CN110163061A - For extracting the method, apparatus, equipment and computer-readable medium of video finger print - Google Patents
For extracting the method, apparatus, equipment and computer-readable medium of video finger print Download PDFInfo
- Publication number
- CN110163061A CN110163061A CN201811353102.6A CN201811353102A CN110163061A CN 110163061 A CN110163061 A CN 110163061A CN 201811353102 A CN201811353102 A CN 201811353102A CN 110163061 A CN110163061 A CN 110163061A
- Authority
- CN
- China
- Prior art keywords
- video
- finger print
- fingerprint characteristic
- video frame
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Collating Specific Patterns (AREA)
Abstract
It discloses a kind of for extracting the method, apparatus, equipment and computer-readable medium of the video finger print of video.The described method includes: extracting multiple video frames in video；For each of the multiple video frame, the video frame is handled using the neural network with multiple layers, it wherein, include at least one convolutional layer in multiple layers, each of at least one described convolutional layer is used for the output to its preceding layer and carries out process of convolution；The convolution processing result that the middle layer of the neural network is exported is as the fingerprint characteristic of the video frame；And the fingerprint characteristic of the multiple video frame of processing, to generate the video finger print of the video.
Description
Technical field
This disclosure relates to field of video processing, and in particular to a kind of method, apparatus for extracting the video finger print of video,
Equipment and computer-readable medium.
Background technique
In recent years, the growth rate of the quantity of network video is very rapid.In order to realize that effective video identifies, Ke Yitong
It crosses and image procossing is carried out to the multiple video frames for including in video, and generate the fingerprint characteristic of these video frames.By to these
Fingerprint characteristic handle the video finger print of the available video.The video finger print that benefit obtains in such a way can be used as
The unique features of video represent a video file.The description of video content is acted on using video finger print, may be implemented similar
The application such as comparison and cluster between video.
Since current video finger print has a wide range of applications during video compares and clusters, therefore, it is necessary to new
Method for extracting video fingerprints is to realize the rapidly extracting of video finger print and the quick comparison of video finger print.
Summary of the invention
For this purpose, the method, apparatus, equipment and computer that present disclose provides a kind of for extracting the video finger print of video can
Read medium.
According to one aspect of the disclosure, a kind of method for extracting the video finger print of video is provided, comprising: extract
Multiple video frames in video；For each of the multiple video frame, using the neural network with multiple layers to the view
Frequency frame is handled, wherein includes at least one convolutional layer in multiple layers, each of at least one described convolutional layer is used for
Process of convolution is carried out to the output of its preceding layer, using the convolution processing result of the output of the middle layer of the neural network as this
The fingerprint characteristic of video frame；And the fingerprint characteristic of the multiple video frame of processing, to generate the video finger print of the video.
In some embodiments, the fingerprint characteristic of the multiple video frame is handled, to generate the video finger print of the video
It include: that dimension-reduction treatment is executed respectively to the fingerprint characteristic of the multiple video frame, to obtain the fingerprint characteristic of multiple dimensionality reductions；Splicing
The video finger print feature of the multiple dimensionality reduction, to generate the video finger print of the video.
In some embodiments, the fingerprint characteristic of the multiple video frame is executed during dimension-reduction treatment includes the following steps
At least one: executing pondization processing to the fingerprint characteristic of the multiple video frame respectively, to obtain the fingerprint characteristic in multiple ponds,
And principal component analysis is carried out to the fingerprint characteristic in the multiple pond, to obtain the fingerprint characteristic of multiple dimensionality reductions.
In some embodiments, extract video in multiple video frames include: selected in the video equal intervals it is more
A frame is as the multiple video frame.
In some embodiments, the neural network by the video frame transformation at the image data with multiple channels,
The method also includes the image datas concurrently to the multiple channel to execute convolution operation.
In some embodiments, the neural network is Mobilenet network.
According to another aspect of the present disclosure, it additionally provides a kind of for extracting the device of video finger print, comprising: video frame mentions
Unit is taken, is configured to extract multiple video frames in video；Finger print characteristic abstract unit is configured in the multiple video frame
Each, is handled the video frame using the neural network with multiple layers, wherein includes at least one volume in multiple layers
Lamination, each of at least one described convolutional layer are used for the output to its preceding layer and carry out process of convolution；By the nerve
Fingerprint characteristic of the convolution processing result of the middle layer output of network as the video frame；And video finger print generation unit, match
It is set to the fingerprint characteristic for handling the multiple video frame, to generate the video finger print of the video.
In some embodiments, the video finger print generation unit further include: dimensionality reduction subelement is configured to the multiple view
The fingerprint characteristic of frequency frame executes dimension-reduction treatment respectively, to obtain the fingerprint characteristic of multiple dimensionality reductions；And splicing subelement, it is configured to
Splice the video finger print feature of the multiple dimensionality reduction, to generate the video finger print of the video.
In some embodiments, the dimension-reduction treatment at least one of includes the following steps: to the multiple video frame
Fingerprint characteristic execute respectively pondization processing, to obtain the fingerprint characteristic in multiple ponds, and the fingerprint to the multiple pond
Feature carries out principal component analysis, to obtain the fingerprint characteristic of multiple dimensionality reductions.
In some embodiments, extract video in multiple video frames include: selected in the video equal intervals it is more
A frame is as the multiple video frame.
In some embodiments, the neural network by the video frame transformation at the image data with multiple channels,
Described device is configured to concurrently execute convolution operation to the image data in the multiple channel.
In some embodiments, the neural network is Mobilenet network.
According to another aspect of the present disclosure, additionally provide a kind of equipment for extracting video finger print, including processor and
Memory is stored with program instruction in the memory, and when executing described program instruction by processor, the processor is matched
It is set to and executes the method for extracting the video finger print of video as previously described.
According to another aspect of the present disclosure, a kind of computer readable storage medium is additionally provided, instruction, institute are stored thereon with
Instruction is stated when being executed by processor, so that the processor executes as previously described for extracting the side of the video finger print of video
Method.
