned image C and image D at this time Supply user.

In disclosure optional embodiment, the executing subject of this method is server or terminal device, if method is held Row main body is server, and the image search result after duplicate removal is supplied to user, comprising:

In practical applications, the determination method of the provided multiimage in the embodiment of the present disclosure can be by server It executes, can also be executed by terminal device.If the executing subject of this method is terminal device, by the picture search knot after duplicate removal When fruit is supplied to user, the image search result after weight directly can be shown to user；If the executing subject of this method is clothes Business device, then server needs for the image search result after duplicate removal to be sent to terminal device, and terminal device upon receipt, will be gone Image search result after weight is shown to user.

In one example, if as shown in figure 5, the executing subject of this method is terminal device, based on picture search key The initial image search result that word " stewardess " obtains includes image A to image H, wherein image A is first in search result Open image.When determining image A into image H with the presence or absence of multiimage, using image A as first figure of the first image group Picture will if the similarity for calculating image B and image A illustrates that image B and image A are not similar images less than the 4th threshold value Image B is as first image in second group of image group；Then the similarity of image C and image A is calculated less than the 4th threshold value, The similarity of image C and image B is greater than the 4th threshold value, image group (i.e. the second figure being at this time divided to image C where image B As group)；Further, the similarity of image D and image A are calculated less than the 4th threshold value, image D and the similarity of image B are small In the 4th threshold value, illustrate that image D and image A and image B are not multiimage, using image D as in third group image group First image；Based on identical mode, successively determine that image E to image H is multiimage with image D, at this time image E to figure Image group (i.e. third image group) as where H is divided to image D；It further, can be by image A, image B and image D It is sent to terminal device, terminal device upon receipt, will go image A, image B and image D to be shown to user and be supplied to user (image shown in Fig. 5 center)；After having received and continuing to check request, i.e., by image C and image E, it is sent to terminal and sets Standby, image C and image E are shown to user again upon receipt by terminal device.

In disclosure optional embodiment, the first image to be processed and the second image to be processed are in video to be processed Video frame images, determine whether the first image to be processed and the second image to be processed repeat to refer to the video frame in video to be processed Whether image repeats；

If current video frame image is greater than the 5th threshold value, current video frame relative to the similarity of associated video frame image Image is the multiimage relative to associated video frame image

In practical applications, when auditing to video content, the video frame images in usual video can also be deposited In a large amount of multiimage, as shown in FIG, but the workload of auditor can be greatly increased at this time, so being often desirable to Only unduplicated image is audited (as depicted in figure 6b), therefore, image to be processed can also be in video to be processed Video frame images.

Wherein, current video frame image refers to the video frame images being presently processing, such as in the third frame to video When image is handled, third video frame images are then current video frame image, when handling the 4th frame image, the 4th frame image It is then current video frame image.Since the first image to be processed and the second figure to be processed are the video frame images in video, due to There are corresponding front-rear position relationship between each video frame images in video, therefore the associated video frame of current video frame image Image then can be determining according to there are corresponding front-rear position relationships between each video frame images, and associated video frame image Quantity does not limit, for example can be the continuous video frame images of an at least frame before current video frame image, and/ Or, it is located at the continuous frame video frame images of an at least frame after current video frame image, and the continuous video frame of an at least frame Image refers to that with the position of current video frame image be continuous.

In one example, if as shown in fig. 7, the video frame images in video include video frame images F1~video frame figure As F7, in the example, 2 frames before the associated video frame image of a frame video frame images includes the video frame images are continuously regarded Frequency frame image and the 1 frame video frame images after video frame images.If currently pending video frame images are video frame Image F3, the continuous video frame images of 2 frames before video frame images F3 are video frame images F1 and video frame images F2 at this time, 1 frame video frame images after video frame images F3 are video frame images F4.

It is understood that in practical applications, if there is no at least frame companies before or after certain video to be processed Continuous video frame images, at least frame before the associated video frame image of the video frame images can be only at this time continuously regard Frequency frame image, or the continuous video frame images of an at least frame later；And if before a certain frame video frame images, or later only There are a frame video frame images, and 2 frames continuously regard before or after associated video frame image is provided that current video frame image Frequency frame image, before or after the associated video frame image of current video frame image can be only current video frame image at this time One frame video frame images.