According to the disclosure provide for extracting the method, apparatus, equipment and computer-readable medium of video finger print, can be with
The fingerprint characteristic for the video frame in video is determined by new method for extracting fingerprint feature, and further special using the fingerprint
Levy the video finger print for determining the video.There is better Clustering Effect using the video finger print that the method that the disclosure provides obtains.
Detailed description of the invention
It, below will be to required use in embodiment description in order to illustrate more clearly of the technical solution of the embodiment of the present disclosure
Attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is only some embodiments of the present disclosure, for this
For the those of ordinary skill of field, without making creative work, it can also be obtained according to these attached drawings other
Attached drawing.The following drawings is not drawn by actual size equal proportion scaling deliberately, it is preferred that emphasis is shows the purport of the disclosure.
Fig. 1 shows the schematic diagram of an illustrative video finger print extraction system according to the disclosure；
Fig. 2 shows a kind of schematical processes for the method extracted for video finger print according to an embodiment of the present disclosure
Figure；
Fig. 3 shows the schematical flow chart of video finger print feature generation step according to an embodiment of the present disclosure；
Fig. 4 shows according to an embodiment of the present disclosure a kind of for extracting the schematical frame of the device of video finger print
Figure；And
Fig. 5 shows the illustrative framework for calculating equipment of one kind according to the disclosure.
Specific embodiment
To keep the purposes, technical schemes and advantages of the disclosure clearer, the disclosure is made by the following examples further
It is described in detail.Obviously, described embodiment is only disclosure a part of the embodiment, instead of all the embodiments.It is based on
Embodiment in the disclosure, it is obtained by those of ordinary skill in the art without making creative efforts all other
Embodiment belongs to the range of disclosure protection.
The video finger print of video is the unique video features generated by the analysis to the image information in video, the video
Fingerprint can represent corresponding video file.Video search may be implemented using video finger print, for example, due to video finger print for
The descriptive of video content can help to position similar video, therefore may be implemented to search answering for video with video using video finger print
With.In another example the quick comparison of two videos may be implemented using video finger print.It is just the same for two contents, but video
The discrepant video of frame per second, resolution ratio, by the Message Digest 5 md5 of video file can not judge two videos whether phase
Together.However, due to video finger print can with describing video contents, by compare video finger print can be found that content is identical
Two videos.Using such characteristic, video finger print may be implemented to operate such as executing duplicate removal to same video.In addition, sharp
With video finger print for the descriptive of video content, the video that user watched can be clustered according to video finger print, from
And the recommendation of similar video can be carried out to the interested classification of user.In addition, during video content is propagated, such as regard
On the shared platform of frequency, can be determined by video finger print and video uplink time in the most morning in the identical or similar video of height
The user of biography, so that it is determined that the author of video content, and pirate video is hit.
Since current video finger print has a wide range of applications during video compares and clusters, therefore, it is necessary to new
Method for extracting video fingerprints is to realize the rapidly extracting of video finger print and the quick comparison of video finger print.
Fig. 1 shows the schematic diagram of an illustrative video finger print extraction system according to the disclosure.As shown in Figure 1,
Video finger print extraction system 100 may include one or more clients 110, network 120, server 130 and database
140.For convenience of description, in the disclosure, video finger print extraction system 100 can be referred to as system 100.
Client 110 can include but is not limited to one of stationary electronic device or mobile electronic equipment or more
Kind.For example, stationary electronic device can include but is not limited to desktop computer, smart home device etc..Mobile electronic device can
To include but is not limited to one of smart phone, smartwatch, laptop, tablet computer, game station etc. or a variety of.
Client 110 can be communicated by network 120 with server, database or other clients, for example, by client local
The video of storage is sent to server 130 or other clients via network by the video that client is shot.For example, can be with
Utilize the camera program or other programs such as browser run in client, barcode scanning built-in in instant messaging (IM) or photograph
Phase program shoots video.In some embodiments of the present disclosure, client 110 can be configured for executing what the disclosure provided
Method for extracting video finger print.For example, client 110 can be directed to the video being locally stored or objective from other by network
Received video executes the method for extracting video finger print that the disclosure provides in family end and/or database.
Network 120 can be the combination of single network or multiple and different networks.For example, network 120 may include but unlimited
Combination in one or more of local area network, wide area network, internet etc..Network 120 can be used to implement client 110, clothes
The data exchange being engaged between device 130 and database 140.
Server 130 is the system that data can be carried out with analysis processing to generate analysis result.Server 130 can be
One individual server or a server farm, each server in group by wired or wireless network into
Row connection.In embodiment of the disclosure, server 130 can be configured for executing that the disclosure provides for extracting video
The method of fingerprint.Server 130 can execute the disclosure to video received from client 110 and/or database 140 and provide
The method for extracting video finger print.
Database 140 can refer to the equipment with store function.Database 140 is mainly used for storage from client 110
The various data for utilizing, producing and exporting in data and server 130 work of collection.For example, database 140 can be deposited
Store up the algorithm parameter for being used to generate video finger print being related in the method being described below.Database 140 can be used for depositing
The video finger print that storage is generated using the method that the disclosure provides.Database 140 can be local or long-range.Database 140
It can be impermanent memory memory or permanent memory memory.Above-mentioned storage equipment only lists some examples,
The storage equipment that the system can be used is not limited thereto.Database 140 can be connected with each other or communicate with network 120, or
Directly and system 100 or part of it (for example, server 130) is connected with each other or communicates or the combination of two ways.One
In a little embodiments, database 140 be can be set on the backstage of server 130.In some embodiments, database 140 can be
It is independent, directly it is connected with network 120.Connection or communication between database 140 and system other equipment can be it is wired,
Or wireless.
It should be noted that in addition to it is above-mentioned including the system of network other than, embodiment of the disclosure also can be implemented in list
In only local computer.
In the following, will be described in connection with the drawings embodiment of the disclosure.
Fig. 2 shows a kind of the schematic of the method that the video finger print for video extracts according to an embodiment of the present disclosure
Flow chart.Method 200 shown in Figure 2 can use client 110 shown in Fig. 1 or server 130 is realized.
As shown in Fig. 2, method 200 may include step S202.In step S202, multiple views in video can be extracted
Frequency frame.In the following description, characteristics of image can be executed for the image information for the multiple video frames extracted in step S202
Extraction, and using the characteristics of image extracted respectively for above-mentioned multiple video frames as the fingerprint characteristic of corresponding video frame.
After fingerprint characteristic mentioned here can be using having multiple layers of neural network to handle video frame, in neural network
The characteristic pattern of any one layer of output.Hereinafter, these video frames for being used to take the fingerprint feature are also referred to as key frame.
In step S202, the effective video segment for extracting video finger print can be determined according to video, and from effective
Key frame is extracted in video clip.
In some embodiments, a part of video can be determined as effective video segment.For example, video can be cut
It is segmented into the segment of predetermined time period (such as 15s or any other possible predetermined time period), and the segment that cutting is obtained
As effective video segment.For example, the segment of the 0 to 15th second of video can be determined as effective video segment, can also incite somebody to action
The segment for the predetermined time period that any position starts in video is determined as effective video segment.In further embodiments, may be used
Entire video is determined as effective video segment.
It is then possible to extract key frame from determining effective video segment.In some embodiments, it will can effectively regard
Each frame in frequency segment is determined as key frame.It in further embodiments, can be by each in effective video segment
Frame is analyzed, and realizes the shot segmentation to effective video segment, and extract according to the result of shot segmentation and be used for each camera lens
Key frame.For example, the first frame of each camera lens determined by shot segmentation can be determined as key frame.In other reality
It applies in example, the method selecting video frame of sampling can be passed through in video as key frame.For example, can be in a manner of equally spaced
(such as every preset time interval of 5s or any) samples effective video segment, and the video frame that sampling is obtained
As key frame.For the effective video segment of a 15 seconds length, when being sampled with 5 seconds for interval, 3 frames work will be obtained
For key frame.In another example can be sampled with arbitrary interval to effective video segment, and the video frame that sampling is obtained as
Key frame.
Then, in step S204, for each of the multiple video frame, the neural network with multiple layers is utilized
The video frame is handled, and using the middle layer characteristic pattern of the middle layer of neural network output as the finger of the video frame
Line feature.The fingerprint characteristic can be used for generating video finger print.
In some embodiments, which can be the neural network obtained using the training of image classification task.This
The neural network of sample can be made of the one or more in convolutional layer, pond layer, full articulamentum and activation primitive.For example, this
The open neural network with multiple layers used is the depth constituted by stacking multiple convolutional layers and at least one pond layer
Neural network.Wherein each convolutional layer is used for the output to its preceding layer and carries out process of convolution, and exports the knot after process of convolution
Fruit.For example, can use the reality that the structures such as Alexnet, Vggnet, Resnet, Googlenet, Mobilenet realize the disclosure
Apply neural network used in example.
The neural fusion image point that training can be made to obtain using suitable image data set training neural network
The function of class.That is, can be defeated using the last layer of the above-mentioned neural network obtained using the training of image classification task
The classifying image features of the input picture of the neural network and the classification results of prediction out.
According to the method that the disclosure provides, the above-mentioned neural network pair obtained using the training of image classification task can be used
Key frame is handled, and using the feature of the middle layer of neural network output as the fingerprint characteristic of currently processed key frame.
By taking Mobilenet structure as an example, a kind of illustrative structure of Mobilenet network is shown in table 1.Below
In, every a line in table 1 is referred to as one layer in Mobilenet network.For example, Mobilenet network shown in table 1
First layer is the convolutional layer that shape is 3 × 3 × 3 × 32, and the second layer is the depth convolutional layer that shape is 3 × 3 × 32.With such
It pushes away, layer second from the bottom is the full articulamentum FC that shape is 1024 × 1000, and the last layer is activation primitive Softmax.
Table 1
In order to adapt to the task of multi-tag, it can use sigmoid and intersect entropy function as in loss function training table 1
The Mobilenet structure shown.
Process of convolution can be carried out to the key frame of input using Mobilenet structure shown in table 1, and is generated and be used for
The characteristics of image of the image information of the key frame is described.In fact, each layer of neural network of output can be used for retouching
State input picture.But, it is however generally that, in deep neural network, the characteristics of image of deeper output compares shallow-layer output
Characteristics of image is to the descriptive much better of input picture.
In addition, as previously mentioned, the effect of video finger print is to be used for describing video contents, and can further refer to using video
The comparison and cluster of line progress similar video.Therefore, the feature which layer exports in each layer in order to determine neural network has
Best Clustering Effect can carry out clustering to the output of layer each in neural network, and be selected according to the result of clustering
Select fingerprint characteristic of the most suitable output as video frame.
For example, cluster point can be carried out by the characteristics of image exported to each layer of neural network structure shown in table 1
Analysis determines the characteristics of image for the output for selecting which layer as the fingerprint characteristic for being used for the key frame.For example, can choose including
The data set of multiple pictures, and the picture in the data set is handled using trained neural network.Due to data set
In include image content be it is known, therefore, can by middle layer different in neural network (layer such as second from the bottom,
One layer of layer 5 reciprocal or any other centre) output characteristics of image execute clustering algorithm (such as K mean cluster, average drifting
Cluster, density clustering method etc.) come the image clustering effect for the characteristics of image realization for examining neural network middle layer to export
Fruit.I.e., it is possible to examine the effect for determining similar pictures using the output of layer each among neural network.In some embodiments, may be used
Cluster is executed with the feature exported using clustering algorithm to each middle layer of neural network, and is judged obtained in above-mentioned cluster process
The consistency of image data in every one kind.
For illustrative neural network structure shown in table 1, such as it can determine the figure of its layer 5 output reciprocal
As the consistency of feature is higher, i.e., Clustering Effect is best, and therefore, as shown in table 1, the characteristics of image that this layer can be exported is made
For the fingerprint characteristic of processed key frame.