Correspondingly, for the current video frame image in video to be processed, it can be according to the Hash of current video frame image The Hash codes of code and the associated video frame image of current video frame image, determine current video frame image and each associated video The similarity of frame image, and the similarity based on current video frame image Yu each associated video frame image, determine current video Similarity of the frame image relative to associated video frame image；Further, if current video frame image is relative to associated video frame The similarity of image is greater than the 5th threshold value, then illustrates current video frame image and associated video frame image is multiimage.Wherein, The 5th specific value of threshold value can be preset according to actual needs, and the embodiment of the present disclosure is without limitation.For example, when wanting essence When spending higher, a biggish numerical value can be set by the 5th threshold value.

In disclosure optional embodiment, if associated video frame image is at least two field pictures, it is based on current video frame The similarity of image and each associated video frame image determines current video frame image relative to the similar of associated video frame image Degree, comprising:

In practical applications, if the associated video frame image of current video frame image is at least two video frame images, It can be according to the Hash codes of current video frame image and the Hash codes of each associated video frame image and preset each association The weight of video frame images determines similarity of the current video frame figure relative to associated video frame image, works as example, can determine Hamming distance between the Hash codes of preceding video frame images and the Hash codes of each associated video frame image, may then based on pre- The weight of first configuration association video frame images determines similarity of the current video frame image relative to associated video frame image.

Wherein, similarity and preset each pass based on current video frame image and each associated video frame image The weight for joining video frame images, determines specific implementation side of the current video frame image relative to the similarity of associated video frame image Formula can be preset according to demand.

In the embodiments of the present disclosure, as a kind of optional mode, based on current video frame image and each associated video The weight of the similarity of frame image and preset each associated video frame image, determines current video frame image relative to pass Join video frame images similarity, may include:

By current video frame image and the similarity of each associated video frame image and corresponding associated video frame image Weight carries out multiplying, obtains the product that current video frame image corresponds to each associated video frame image；

The product that current video frame image corresponds to each associated video frame image is subjected to summation operation, by obtain and value Similarity as current video frame image relative to associated video frame image

That is, can be by the similarity and the associated video of current video frame image and each associated video frame image The corresponding multiplied by weight of frame image obtains the product that current video frame image corresponds to each associated video frame image, then will work as Preceding video frame images product corresponding with each associated video frame image is added, and obtained sum is opposite as current video frame image In the similarity of associated video frame image.

It is video frame images F3 in current video frame image, and video frame images F3 is transformation in example as shown in Figure 7 Video frame images near scene, the associated video frame image of video frame images F3 are F1, F2, F4, at this time video frame images F3 Similarity relative to associated video frame image can be determined by following equation:

sim₃=α sim_{1_3}+βsim_{2_3}+γsim_{3_4}

Wherein, sim₃Indicate similarity of the video frame images F3 relative to associated video frame image, sim_1-3Indicate video frame The similarity of image F3 and video frame images F1, α indicate the weight of preset associated video frame image F1；sim_2-3Indicate video The similarity of frame image F3 and video frame images F2, β indicate the weight of preset associated video frame image F2；sim_3-4Indicate view The similarity of frequency frame image F3 and video frame images F4, γ indicate the weight of preset associated video frame image F3.

In practical applications, usually not attached to scene change when determining the multiimage in video frame images to be processed Close video frame images disappear again, because the video frame images near scene change are easiest to go wrong, need to enter audit.Cause This if the scene of video frame images F3 and video frame images F4 two field pictures is different scenes, is being counted in the embodiments of the present disclosure When calculating similarity, the similarity of video frame images F3 and video frame images F4 will be very low, and video frame images 4 will be saved down Come.

Further, since the similarity calculation of every frame video frame images is not symmetrically, i.e., as that can not introduce view in Fig. 7 The calculating of frequency two frame of frame image F3 and video frame images F5 can be promoted under the premise of not influencing accurately in this way by verifying The amount of recalling (as soon as such as the frame image second from the bottom under scene can be disappeared weight, and the calculating for introducing video frame images may disappear Do not fall), further, due to only calculating the similarity between current video frame image and associated video frame, it is not necessarily at this time All videos frame is wanted to load out time delay can be effectively reduced.