It is understood that above only describing for key frame for the Mobilenet structure shown in the table 1
A kind of illustrative determining method of fingerprint characteristic.Middle layer for exporting the fingerprint characteristic of key frame is not limited to neural network
Layer 5 reciprocal.It will be understood by those skilled in the art that when the figure for selecting other neural network structures to be used to extract key frame
It, can be according to the output of a middle layer of above method neural network selected to use as processed key when as feature
The fingerprint characteristic of frame.Those skilled in the art can according to the actual situation, the figure for the middle layer output for selecting Clustering Effect best
Fingerprint characteristic as feature as processed key frame.In addition, those skilled in the art can also make according to video finger print
With scene, is examined by other standards and select the characteristics of image of any middle layer output in neural network structure as processed
Key frame fingerprint characteristic.
Neural network used in step S204 may include depth convolution (depthwise convolution) layer.Example
Such as, as shown in table 1, the layers such as the 2nd, 4,6 are depth convolutional layers.In deep neural network, processed input picture can be with
It is transformed into the characteristics of image of the image data with multiple channels.For depth convolution, the volume of depth convolutional layer setting
The port number of product core is identical as the port number of processed characteristics of image, also, the convolution kernel in each channel is respectively used to this
The image data of respective channel carries out convolution in characteristics of image.Then, each channel convolution can be combined using 1 × 1 convolution kernel
Output.Such as the layers such as the 3rd, 5,7 shown in table 1.Pass through the separable volume of depth level using depth convolution sum point convolution composition
Product can replace Standard convolution traditional in neural network structure, while the computational efficiency of neural network can be improved, and reduce
Number of parameters in neural network structure.
When in neural network structure including the structure of depth convolution, in some embodiments, processor can use (such as
CPU single-instruction multiple-data stream (SIMD) (SIMD) characteristic) accelerates the speed of service of depth convolution.When needing to be implemented convolution operation, by
In on the different channels to processed characteristics of image data execute convolution operation instruction be it is identical, therefore, Ke Yili
With SIMD characteristic, each instruction is enabled to execute operation to the data parallel on multiple channels simultaneously, so as to improve processor
Handling capacity.In addition when the size of convolution kernel is greater than 1, convolution kernel has part when sliding on characteristic pattern (feature map)
Overlapping, the data for retaining lap in a register can reduce internal storage access number, to improve computational efficiency.It can make
The SIMD characteristic of processor is called with assembly instruction.
It optionally, can be to the execution of key frame before carrying out image characteristics extraction to key frame using neural network
Size change over (such as up-sampling or down-sampling), so that the size of key frame meets the input size of neural network.With table 1
Shown in for structure, the size of the picture frame of input can be 224 × 224 × 3.
Utilize the fingerprint characteristic of all key frames of the video extracted in the available step S202 of step S204.Then,
As shown in Fig. 2, can handle the fingerprint characteristic of the multiple video frame in step S206, to generate the video of the video
Fingerprint.In some embodiments, it can determine that the video of the video refers to by merging the fingerprint characteristic of the multiple video frame
Line.That is, the information of the fingerprint characteristic in video finger print including all key frames.
Fig. 3 shows the schematical process of video finger print feature generation step S206 according to an embodiment of the present disclosure
Figure.
As shown in figure 3, step S206 may further include step S2062.It, can be to described more in step S2062
The fingerprint characteristic of a video frame executes dimension-reduction treatment respectively, to obtain the fingerprint characteristic of multiple dimensionality reductions.
In some embodiments, dimension-reduction treatment may include pond step and principal component analysis step.For example, can to institute
The fingerprint characteristic for stating multiple video frames executes pondization processing respectively, to obtain the fingerprint characteristic in multiple ponds.For example, pondization is handled
It can be one of average pond, maximum pond or minimum pond or a variety of.With Mobilenet network knot shown in table 1
For structure, as previously mentioned, the feature that the layer 5 reciprocal of network shown in table 1 can be exported is as processed key frame
Fingerprint characteristic.As shown in table 1, the size of the fingerprint characteristic of this layer output is 7 × 7 × 1024.Then, pass through pond
Processing can carry out dimensionality reduction to the fingerprint characteristic that this layer exports.For example, the size of the fingerprint characteristic of Chi Huahou can be 1 × 1 ×
1024。
It is then possible to which the fingerprint characteristic to the multiple pond carries out principal component analysis, to obtain the fingerprint of multiple dimensionality reductions
Feature.Principal component analysis may include carrying out matrix respectively using fingerprint characteristic of the preset projection matrix to above-mentioned multiple ponds
Multiplication, to carry out dimensionality reduction to fingerprint characteristic.In some embodiments, projection matrix can be converted to corresponding convolution kernel
Form, and convolution is carried out to the fingerprint characteristic in above-mentioned pond, to obtain result identical with matrix multiplication.In some embodiments
In, it can use preset image data set and determine the projection matrix for being used for principal component analysis.For example, can use including a variety of
The d dimensional data image collection of the image information of type calculates the covariance matrix for being used for the data set.It is hereinbefore retouched in the disclosure
In the example stated, d=1024.It is then possible to calculate the characteristic value and corresponding feature vector of the covariance matrix, and therefrom select
Select feature vector corresponding with preceding k maximum eigenvalue.Wherein k is less than the integer of d.By corresponding to first k maximum feature
K feature vector of value can construct a projection matrix.Using the projection matrix can by d tie up fingerprint characteristic dimensionality reduction at
K dimension.In one example, k=128.It is understood that the value of k can be arranged in those skilled in the art according to the actual situation,
To obtain the fingerprint characteristic of various sizes of dimensionality reduction.
Although the fingerprint characteristic after dimensionality reduction has lost parts of images characteristic information, however, the drop generated using the above method
Fingerprint characteristic after dimension compares fingerprint characteristic of the successful accuracy rate only than non-dimensionality reduction with similar video and has dropped by examining
0.05%, still realize 96% comparison accuracy rate.It is therefore contemplated that fingerprint characteristic after dimensionality reduction and non-dimensionality reduction is complete
Fingerprint characteristic be identical to the descriptive of video content.Since the data volume in the fingerprint characteristic after dimensionality reduction is greatly reduced,
The comparison speed of fingerprint between video can thus be accelerated.
Step S206 can also include step S2064.In step S2064, the video of the multiple dimensionality reduction can be spliced
Fingerprint characteristic, to generate the video finger print of the video.For example, 3 can will be directed to when the crucial number of frames of video is 3
The fingerprint characteristic for the dimensionality reduction that key frame generates respectively is spliced.It is spliced available by taking the fingerprint characteristic of 128 dimensions as an example
The feature of one 384 dimension.The spliced feature can be used as the video finger print of the video.
Alternatively, in step S2064, can also video finger print feature to the multiple dimensionality reduction execute transformation with life
At the video finger print of the video.For example, can video finger print feature to the multiple dimensionality reduction execute feature hash conversion, and
The result of feature Hash is determined as to the video finger print of the video.
The method for extracting video finger print provided using the disclosure, can be by only to the partial video frame in video
The operation for executing the feature that takes the fingerprint, improves the formation speed of video finger print.In addition, by examining each layer output of neural network
The Clustering Effect of characteristics of image, the output that can choose the middle layer with best Clustering Effect are used to generate video finger print,
So that the video finger print generated has better effect during identifying similar video.Further, by mind
Output through network executes principal component analysis, can compare successful accuracy rate guaranteeing the video of video finger print and Similar content
While, the video decoded time is removed, the time CPU monokaryon that entire video generates fingerprint only needs 160 milliseconds.By to generation
The dimension of video finger print feature carry out dimensionality reduction so that the versus speed of fingerprint has obtained fast lifting between video, therefore can be with
Support the video search application based on video content.
Fig. 4 shows according to an embodiment of the present disclosure a kind of for extracting the schematical process of the device of video finger print
Figure.Client and/or server shown in Fig. 1 can be implemented as shown in Fig. 4 for extracting the device of video finger print.
As shown in figure 4, device 400 may include video frame extraction unit 410, it is configured to extract multiple views in video
Frequency frame.It is special image can be executed for the image information for multiple video frames that video frame extraction unit 410 extracts using device 400
The extraction of sign, and the characteristics of image extracted respectively according to above-mentioned multiple video frames is respectively as the fingerprint characteristic for corresponding to video frame.
These video frames for being used to take the fingerprint feature are also referred to as key frame.
In some embodiments, video frame extraction unit 410 can be configured to be determined according to video and refer to for extracting video
The effective video segment of line, and key frame is extracted from effective video segment.It in some embodiments, can be by one of video
Divide and is determined as effective video segment.For example, the piece that video can be cut into the segment of predetermined time period, and cutting is obtained
Duan Zuowei effective video segment.In further embodiments, entire video can be determined as effective video segment.
Video frame extraction unit 410 may be configured to extract key frame from determining effective video segment.Some
In embodiment, each frame in effective video segment can be determined as key frame.It in further embodiments, can be by right
Each frame in effective video segment is analyzed, and realizes the shot segmentation to effective video segment, and according to shot segmentation
As a result the key frame for being used for each camera lens is determined.Sampling can be passed through in effective video segment in yet other embodiments,
Method selecting video frame is as key frame.For example, can be sampled in a manner of equally spaced to effective video segment, and will adopt
The video frame that sample obtains is as key frame.In another example can be sampled with arbitrary interval to effective video segment, and will sampling
Obtained video frame is as key frame.
Device 400 can also include Finger print characteristic abstract unit 420, can be configured to for the multiple video frame
Each, is handled the video frame using the neural network with multiple layers, wherein includes at least one volume in multiple layers
Lamination, each of at least one described convolutional layer is used for the output to its preceding layer and carries out process of convolution, by the nerve
Fingerprint characteristic of the convolution processing result of the middle layer output of network as the video frame.
In some embodiments, which can be the neural network obtained using the training of image classification task.This
The neural network of sample can be made of the one or more in convolutional layer, pond layer, full articulamentum and activation primitive.For example, this
The open neural network with multiple layers used is the depth constituted by stacking multiple convolutional layers and at least one pond layer
Neural network.For example, can use the structures such as Alexnet, Vggnet, Resnet, Googlenet, Mobilenet realizes this public affairs
Neural network used in the embodiment opened.In the example being described below, using Mobilenet network shown in table 1 as
Example describes the device for being used to extract video finger print of disclosure offer.
As previously mentioned, carrying out cluster point by the characteristics of image exported to each layer of neural network structure shown in table 1
Analysis determines the characteristics of image for the output for selecting which layer as the fingerprint characteristic for being used for the key frame.That is, can pass through
To the image of middle layer (one layer of layer, layer 5 reciprocal or any other centre such as second from the bottom) output different in neural network
Feature executes clustering algorithm (such as K mean cluster, mean shift clustering, density clustering method) to examine neural network
The image clustering effect that the characteristics of image of middle layer output is realized.As previously mentioned, for illustrative nerve net shown in table 1
Network structure, such as can determine that the Clustering Effect of the characteristics of image of its layer 5 output reciprocal is best, therefore, as shown in table 1,
The characteristics of image that this layer can be exported is as the fingerprint characteristic of processed key frame.
It will be understood by those skilled in the art that when the characteristics of image for selecting other neural network structures to be used to extract key frame
When, it can be according to the finger of a middle layer of above method neural network selected to use exported as processed key frame
Line feature.Those skilled in the art can select according to the actual situation or the image of middle layer output that Clustering Effect is best is special
Levy the fingerprint characteristic as processed key frame.In addition, those skilled in the art can also use field according to video finger print
Scape is examined by other standards and selects the characteristics of image of any middle layer output in neural network structure as processed pass
The fingerprint characteristic of key frame.
In some embodiments, Finger print characteristic abstract unit 420 may be configured to singly referring to using processor (such as CPU)
Multiple data stream (SIMD) characteristic is enabled to accelerate the speed of service of depth convolution.When needing to be implemented convolution operation, due to processed
Characteristics of image different channels on data execute convolution operation instruction be it is identical, therefore, can use SIMD characteristic,
Each instruction is enabled to execute operation to the data parallel on multiple channels simultaneously, so as to improve the handling capacity of processor.In addition
When the size of convolution kernel is greater than 1, convolution kernel can overlap when sliding on characteristic pattern (feature map), handle
The data for retaining lap in the register of device can reduce internal storage access number, to improve computational efficiency.It can be used
The SIMD characteristic of assembly instruction calling processor.