It should be noted that determining phase of other video frame images relative to associated video frame image in video to be processed Like degree formula and principle among the above determine video frame images F3 relative to the similarity of associated video frame image formula and Principle is similar, in specific implementation can be with reference to determining video frame images F3 among the above relative to the similar of associated video frame image The implementation of degree, details are not described herein again for the embodiment of the present disclosure.

In disclosure optional embodiment, the weight of each associated video frame image is based on each associated video frame figure As what is determined with the positional relationship of current video frame image.

In disclosure optional embodiment, positional relationship includes adjacent position relationship, and adjacent position relationship is direct phase Adjacent or indirect neighbor, wherein be greater than with the weight of the associated video frame image of current video frame image direct neighbor and work as forward sight The weight of the associated video frame image of Pin Zheng image indirect neighbor；

That is, as an optional way, with current video frame image closer to video frame images, weight is opposite Bigger, when identical as the relative positional relationship of current video frame image, before current video frame image video frame figure The weight of picture is greater than the weight for the video frame images being located at after current video frame.

It in practical applications, can be according to each associated video frame image when the weight of each associated video frame image is set It is determined with the positional relationship of current video frame image, and the position of each associated video frame image and current video frame image is closed System may include usually direct neighbor or indirect neighbor.

Wherein, direct neighbor refers to being located at the frame image before or after current video frame, and indirect neighbor refers to It is not the image of direct neighbor in positional relationship with current video frame.Such as in Fig. 7, current video frame image is video frame figure As F3, video frame images F2's and video frame images F4 is video frame images with video frame images F3 direct neighbor at this time, And video frame images F1 and video frame images F5 etc. are the video frame images with video frame images F3 indirect neighbor.

It in practical applications, usually can will be with current video frame image in the weight of default associated video frame image The weight of the associated video frame image of direct neighbor is greater than and the associated video frame image of current video frame image indirect neighbor Weight；If the adjacent position relationship of associated video frame image is identical, such as when being direct neighbor, then it is located at current video frame figure The weight of associated video frame image before picture is greater than the weight for the associated video frame image being located at after current video frame image.

For example, current video frame image is video frame images F3, the associated video frame of video frame images F3 shown in Fig. 7 Image is F1, F2 and F4, and the adjacent position relationship of associated video frame image F1 and video frame images F3 is indirect neighbor at this time, is closed The adjacent position relationship for joining video frame images F2 and F4 and video frame images F3 is direct neighbor, so associated video frame image The value of the weight α of F1 is respectively less than the value of the weight beta of associated video frame image F2 and the weight γ of associated video frame image F2 Value；Further, since associated video frame image F2 and F4 are the identical associated video frame image of adjacent position relationship, close at this time Join the value of weight γ of the value greater than associated video frame image F4 of the weight beta of video frame images F2.That is, β's takes Value > α value of value > γ.

In the embodiments of the present disclosure, due to the weight of current video frame image and the similarity of each associated video frame image Be provided that it is different, therefore calculated similarity calculation be it is asymmetric, in this way can not influence accurately under the premise of, The promotion amount of recalling.

Based on principle identical with method shown in Fig. 1, a kind of neural network mould is additionally provided in embodiment of the disclosure The training device 30 of type, as shown in figure 8, the training device 30 of the neural network model may include 310 He of data acquisition module Model training module 320, in which:

Data acquisition module 310, for obtaining training sample data, training sample data include the first image set and second Image set, wherein include different original image in the first image set, the second image set includes the Transformation Graphs of original image Picture；

Model training module 320, for being trained based on training sample data to neural network model, until meeting pre- If training termination condition；Wherein, the input of neural network model is image, exports the Hash codes for image, and training terminates item Part includes:

In disclosure optional embodiment, which includes:

Alternatively,

One kind provided by embodiment of the disclosure can be performed in the training device of the neural network model of the embodiment of the present disclosure The training method of neural network model, realization principle is similar, the training of the neural network model in each embodiment of the disclosure Movement performed by each module in device is and the step in the training method of the neural network model in each embodiment of the disclosure It is rapid corresponding, the detailed functions description of each module of the training device of neural network model specifically may refer to hereinbefore Shown in corresponding neural network model training method in description, details are not described herein again.