Optionally, before carrying out image characteristics extraction to key frame using neural network, Finger print characteristic abstract unit 420
It may be configured to the execution size change over (such as up-sampling or down-sampling) to key frame, so that the size of key frame accords with
Close the input size of neural network.
Device 400 can also include video finger print generation unit 330, can be configured to handle the multiple video frame
Fingerprint characteristic, to generate the video finger print of the video.It in some embodiments, can be by merging the multiple video frame
Fingerprint characteristic determines the video finger print of the video.That is, including the fingerprint characteristic of all key frames in video finger print
Information.
As shown in figure 4, video finger print generation unit 430 may include dimensionality reduction subelement 431 and splicing subelement 432.Its
In, dimensionality reduction subelement 431 can be configured to execute dimension-reduction treatment respectively to the fingerprint characteristic of the multiple video frame, more to obtain
The fingerprint characteristic of a dimensionality reduction.
In some embodiments, dimension-reduction treatment may include pond step and principal component analysis step.For example, can to institute
The fingerprint characteristic for stating multiple video frames executes pondization processing respectively, to obtain the fingerprint characteristic in multiple ponds.For example, pondization is handled
It can be one of average pond, maximum pond or minimum pond or a variety of.As shown in table 1, layer 5 reciprocal is defeated
The size of fingerprint characteristic out is 7 × 7 × 1024.Then, the fingerprint characteristic dimensionality reduction that this layer can be exported is handled by pondization
To 1 × 1 × 1024 size.
Principal component analysis step may include utilizing the finger for being used for the projection matrix of principal component analysis to the multiple pond
Line feature carries out principal component analysis, to obtain the fingerprint characteristic of multiple dimensionality reductions.In some embodiments, above-mentioned principal component analysis
Projection matrix is determined using preset image data set.For example, can use the d dimension including a plurality of types of image informations
Image data set calculates the covariance matrix for being used for the data set.It is then possible to calculate the characteristic value and phase of the covariance matrix
The feature vector answered, and therefrom select feature vector corresponding with preceding k maximum eigenvalue.Wherein k is less than the integer of d.It is logical
A projection matrix can be constructed corresponding to k feature vector of preceding k maximum eigenvalue by crossing.It can be with using the projection matrix
The d fingerprint characteristic dimensionality reduction tieed up is tieed up at k.It is understood that those skilled in the art can be arranged according to the actual situation k's
Value, to obtain the fingerprint characteristic of various sizes of dimensionality reduction.
Splicing subelement 432 can be configured to the video finger print feature for splicing the multiple dimensionality reduction, to generate the video
Video finger print.For example, when the crucial number of frames of video is 3, it can be by the finger of the dimensionality reduction generated respectively for 3 key frames
Line feature is spliced.By taking the fingerprint characteristic obtained after dimensionality reduction is 128 dimensions as an example, spliced available one 384 dimension
Feature.The spliced feature can be used as the video finger print of the video.
Alternatively, splicing subelement 432 also can be configured to execute change to the video finger print feature of the multiple dimensionality reduction
It changes to generate the video finger print of the video.For example, can video finger print feature to the multiple dimensionality reduction execute feature Hash
It converts, and the result of feature Hash is determined as to the video finger print of the video.
The device for being used to extract video finger print provided using the disclosure, can be by only to the partial video frame in video
The operation for executing the feature that takes the fingerprint, improves the formation speed of video finger print.In addition, by examining each layer output of neural network
The Clustering Effect of characteristics of image, the output that can choose the middle layer with best Clustering Effect are used to generate video finger print,
So that the video finger print generated has better effect during identifying similar video.Further, by mind
Output through network executes principal component analysis, can compare successful accuracy rate guaranteeing the video of video finger print and Similar content
While, the video decoded time is removed, the time CPU monokaryon that entire video generates fingerprint only needs 160 milliseconds.By to generation
The dimension of video finger print feature carry out dimensionality reduction so that the versus speed of fingerprint has obtained fast lifting between video, therefore can be with
Support the video search application based on video content.
In addition, can also be by means of the framework shown in fig. 5 for calculating equipment according to the method or apparatus of the embodiment of the present disclosure
To realize.Fig. 5 shows the framework of the calculating equipment.As shown in figure 5, calculate equipment 500 may include bus 510, one or
Multiple CPU 520, read-only memory (ROM) 530, random access memory (RAM) 540, the communication port for being connected to network
550, input output assembly 560, hard disk 570 etc..The storage equipment in equipment 500 is calculated, such as ROM 530 or hard disk 570 can
With store processing and/or the various data that use of communication or the file of the method for disclosure offer extracted for video finger print with
And program instruction performed by CPU.Calculating equipment 500 can also include user interface 580.Certainly, framework shown in fig. 5 is
Illustratively, when realizing different equipment, according to actual needs, it is convenient to omit one in calculating equipment shown in Fig. 5 or
Multiple components.
Embodiment of the disclosure also may be implemented as computer readable storage medium.According to the calculating of the embodiment of the present disclosure
Computer-readable instruction is stored on machine readable storage medium storing program for executing.It, can be with when the computer-readable instruction is run by processor
Execute the method according to the embodiment of the present disclosure referring to the figures above description.The computer readable storage medium includes but unlimited
In such as volatile memory and/or nonvolatile memory.The volatile memory for example may include that arbitrary access is deposited
Reservoir (RAM) and/or cache memory (cache) etc..The nonvolatile memory for example may include read-only storage
Device (ROM), hard disk, flash memory etc..
It will be appreciated by those skilled in the art that a variety of variations and modifications can occur in content disclosed by the disclosure.For example,
Various equipment described above or component can also pass through one in software, firmware or three by hardware realization
A little or whole combinations is realized.
In addition, as shown in the disclosure and claims, unless context clearly prompts exceptional situation, " one ", " one
It is a ", the words such as "an" and/or "the" not refer in particular to odd number, may also comprise plural number.It is, in general, that term " includes " and "comprising"
Only prompt included the steps that clearly identified and element, and these steps and element do not constitute one it is exclusive enumerate, method
Or equipment the step of may also including other or element.
In addition, although the disclosure is made that various references to certain units in system according to an embodiment of the present disclosure,
However, any amount of different units can be used and be operated on client and/or server.The unit is only explanation
Property, and different units can be used in the different aspect of the system and method.
In addition, flow chart has been used to be used to illustrate behaviour performed by system according to an embodiment of the present disclosure in the disclosure
Make.It should be understood that front or following operate not necessarily accurately carry out in sequence.On the contrary, can according to inverted order or
Various steps are handled simultaneously.It is also possible to during other operations are added to these, or from these processes remove a certain step
Or number step operation.
Unless otherwise defined, all terms (including technical and scientific term) used herein have leads with belonging to the present invention
The identical meanings that the those of ordinary skill in domain is commonly understood by.It is also understood that those of definition term such as in usual dictionary
The meaning consistent with their meanings in the context of the relevant technologies should be interpreted as having, without application idealization or
The meaning of extremely formalization explains, unless being clearly defined herein.
The above is the description of the invention, and is not considered as limitation ot it.Notwithstanding of the invention several
Exemplary embodiment, but those skilled in the art will readily appreciate that, before without departing substantially from teaching and advantage of the invention
Many modifications can be carried out to exemplary embodiment by putting.Therefore, all such modifications are intended to be included in claims institute
In the scope of the invention of restriction.It should be appreciated that being the description of the invention above, and it should not be considered limited to disclosed spy
Determine embodiment, and the model in the appended claims is intended to encompass to the modification of the disclosed embodiments and other embodiments
In enclosing.The present invention is limited by claims and its equivalent.
Claims (14)
1. a kind of method for extracting the video finger print of video, comprising:
Extract multiple video frames in video；
For each of the multiple video frame, the video frame is handled using the neural network with multiple layers,
In, it include at least one convolutional layer in multiple layers, each of at least one described convolutional layer is used for the defeated of its preceding layer
Process of convolution is carried out out；
The convolution processing result that the middle layer of the neural network is exported is as the fingerprint characteristic of the video frame；And
The fingerprint characteristic of the multiple video frame is handled, to generate the video finger print of the video.
2. according to the method described in claim 1, wherein, the fingerprint characteristic of the multiple video frame is handled, to generate the view
The video finger print of frequency includes:
Dimension-reduction treatment is executed respectively to the fingerprint characteristic of the multiple video frame, to obtain the fingerprint characteristic of multiple dimensionality reductions；
Splice the video finger print feature of the multiple dimensionality reduction, to generate the video finger print of the video.
3. according to the method described in claim 2, wherein, executing dimension-reduction treatment to the fingerprint characteristic of the multiple video frame includes
At least one of following steps:
Pondization processing is executed respectively to the fingerprint characteristic of the multiple video frame, to obtain the fingerprint characteristic in multiple ponds, and
Principal component analysis is carried out to the fingerprint characteristic in the multiple pond, to obtain the fingerprint characteristic of multiple dimensionality reductions.
4. method according to claim 1-3, wherein extract video in multiple video frames include:
Select multiple frames as the multiple video frame in the video equal intervals.
5. method according to claim 1-4, wherein the neural network is by the video frame transformation at having
The image data in multiple channels, the method also includes the image datas concurrently to the multiple channel to execute convolution operation.
6. method according to claim 1-5, wherein the neural network is Mobilenet network.
7. a kind of for extracting the device of the video finger print of video, comprising:
Video frame extraction unit is configured to extract multiple video frames in video；
Finger print characteristic abstract unit, be configured to for the multiple video frame each, utilize the nerve net with multiple layers
Network handles the video frame, wherein and it include at least one convolutional layer in multiple layers, it is every at least one described convolutional layer
One carries out process of convolution for the output to its preceding layer, the convolution processing result that the middle layer of the neural network is exported
Fingerprint characteristic as the video frame；And
Video finger print generation unit is configured to handle the fingerprint characteristic of the multiple video frame, to generate the video of the video
Fingerprint.
8. device according to claim 7, wherein the video finger print generation unit further include:
Dimensionality reduction subelement, configuration executes dimension-reduction treatment to the fingerprint characteristic of the multiple video frame respectively, to obtain multiple dimensionality reductions
Fingerprint characteristic；And
Splice subelement, is configured to splice the video finger print feature of the multiple dimensionality reduction, to generate the video finger print of the video.
9. device according to claim 8, wherein the dimension-reduction treatment at least one of includes the following steps:
Pondization processing is executed respectively to the fingerprint characteristic of the multiple video frame, to obtain the fingerprint characteristic in multiple ponds, and
Principal component analysis is carried out to the fingerprint characteristic in the multiple pond, to obtain the fingerprint characteristic of multiple dimensionality reductions.
10. according to the described in any item devices of claim 7-9, wherein extract video in multiple video frames include:
Select multiple frames as the multiple video frame in the video equal intervals.
11. according to the described in any item devices of claim 7-10, wherein the neural network is by the video frame transformation at tool
There is the image data in multiple channels, described device is configured to concurrently execute convolution behaviour to the image data in the multiple channel
Make.
12. according to the described in any item devices of claim 7-11, wherein the neural network is Mobilenet network.
13. a kind of equipment for extracting video finger print, including processor and memory, program is stored in the memory and is referred to
It enables, when executing described program instruction by processor, the processor is configured to execute any one of -6 institute according to claim 1
The method for the video finger print for extracting video stated.