Based on principle identical with method shown in Fig. 5, a kind of multiimage is additionally provided in embodiment of the disclosure Determining device 40, as shown in figure 9, the determining device 40 of the multiimage may include image collection module 410, Hash codes determination Module 420 and multiimage determining module 430, in which:

Image collection module 410, for obtaining the first image to be processed and the second image to be processed；

Hash codes determining module 420, for determining the first Hash codes and the second image to be processed of the first image to be processed The second Hash codes, wherein the first Hash codes and the second Hash codes are determined by neural network model, neural network model It is by training the method training of neural network model to obtain among the above；

Multiimage determining module 430 determines the first image to be processed for being based on the first Hash codes and the second Hash codes Whether repeated with the second image to be processed.

In disclosure optional embodiment, which further includes retrieval module, is specifically used for:

Determine whether the image in initial image search result repeats；

Multiimage determining module is also used to:

In disclosure optional embodiment, image of the multiimage determining module in the initial image search result of determination When whether repeating, it is specifically used for:

Wherein, the image in image group including at least two images is multiimage.

In disclosure optional embodiment, the image search result after duplicate removal is being supplied to use by multiimage determining module When family, it is specifically used for:

In disclosure optional embodiment, the executing subject of method is server or terminal device, if the execution master of method Body is server, and multiimage determining module is specifically used for when the image search result after duplicate removal is supplied to user:

In disclosure optional embodiment, the first image to be processed and the second image to be processed are the view in video to be processed Frequency frame image, determines whether the first image to be processed and the second image to be processed repeat to refer to the video determined in video to be processed Whether frame image repeats；

In disclosure optional embodiment, if associated video frame image is at least two field pictures, multiimage determining module In the similarity based on current video frame image Yu each associated video frame image, determine current video frame image relative to association When the similarity of video frame images, it is specifically used for:

In disclosure optional embodiment, the weight of each associated video frame image is based on each associated video frame image It is determined with the positional relationship of current video frame image.

In disclosure optional embodiment, positional relationship includes adjacent position relationship, and adjacent position relationship is direct neighbor Or indirect neighbor, wherein be greater than with the weight of the associated video frame image of current video frame image direct neighbor and current video The weight of the associated video frame image of frame image indirect neighbor；

A kind of repetition provided by embodiment of the disclosure can be performed in the determining device of the multiimage of the embodiment of the present disclosure The determination method of image, realization principle is similar, each mould in the determining device of the multiimage in each embodiment of the disclosure Movement performed by block be it is corresponding with the step in the determination method of the multiimage in each embodiment of the disclosure, for weight Each module of the determining device of complex pattern detailed functions description specifically may refer to hereinbefore shown in corresponding multiimage Determination method in description, details are not described herein again.

Based on principle identical with method shown in embodiment of the disclosure, one is additionally provided in embodiment of the disclosure Kind electronic equipment, the electronic equipment can include but is not limited to: processor and memory；Memory, for storing computer behaviour It instructs；Processor, for by calling computer operation instruction to execute method shown in embodiment.

Based on principle identical with method shown in embodiment of the disclosure, one is additionally provided in embodiment of the disclosure Kind computer readable storage medium, the computer-readable recording medium storage have at least one instruction, at least a Duan Chengxu, code Collection or instruction set, at least one instruction, an at least Duan Chengxu, code set or instruction set are loaded by processor and are executed on to realize Method shown in embodiment is stated, details are not described herein.

Scheme in embodiment of the disclosure is adapted to carry out the disclosure it illustrates one kind and implements below with reference to Figure 10 The structural schematic diagram of the electronic equipment 500 of example, which can be terminal device or server.Wherein, terminal device It can include but is not limited to such as mobile phone, laptop, digit broadcasting receiver, PDA (personal digital assistant), PAD The mobile terminal of (tablet computer), PMP (portable media player), car-mounted terminal (such as vehicle mounted guidance terminal) etc. with And the fixed terminal of such as number TV, desktop computer etc..Electronic equipment shown in Figure 10 is only an example, should not be right The function and use scope of the embodiment of the present disclosure bring any restrictions.

As shown in Figure 10, electronic equipment 500 may include processing unit (such as central processing unit, graphics processor etc.) 501, random access can be loaded into according to the program being stored in read-only memory (ROM) 502 or from storage device 508 Program in memory (RAM) 503 and execute various movements appropriate and processing.In RAM 503, it is also stored with electronic equipment Various programs and data needed for 500 operations.Processing unit 501, ROM 502 and RAM 503 pass through the phase each other of bus 504 Even.Input/output (I/O) interface 505 is also connected to bus 504.

In general, following device can connect to I/O interface 505: including such as touch screen, touch tablet, keyboard, mouse, taking the photograph As the input unit 506 of head, microphone, accelerometer, gyroscope etc.；Including such as liquid crystal display (LCD), loudspeaker, vibration The output device 507 of dynamic device etc.；Storage device 508 including such as tape, hard disk etc.；And communication device 509.Communication device 509, which can permit electronic equipment 500, is wirelessly or non-wirelessly communicated with other equipment to exchange data.Although Figure 10 shows tool There is the electronic equipment 500 of various devices, it should be understood that being not required for implementing or having all devices shown.It can be with Alternatively implement or have more or fewer devices.

Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be carried on computer-readable medium On computer program, which includes the program code for method shown in execution flow chart.In such reality It applies in example, which can be downloaded and installed from network by communication device 509, or from storage device 508 It is mounted, or is mounted from ROM 502.When the computer program is executed by processing unit 501, the embodiment of the present disclosure is executed Method in the above-mentioned function that limits.

It should be noted that the above-mentioned computer-readable medium of the disclosure can be computer-readable signal media or meter Calculation machine readable storage medium storing program for executing either the two any combination.Computer readable storage medium for example can be --- but not Be limited to --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor system, device or device, or any above combination.Meter The more specific example of calculation machine readable storage medium storing program for executing can include but is not limited to: have the electrical connection, just of one or more conducting wires Taking formula computer disk, hard disk, random access storage device (RAM), read-only memory (ROM), erasable type may be programmed read-only storage Device (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), light storage device, magnetic memory device, Or above-mentioned any appropriate combination.In the disclosure, computer readable storage medium can be it is any include or storage journey The tangible medium of sequence, the program can be commanded execution system, device or device use or in connection.And at this In open, computer-readable signal media may include in a base band or as the data-signal that carrier wave a part is propagated, In carry computer-readable program code.The data-signal of this propagation can take various forms, including but not limited to Electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be computer-readable and deposit Any computer-readable medium other than storage media, the computer-readable signal media can send, propagate or transmit and be used for By the use of instruction execution system, device or device or program in connection.Include on computer-readable medium Program code can transmit with any suitable medium, including but not limited to: electric wire, optical cable, RF (radio frequency) etc. are above-mentioned Any appropriate combination.

Above-mentioned computer-readable medium can be included in above-mentioned electronic equipment；It is also possible to individualism, and not It is fitted into the electronic equipment.

Above-mentioned computer-readable medium carries one or more program, when said one or multiple programs are by the electricity When sub- equipment executes, so that the electronic equipment executes method shown in above-described embodiment.

The calculating of the operation for executing the disclosure can be write with one or more programming languages or combinations thereof Machine program code, above procedure design language include object oriented program language-such as Java, Smalltalk, C+ +, it further include conventional procedural programming language-such as " C " language or similar programming language.Program code can Fully to execute, partly execute on the user computer on the user computer, be executed as an independent software package, Part executes on the remote computer or executes on a remote computer or server completely on the user computer for part. In situations involving remote computers, remote computer can pass through the network of any kind --- including local area network (LAN) Or wide area network (WAN)-is connected to subscriber computer, or, it may be connected to outer computer (such as utilize Internet service Provider is connected by internet).

Flow chart and block diagram in attached drawing are illustrated according to the system of the various embodiments of the disclosure, method and computer journey The architecture, function and operation in the cards of sequence product.In this regard, each box in flowchart or block diagram can generation A part of one module, program segment or code of table, a part of the module, program segment or code include one or more use The executable instruction of the logic function as defined in realizing.It should also be noted that in some implementations as replacements, being marked in box The function of note can also occur in a different order than that indicated in the drawings.For example, two boxes succeedingly indicated are actually It can be basically executed in parallel, they can also be executed in the opposite order sometimes, and this depends on the function involved.Also it to infuse Meaning, the combination of each box in block diagram and or flow chart and the box in block diagram and or flow chart can be with holding The dedicated hardware based system of functions or operations as defined in row is realized, or can use specialized hardware and computer instruction Combination realize.

Being described in unit involved in the embodiment of the present disclosure can be realized by way of software, can also be by hard The mode of part is realized.Wherein, the title of unit does not constitute the restriction to the unit itself under certain conditions, for example, the One acquiring unit is also described as " obtaining the unit of at least two internet protocol addresses ".

Above description is only the preferred embodiment of the disclosure and the explanation to institute's application technology principle.Those skilled in the art Member is it should be appreciated that the open scope involved in the disclosure, however it is not limited to technology made of the specific combination of above-mentioned technical characteristic Scheme, while should also cover in the case where not departing from design disclosed above, it is carried out by above-mentioned technical characteristic or its equivalent feature Any combination and the other technical solutions formed.Such as features described above has similar function with (but being not limited to) disclosed in the disclosure Can technical characteristic replaced mutually and the technical solution that is formed.

Claims

1. a kind of training method of neural network model characterized by comprising

Training sample data are obtained, the training sample data include the first image set and the second image set, wherein described first It include different original image in image set, second image set includes the changing image of the original image；

Neural network model is trained based on the training sample data, until meeting preset trained termination condition；Its In, the input of the neural network model is image, exports the Hash codes for image, the trained termination condition includes:

First similarity and the second similarity meet preset condition, wherein first similarity be different original images it Between similarity, similarity of second similarity between each original image and the changing image of the original image, institute It states the first similarity and the second similarity is the Hash codes determination based on image.

2. the method according to claim 1, wherein the preset condition includes:

First prediction accuracy is not less than first threshold, and the second prediction accuracy is not less than second threshold, wherein described first Less than the first similarity of the first setting value in calculated all first similarities when prediction accuracy is each training The accounting of quantity is greater than second in calculated all second similarities when the second prediction similarity is each training and sets The accounting of the quantity of second similarity of definite value；

Alternatively,

Third prediction accuracy is greater than third threshold value, wherein the third prediction accuracy all similarities when being training every time The accounting of the quantity of the similarity pair met the requirements in, each similarity is to similar for corresponding to an original image first Minimum value in second similarity corresponding to maximum value and an original image in degree, it is described to require to be the maximum Value is less than the minimum value.

3. a kind of determination method of multiimage characterized by comprising

Obtain the first image to be processed and the second image to be processed；

Determine the first Hash codes of the described first image to be processed and the second Hash codes of second image to be processed, wherein First Hash codes and second Hash codes are determined by neural network model, and the neural network model is to pass through What the method training in claim 1 or claim 2 obtained；

Based on first Hash codes and second Hash codes, the described first image to be processed and described second to be processed is determined Whether image repeats.

4. according to the method described in claim 3, it is characterized in that, the first image to be processed of the acquisition and the second figure to be processed Picture, comprising:

Obtain picture search keyword；

Initial image search result, first image to be processed and described second are obtained based on described image search key Image to be processed is the image in the initial image search result；

Whether the determination described first image to be processed and second image to be processed repeat, comprising:

Determine whether the image in the initial image search result repeats；

The method also includes:

Based on the multiimage determined in the initial image search result, the image search result after duplicate removal is supplied to User.

5. according to the method described in claim 4, it is characterized in that, figure in the determination initial image search result It seem no repetition, comprising:

According to the image sequence in the initial image search result, by first in the initial image search result First image of the image as the first image group executes following operation since second image:

If there is the similarity greater than the 4th threshold value in calculated similarity, the phase greater than the 4th threshold value is added the image to Like in the corresponding image group of degree；

If calculated similarity creates an image group no more than the 4th threshold value, using image as newly-built image group In first image；

Wherein, the image in image group including at least two images is multiimage.

6. according to the method described in claim 5, it is characterized in that, the image search result by after duplicate removal is supplied to use Family, comprising:

7. according to the method described in claim 6, it is characterized in that, the first image using in each image group is as duplicate removal Image search result afterwards is supplied to after user, further includes:

Receive continue to check request when, next image of the image for being supplied to user last in every group of image group is mentioned Supply user.

8. method according to any one of claims 4 to 7, which is characterized in that the executing subject of the method is service Device or terminal device, if the executing subject of the method is the server, the image search result by after duplicate removal is provided To user, comprising:

Image search result after duplicate removal is sent to the terminal device, so that equipment will be after the duplicate removal in the terminal Image search result is shown to user.

9. according to the method described in claim 3, it is characterized in that, first image to be processed and second figure to be processed As being the video frame images in video to be processed, the determination described first image to be processed and second image to be processed are No repetition, which refers to, determines whether the video frame images in the video to be processed repeat；

It is described be based on first Hash codes and second Hash codes, determine the described first image to be processed and described second to Whether processing image repeats, comprising:

For the current video frame image in the video to be processed, according to the Hash codes of the current video frame image with it is described The Hash codes of the associated video frame image of current video frame image, determine the current video frame image and each associated video frame The similarity of image；

Similarity based on the current video frame image Yu each associated video frame image, determines the current video frame image Similarity relative to associated video frame image；

Wherein, the continuous video frame images before the associated video frame image includes the current video frame image, and/ Or, the continuous video frame images after the current video frame image；

If the current video frame image is greater than the 5th threshold value relative to the similarity of associated video frame image, described to work as forward sight Frequency frame image is the multiimage relative to the associated video frame image.

10. according to the method described in claim 9, it is characterized in that, if the associated video frame image is at least two field pictures, The similarity based on the current video frame image Yu each associated video frame image, determines the current video frame image Similarity relative to associated video frame image, comprising:

Similarity and preset each associated video based on the current video frame image and each associated video frame image The weight of frame image determines similarity of the current video frame image relative to associated video frame image.

11. according to the method described in claim 10, it is characterized in that, the weight of each associated video frame image is to be based on What the positional relationship of each associated video frame image and the current video frame image determined.

12. according to the method for claim 11, which is characterized in that the positional relationship includes adjacent position relationship, described Adjacent position relationship is direct neighbor or indirect neighbor, wherein the associated video with the current video frame image direct neighbor The weight of frame image is greater than the weight with the associated video frame image of current video frame image indirect neighbor；

When the adjacent position relationship is identical, the weight of the associated video frame image before the current video frame image Greater than the weight of the associated video frame image after the current video frame image.

13. a kind of training device of neural network model characterized by comprising

Data acquisition module, for obtaining training sample data, the training sample data include the first image set and the second figure Image set, wherein it includes different original image that the first image, which is concentrated, and second image set includes the change of original image Change image；

Model training module, for being trained based on the training sample data to neural network model, until meeting default Training termination condition；Wherein, the input of the neural network model is image, exports the Hash codes for image, the training Termination condition includes:

First similarity and the second similarity meet preset condition, wherein first similarity be different original images it Between similarity, similarity of second similarity between each original image and the changing image of the original image；Its In, first similarity and the second similarity are determined based on the Hash codes of image.

14. a kind of determining device of multiimage characterized by comprising

Hash codes determining module, for determine the described first image to be processed the first Hash codes and second image to be processed The second Hash codes, wherein first Hash codes and second Hash codes are determined by neural network model, described Neural network model is obtained by the method training in claim 1 or claim 2；

Multiimage determining module is determined for being based on first Hash codes and second Hash codes described first wait locate Whether reason image and second image to be processed repeat.

15. a kind of electronic equipment characterized by comprising

Processor and memory；

The memory, for storing computer operation instruction；

The processor, for by calling the computer operation instruction, perform claim to be required described in any one of 1 to 12 Method.

16. a kind of computer readable storage medium, which is characterized in that the readable storage medium storing program for executing be stored at least one instruction, At least a Duan Chengxu, code set or instruction set, at least one instruction, an at least Duan Chengxu, the code set or refer to Collection is enabled to be loaded as processor and executed to realize method described in any one of claims 1 to 12.