14. a kind of computer readable storage medium is stored thereon with instruction, described instruction is when being executed by processor, so that institute
It states processor and executes the method according to claim 1-6 for extracting the video finger print of video.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811353102.6A CN110163061B (en) | 2018-11-14 | 2018-11-14 | Method, apparatus, device and computer readable medium for extracting video fingerprint |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811353102.6A CN110163061B (en) | 2018-11-14 | 2018-11-14 | Method, apparatus, device and computer readable medium for extracting video fingerprint |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110163061A true CN110163061A (en) | 2019-08-23 |
| CN110163061B CN110163061B (en) | 2023-04-07 |
Family
ID=67645250
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811353102.6A Active CN110163061B (en) | 2018-11-14 | 2018-11-14 | Method, apparatus, device and computer readable medium for extracting video fingerprint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163061B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781818A (en) * | 2019-10-25 | 2020-02-11 | Oppo广东移动通信有限公司 | Video classification method, model training method, device and equipment |
| CN110830938A (en) * | 2019-08-27 | 2020-02-21 | 武汉大学 | Fingerprint positioning quick implementation method for indoor signal source deployment scheme screening |
| CN110996124A (en) * | 2019-12-20 | 2020-04-10 | 北京百度网讯科技有限公司 | Original video determination method and related equipment |
| CN111614991A (en) * | 2020-05-09 | 2020-09-01 | 咪咕文化科技有限公司 | Video progress determination method and device, electronic equipment and storage medium |
| CN113313065A (en) * | 2021-06-23 | 2021-08-27 | 北京奇艺世纪科技有限公司 | Video processing method and device, electronic equipment and readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008395A (en) * | 2014-05-20 | 2014-08-27 | 中国科学技术大学 | Intelligent bad video detection method based on face retrieval |
| CN106686472A (en) * | 2016-12-29 | 2017-05-17 | 华中科技大学 | High-frame-rate video generation method and system based on depth learning |
| CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
| CN107665261A (en) * | 2017-10-25 | 2018-02-06 | 北京奇虎科技有限公司 | Video duplicate checking method and device |
| CN108198202A (en) * | 2018-01-23 | 2018-06-22 | 北京易智能科技有限公司 | A kind of video content detection method based on light stream and neural network |
| CN108280233A (en) * | 2018-02-26 | 2018-07-13 | 南京邮电大学 | A kind of VideoGIS data retrieval method based on deep learning |
| CN108431823A (en) * | 2015-11-05 | 2018-08-21 | 脸谱公司 | With the system and method for convolutional neural networks process content |
-
2018
- 2018-11-14 CN CN201811353102.6A patent/CN110163061B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008395A (en) * | 2014-05-20 | 2014-08-27 | 中国科学技术大学 | Intelligent bad video detection method based on face retrieval |
| CN108431823A (en) * | 2015-11-05 | 2018-08-21 | 脸谱公司 | With the system and method for convolutional neural networks process content |
| CN106686472A (en) * | 2016-12-29 | 2017-05-17 | 华中科技大学 | High-frame-rate video generation method and system based on depth learning |
| CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
| CN107665261A (en) * | 2017-10-25 | 2018-02-06 | 北京奇虎科技有限公司 | Video duplicate checking method and device |
| CN108198202A (en) * | 2018-01-23 | 2018-06-22 | 北京易智能科技有限公司 | A kind of video content detection method based on light stream and neural network |
| CN108280233A (en) * | 2018-02-26 | 2018-07-13 | 南京邮电大学 | A kind of VideoGIS data retrieval method based on deep learning |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110830938A (en) * | 2019-08-27 | 2020-02-21 | 武汉大学 | Fingerprint positioning quick implementation method for indoor signal source deployment scheme screening |
| CN110781818A (en) * | 2019-10-25 | 2020-02-11 | Oppo广东移动通信有限公司 | Video classification method, model training method, device and equipment |
| CN110781818B (en) * | 2019-10-25 | 2023-04-07 | Oppo广东移动通信有限公司 | Video classification method, model training method, device and equipment |
| CN110996124A (en) * | 2019-12-20 | 2020-04-10 | 北京百度网讯科技有限公司 | Original video determination method and related equipment |
| CN110996124B (en) * | 2019-12-20 | 2022-02-08 | 北京百度网讯科技有限公司 | Original video determination method and related equipment |
| CN111614991A (en) * | 2020-05-09 | 2020-09-01 | 咪咕文化科技有限公司 | Video progress determination method and device, electronic equipment and storage medium |
| CN111614991B (en) * | 2020-05-09 | 2022-11-22 | 咪咕文化科技有限公司 | Video progress determination method and device, electronic equipment and storage medium |
| CN113313065A (en) * | 2021-06-23 | 2021-08-27 | 北京奇艺世纪科技有限公司 | Video processing method and device, electronic equipment and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110163061B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110163061A (en) | For extracting the method, apparatus, equipment and computer-readable medium of video finger print | |
| CN103927387B (en) | Image indexing system and its correlation technique and device | |
| CN110889855B (en) | Certificate photo matting method and system based on end-to-end convolution neural network | |
| CN109117781B (en) | Multi-attribute identification model establishing method and device and multi-attribute identification method | |
| Kucer et al. | Leveraging expert feature knowledge for predicting image aesthetics | |
| Bianco et al. | Predicting image aesthetics with deep learning | |
| CN109413510B (en) | Video abstract generation method and device, electronic equipment and computer storage medium | |
| Tan et al. | Selective dependency aggregation for action classification | |
| Zhu et al. | Depth2action: Exploring embedded depth for large-scale action recognition | |
| CN104346630B (en) | A kind of cloud flowers recognition methods of heterogeneous characteristic fusion | |
| CN108198177A (en) | Image acquiring method, device, terminal and storage medium | |
| Douze et al. | The 2021 image similarity dataset and challenge | |
| TWI719512B (en) | Method and system for algorithm using pixel-channel shuffle convolution neural network | |
| CN110958469A (en) | Video processing method and device, electronic equipment and storage medium | |
| CN108492160A (en) | Information recommendation method and device | |
| CN111985281A (en) | Image generation model generation method and device and image generation method and device | |
| CN112132279A (en) | Convolutional neural network model compression method, device, equipment and storage medium | |
| CN108875751A (en) | Image processing method and device, the training method of neural network, storage medium | |
| CN113673613A (en) | Multi-modal data feature expression method, device and medium based on contrast learning | |
| Mejjati et al. | Look here! a parametric learning based approach to redirect visual attention | |
| CN115687670A (en) | Image searching method and device, computer readable storage medium and electronic equipment | |
| Luo et al. | Image steganalysis with convolutional vision transformer | |
| CN116206334A (en) | Wild animal identification method and device | |
| CN117372782A (en) | Small sample image classification method based on frequency domain analysis | |
| CN109697240A (en) | A kind of image search method and device based on feature |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |