CN109978067A - A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform - Google Patents
A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform Download PDFInfo
- Publication number
- CN109978067A CN109978067A CN201910259887.9A CN201910259887A CN109978067A CN 109978067 A CN109978067 A CN 109978067A CN 201910259887 A CN201910259887 A CN 201910259887A CN 109978067 A CN109978067 A CN 109978067A
- Authority
- CN
- China
- Prior art keywords
- trademark image
- feature
- image
- convolutional neural
- neural networks
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biodiversity & Conservation Biology (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The disclosure is directed to a kind of trade-mark searching method based on convolutional neural networks and Scale invariant features transform, device, electronic equipment and storage mediums.Wherein, this method comprises: crawling the other trademark image of universal class of the first preset quantity on network by crawler, trademark image database is established;The other trademark image of universal class in the trademark image database is subjected to training, verifying and test based on convolutional neural networks respectively, extracts convolutional neural networks feature;It determines key point and the point of interest direction of trademark image, and scale invariant feature is extracted by trademark image descriptor computation;Convolutional neural networks feature and scale invariant feature head and the tail are spliced, the fusion feature of trademark image is obtained;The Euclidean distance of the feature vector of trademark image judges similarity in fusion feature and the trademark image database by calculating the trademark image.The disclosure realizes the more acurrate identification of trade mark in such a way that a variety of trademark image Fusion Features compare.
Description
Technical field
This disclosure relates to which field of computer technology, is based on convolutional neural networks and Scale invariant in particular to one kind
Trade-mark searching method, device, electronic equipment and the computer readable storage medium of eigentransformation.
Background technique
Trade mark is the important component of intangible assets in enterprises, closely bound up with brand names, and trade mark is the mark in brand
Will and name portion identify that product and brand provide conveniently for consumer, and one unusual means a great for brand
And the trade mark convenient for identification has important role to enterprise.But China is big there are trade mark radix in terms of trade mark and exists
The problems such as a large amount of invalid applications, trade mark are supervised there are loophole, and mountain vallage trade mark is prevailing.Therefore, for how quickly and accurately from quotient
The information that similar brand is returned in mark library is a problem to be solved.
In recent years, trade mark retrieval usually utilizes shape or color characteristic, is such as accurately matched using scale invariant feature,
Similarity measure is completed by Euclidean distance, returns to similarity trade mark result in the top.But for more complex trade mark, by
Shape information cannot be fully described in scale invariant feature, so the retrieval effectiveness of the technology can't be satisfactory, makes its inspection
Suo Jingdu has to be optimized.
From the above, it can be seen that, it is desirable to provide one or more technical solutions for being at least able to solve the above problem.
It should be noted that information is only used for reinforcing the reason to the background of the disclosure disclosed in above-mentioned background technology part
Solution, therefore may include the information not constituted to the prior art known to persons of ordinary skill in the art.
Summary of the invention
Being designed to provide for the disclosure is a kind of based on the retrieval of the trade mark of convolutional neural networks and Scale invariant features transform
Method, apparatus, electronic equipment and computer readable storage medium, and then overcome at least to a certain extent due to the relevant technologies
Limitation and defect caused by one or more problem.
According to one aspect of the disclosure, a kind of trade mark based on convolutional neural networks and Scale invariant features transform is provided
Search method, comprising:
Trademark image database step is established, crawls the other trademark image of universal class of the first preset quantity on network by crawler
Picture establishes trademark image database;
Convolutional neural networks characteristic extraction step, the universal class of the first preset quantity in the trademark image database is other
Trademark image cutting is training set, verifying collection and test set, carries out training, verifying and survey based on convolutional neural networks respectively
Examination, extracts convolutional neural networks feature;
Scale invariant feature extraction step, determines key point and the point of interest direction of trademark image, and passes through trademark image
Descriptor computation extracts scale invariant feature;
Convolutional neural networks feature and scale invariant feature head and the tail are spliced, obtain trademark image by fusion feature generation step
The fusion feature of picture;
Fusion feature compares step, in the fusion feature and the trademark image database by calculating the trademark image
The Euclidean distance of the feature vector of trademark image judges similarity.
In a kind of exemplary embodiment of the disclosure, the convolutional neural networks characteristic extraction step further include:
It is the second present count by the other trademark image cutting of the universal class of the first preset quantity in the trademark image database
The test set of the training set of trademark image, the verifying collection of third preset quantity trademark image, the 4th preset quantity trademark image is measured,
The sum of second preset quantity, third preset quantity and the 4th preset quantity are the first preset quantity.
In a kind of exemplary embodiment of the disclosure, the method also includes:
Second preset quantity trademark image progress data are enhanced to obtain the 5th preset quantity trademark image training set:
The data enhancing includes five kinds of modes: horizontal mirror image being used to erect image left-half and right half part with image
Straight central axes are that center axis is exchanged;Cut out image in a certain range at random in original image;Image level position translation;Image
Upper and lower position translation;Image random small angle rotation in [0,15] degree range;
5th preset quantity is 5 times of the second preset quantity.
In a kind of exemplary embodiment of the disclosure, the convolutional neural networks characteristic extraction step further include:
It is special that convolutional neural networks are extracted by the depth residual error network based on ResNet32 and the activation primitive based on PReLU
Sign;
The depth residual error network includes 5 layers of convolutional layer, 4 layers of residual error layer, 1 layer of full articulamentum and SoftmaxWithLoss
Classifier layer;
The activation primitive coefficient value based on PReLU is 0.25.
In a kind of exemplary embodiment of the disclosure, the scale invariant feature extraction step further include:
Trademark image scale space and critical point detection are used by obtaining sigma value using Gaussian operator
The LOG filtering image of sigma value positions key point for creating scale space, by pixel intensity value.
Calculate key vertex neighborhood interior derivative amplitude F (x, y) and direction θ (x, y):
Wherein, L is gray level
Distribution weight by gradient magnitude as pixel orientation, the highest pixel orientation of weight are point of interest direction;
Trademark image descriptor is calculated by gradient direction, extracts scale invariant feature.
It is described that trademark image descriptor step is calculated by gradient direction in a kind of exemplary embodiment of the disclosure
Further include:
Choose 16 × 16 window of the trademark image centered on key point;
By the pixel in window according to Gaussian weighting be divided into 4 × 4 region;
The relative direction of pixel and key point is subjected to Gaussian weighting and is put into the direction histogram containing 8 pixels,
Obtain scale invariant feature.
In one aspect of the present disclosure, it provides a kind of based on the inspection of the trade mark of convolutional neural networks and Scale invariant features transform
Rope device, comprising:
Trademark image database module is established, for crawling the other quotient of universal class of the first preset quantity on network by crawler
Logo image establishes trademark image database;
Convolutional neural networks characteristic extracting module, for by the complete of the first preset quantity in the trademark image database
Classification trademark image cutting is training set, verifying collection and test set, carry out respectively training based on convolutional neural networks, verifying and
Test, extracts convolutional neural networks feature;
Scale invariant feature extraction module for determining key point and the point of interest direction of trademark image, and passes through trade mark
Image descriptor calculating extracts scale invariant feature;
Fusion feature generation module obtains quotient for splicing convolutional neural networks feature and scale invariant feature head and the tail
The fusion feature of logo image;
Fusion feature contrast module, for passing through the fusion feature for calculating the trademark image and the trademark image data
The Euclidean distance of the feature vector of trademark image judges similarity in library.
In one aspect of the present disclosure, a kind of electronic equipment is provided, comprising:
Processor;And
Memory is stored with computer-readable instruction on the memory, and the computer-readable instruction is by the processing
The method according to above-mentioned any one is realized when device executes.
In one aspect of the present disclosure, a kind of computer readable storage medium is provided, computer program is stored thereon with, institute
State realization method according to above-mentioned any one when computer program is executed by processor.
Being retrieved based on the trade mark of convolutional neural networks and Scale invariant features transform in the exemplary embodiment of the disclosure
Method is crawled the other trademark image of universal class of the first preset quantity on network by crawler, establishes trademark image database;By institute
It states the other trademark image of universal class in trademark image database and carries out training, verifying and test based on convolutional neural networks respectively, mention
Take out convolutional neural networks feature;It determines key point and the point of interest direction of trademark image, and passes through trademark image descriptor meter
Calculation extracts scale invariant feature;Convolutional neural networks feature and scale invariant feature head and the tail are spliced, trademark image is obtained
Fusion feature;In fusion feature and the trademark image database by calculating the trademark image feature of trademark image to
The Euclidean distance of amount judges similarity.On the one hand, due to using based on convolutional neural networks and Scale invariant features transform knot
The trademark image identification method of conjunction, improves the accuracy of identification;On the other hand, depth residual error is used based on convolutional neural networks
Network greatly reduces the calculation amount of framework, accelerates the models fitting time, more accurately expresses the feature of vector data, thus
Improve category of model accurate rate.
It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not
The disclosure can be limited.
Detailed description of the invention
Its example embodiment is described in detail by referring to accompanying drawing, the above and other feature and advantage of the disclosure will become
It is more obvious.
Fig. 1 is shown according to one exemplary embodiment of the disclosure based on convolutional neural networks and Scale invariant features transform
Trade-mark searching method flow chart;
Fig. 2 shows according to one exemplary embodiment of the disclosure based on convolutional neural networks and Scale invariant features transform
Trade mark retrieval device schematic block diagram;
Fig. 3 diagrammatically illustrates the block diagram of the electronic equipment according to one exemplary embodiment of the disclosure;And
Fig. 4 diagrammatically illustrates the schematic diagram of the computer readable storage medium according to one exemplary embodiment of the disclosure.
Specific embodiment
Example embodiment is described more fully with reference to the drawings.However, example embodiment can be real in a variety of forms
It applies, and is not understood as limited to embodiment set forth herein;On the contrary, thesing embodiments are provided so that the disclosure will be comprehensively and complete
It is whole, and the design of example embodiment is comprehensively communicated to those skilled in the art.Identical appended drawing reference indicates in figure
Same or similar part, thus repetition thereof will be omitted.
In addition, described feature, structure or characteristic can be incorporated in one or more implementations in any suitable manner
In example.In the following description, many details are provided to provide and fully understand to embodiment of the disclosure.However,
It will be appreciated by persons skilled in the art that can be with technical solution of the disclosure without one in the specific detail or more
It is more, or can be using other methods, constituent element, material, device, step etc..In other cases, it is not shown in detail or describes
Known features, method, apparatus, realization, material or operation are to avoid fuzzy all aspects of this disclosure.
Block diagram shown in the drawings is only functional entity, not necessarily must be corresponding with physically separate entity.
I.e., it is possible to realize these functional entitys using software form, or these are realized in the module of one or more softwares hardening
A part of functional entity or functional entity, or realized in heterogeneous networks and/or processor device and/or microcontroller device
These functional entitys.
In this exemplary embodiment, a kind of quotient based on convolutional neural networks and Scale invariant features transform is provided firstly
Search method is marked, can be applied to the electronic equipments such as control system of electric automobile;With reference to shown in Fig. 1, convolutional Neural should be based on
The trade-mark searching method of network and Scale invariant features transform may comprise steps of:
Trademark image database step S110 is established, crawls the other quotient of universal class of the first preset quantity on network by crawler
Logo image establishes trademark image database;
Convolutional neural networks characteristic extraction step S120, by the complete of the first preset quantity in the trademark image database
Classification trademark image cutting is training set, verifying collection and test set, carry out respectively training based on convolutional neural networks, verifying and
Test, extracts convolutional neural networks feature;
Scale invariant feature extraction step S130, determines key point and the point of interest direction of trademark image, and passes through trade mark
Image descriptor calculating extracts scale invariant feature;
Convolutional neural networks feature and scale invariant feature head and the tail are spliced, obtain quotient by fusion feature generation step S140
The fusion feature of logo image;
Fusion feature compares step S150, fusion feature and the trademark image data by the calculating trademark image
The Euclidean distance of the feature vector of trademark image judges similarity in library.
According to the trade-mark searching method based on convolutional neural networks and Scale invariant features transform in this example embodiment,
On the one hand, it due to using the trademark image identification method based on convolutional neural networks in conjunction with Scale invariant features transform, mentions
The high accuracy of identification;On the other hand, depth residual error network is used based on convolutional neural networks, greatly reduces the meter of framework
Calculation amount is accelerated the models fitting time, the feature of vector data is more accurately expressed, to improve category of model accurate rate.
In the following, by being retrieved based on the trade mark of convolutional neural networks and Scale invariant features transform in this example embodiment
Method is further detailed.
In establishing trademark image database step S110, the first preset quantity can be crawled on network by crawler
The other trademark image of universal class, establishes trademark image database.
In this exemplary embodiment, trademark image data are collected, crawl downloading trademark image from network.Collect one compared with
It is a wide range of to cover database of the trademark image library of each classification of clothing, food, lodging and transportion -- basic necessities of life as experiment comprising each industry of row, 210,000 are collected altogether
Width image.
In convolutional neural networks characteristic extraction step S120, first in the trademark image database can be preset
The other trademark image cutting of the universal class of quantity is training set, verifying collection and test set, carries out the instruction based on convolutional neural networks respectively
Practice, verify and test, extracts convolutional neural networks feature.
In this exemplary embodiment, the convolutional neural networks characteristic extraction step further include:
It is the second present count by the other trademark image cutting of the universal class of the first preset quantity in the trademark image database
The test set of the training set of trademark image, the verifying collection of third preset quantity trademark image, the 4th preset quantity trademark image is measured,
The sum of second preset quantity, third preset quantity and the 4th preset quantity are the first preset quantity.
In this exemplary embodiment, the method also includes:
Second preset quantity trademark image progress data are enhanced to obtain the 5th preset quantity trademark image training set:
The data enhancing includes five kinds of modes: horizontal mirror image being used to erect image left-half and right half part with image
Straight central axes are that center axis is exchanged;Cut out image in a certain range at random in original image;Image level position translation;Image
Upper and lower position translation;Image random small angle rotation in [0,15] degree range;
5th preset quantity is 5 times of the second preset quantity.
In this exemplary embodiment, the convolutional neural networks characteristic extraction step further include:
It is special that convolutional neural networks are extracted by the depth residual error network based on ResNet32 and the activation primitive based on PReLU
Sign;
The depth residual error network includes 5 layers of convolutional layer, 4 layers of residual error layer, 1 layer of full articulamentum and SoftmaxWithLoss
Classifier layer;
The activation primitive coefficient value based on PReLU is 0.25.
In this exemplary embodiment, 10,000 width images are chosen from 210,000 width images and carry out data enhancing conduct to it
ResNet training, verifying collection.Data enhancing use horizontal mirror image i.e. by image left-half and right half part with image it is vertical in
Axis is that center axis is exchanged;It cuts out at random and cuts out image in a certain range in original image;Horizontal position is appraised through discussion and up and down
Position translation;Image small angle rotation, the random rotation in [0,15] degree range.Such 10,000 width image enhances by data
For 50,000 width.
Cutting is carried out to 50,000 width images, cutting is training set, verifying collection, for following model training, verifying.Do not pass through
200,000 data sets of data enhancing are crossed as test set.It is 40,000 width, verifying collection number that three parts, which are followed successively by training set data, after cutting
According to for 10,000 width, test set data are 200,000 width.
It constructs and the training deep neural network ResNet model by the way of transfer learning, the depth residual error network architecture is
ResNet32, also comprising part normalization layer (Batch Norm and Scale), activation primitive layer (PReLU), residual error layer in network
(Eltwise), full articulamentum (Fc) and dropout layers (drop).
Depth residual error network using the image of 224 × 224 sizes as input, network include 5 layers of convolutional layer, 4 layers of residual error layer,
1 layer of full articulamentum and SoftmaxWithLoss classifier layer.After level 1 volume product, characteristic pattern size is 112 × 112,
Totally 64 characteristic patterns;After level 2 volume product, characteristic pattern size is 56 × 56, totally 64 characteristic patterns;By the 3rd layer of convolution
Later, characteristic pattern size is 28 × 28, totally 128 characteristic patterns;After the 4th layer of convolution, characteristic pattern size is 14 × 14, altogether
256 characteristic patterns;After the 5th layer of convolution, characteristic pattern size is 7 × 7, totally 512 characteristic patterns.The size of convolution kernel is 7
× 7 perhaps 3 × 3 step sizes be 1 or 2.512 × 1 dimensional feature of final output.
It is too deep in order to solve network, over-fitting is caused, using depth residual error network structure, it allows network as far as possible
Intensification, wherein ResNet proposes two kinds of mapping: one is identity mapping, what is referred to is exactly in the following figure " it is winding
Curve ", another residual mapping, refer to be exactly in addition to " winding curve " part, so last output
It is H (x)=F (x)+x.
Obviously F (x) is easier than H (x) acquires, extreme case of justing think, and (exports when this section of network does not extract feature to data
Data do not change compared to output data), then H (x)=x, if it is residual error network, as long as then learning the table of F (x)=0
Up to formula, and general network will then learn H (x)=x expression formula, and the expression formula of F (x)=0 is more held for computer
It is easy to handle.Therefore the residual error network can greatly reduce the calculation amount of framework, accelerate the models fitting time, more accurately expression is sweared
The feature for measuring data, to improve category of model accurate rate.
Here activation primitive uses PReLU (Parametric Rectified Linear Unit), as its name suggests: band
The ReLU of parameter.If coefficient a=0, it is ReLU that PReLU, which degenerates,;If a is fixed value (such as a=an of very little
0.01) it is Leaky ReLU (LReLU) that, then PReLU, which degenerates,.Here choosing a is 0.25.
In scale invariant feature extraction step S130, key point and the point of interest direction of trademark image can be determined, and
Scale invariant feature is extracted by trademark image descriptor computation.
In this exemplary embodiment, the scale invariant feature extraction step further include:
Trademark image scale space and critical point detection are used by obtaining sigma value using Gaussian operator
The LOG filtering image of sigma value positions key point for creating scale space, by pixel intensity value.
Calculate key vertex neighborhood interior derivative amplitude F (x, y) and direction θ (x, y):
Wherein, L is gray level
Distribution weight by gradient magnitude as pixel orientation, the highest pixel orientation of weight are point of interest direction;
Trademark image descriptor is calculated by gradient direction, extracts scale invariant feature.
It is described that trademark image descriptor step is calculated by gradient direction in this exemplary embodiment further include:
Choose 16 × 16 window of the trademark image centered on key point;
By the pixel in window according to Gaussian weighting be divided into 4 × 4 region;
The relative direction of pixel and key point is subjected to Gaussian weighting and is put into the direction histogram containing 8 pixels,
Obtain scale invariant feature.
In this exemplary embodiment, scale space and critical point detection are carried out first.By applying Gaussian operator
Canny or Laplacian technology, can be readily available correct sigma (σ) value carry out multiscale analysis, therefore this
In using the LOG filtering image of sigma value for creating scale space.Key point is positioned using 27 pixel intensity values, really
Determine whether certain point is key point.
Secondly, determining direction.After being stablized, the key point of noise jamming can be resisted, the another of rotational invariance is realized
One key feature is principal direction, so that feature is formed in principal direction, to overcome the otherness occurred due to rotation.Every
In a key point or point of interest, key vertex neighborhood interior derivative amplitude F (x, y) and direction θ (x, y) are calculated, is indicated are as follows:
In formula: L is gray level.8 receivers are created in desired neighborhood direction histogram, each receiver includes
10 continuous angles.Distribution weight by gradient magnitude as pixel orientation, the highest pixel orientation of weight are considered as
The direction of point-of-interest.
Finally, the trademark image descriptor of key point.After determining key point, then their feature descriptor is determined, to protect
The stability for holding feature calculates its descriptor using gradient direction.The key step that SIFT descriptor determines is as described below.
1) 16 × 16 window centered on key point is chosen.
2) pixel in window is divided into 4 × 4 region according to Gaussian weighting.
3) relative direction of pixel and key point is subjected to Gaussian weighting and is put into the direction Histogram containing 8 pixels
Figure, so that SIFT keeps rotational invariance.
4) the SIFT descriptor of 4 × 4 × 8=128 dimension is obtained by step 2 and 3.
In fusion feature generation step S140, convolutional neural networks feature and scale invariant feature head and the tail can be spelled
It connects, obtains the fusion feature of trademark image.
In this exemplary embodiment, the feature of ResNet model extraction and SIFT feature are subjected to ending splicing,
The characteristic dimension of ResNet model extraction is that the characteristic dimension that 512 × 1, SIFT is extracted is 128 × 1, is spliced to form 640 × 1 spies
Sign.
In fusion feature comparison step S150, the fusion feature and the trade mark of the calculating trademark image can be passed through
The Euclidean distance of the feature vector of trademark image judges similarity in image data base.
In this exemplary embodiment, to further increase retrieval precision, using residual error framework convolutional neural networks with
The mode that SIFT algorithm combines carries out the extraction of feature, has certain increase in such characteristic dimension, but in precision further
It improves, avoids missing inspection, false retrieval.
It should be noted that although describing each step of method in the disclosure in the accompanying drawings with particular order,
This does not require that or implies must execute these steps in this particular order, or have to carry out step shown in whole
Just it is able to achieve desired result.Additional or alternative, it is convenient to omit multiple steps are merged into a step and held by certain steps
Row, and/or a step is decomposed into execution of multiple steps etc..
In addition, in this exemplary embodiment, additionally providing a kind of based on convolutional neural networks and Scale invariant features transform
Trade mark retrieve device.Referring to shown in Fig. 2, device should be retrieved based on the trade mark of convolutional neural networks and Scale invariant features transform
200 may include: to establish trademark image database module 210, convolutional neural networks characteristic extracting module 220, Scale invariant spy
Levy extraction module 230, fusion feature generation module 240 and fusion feature contrast module 250.Wherein:
Trademark image database module 210 is established, for crawling the universal class of the first preset quantity on network by crawler
Other trademark image, establishes trademark image database;
Convolutional neural networks characteristic extracting module 220, for by the first preset quantity in the trademark image database
The other trademark image cutting of universal class be training set, verifying collection and test set, carry out the training based on convolutional neural networks respectively, test
Card and test, extract convolutional neural networks feature;
Scale invariant feature extraction module 230 for determining key point and the point of interest direction of trademark image, and passes through quotient
Logo image descriptor computation extracts scale invariant feature;
Fusion feature generation module 240 is obtained for splicing convolutional neural networks feature and scale invariant feature head and the tail
The fusion feature of trademark image;
Fusion feature contrast module 250, for the fusion feature and the trademark image by calculating the trademark image
The Euclidean distance of the feature vector of trademark image judges similarity in database.
Among the above respectively based on the specific thin of the trade mark of convolutional neural networks and Scale invariant features transform retrieval apparatus module
Section is described in detail in corresponding audio paragraph recognition methods, therefore details are not described herein again.
It should be noted that although being referred in the above detailed description based on convolutional neural networks and Scale invariant features transform
Trade mark retrieval device 200 several modules or unit, but it is this divide it is not enforceable.In fact, according to this public affairs
The feature and function of the embodiment opened, two or more above-described modules or unit can be in a module or list
It is embodied in member.Conversely, an above-described module or the feature and function of unit can be by multiple with further division
Module or unit embody.
In addition, in an exemplary embodiment of the disclosure, additionally providing a kind of electronic equipment that can be realized the above method.
Person of ordinary skill in the field it is understood that various aspects of the invention can be implemented as system, method or
Program product.Therefore, various aspects of the invention can be embodied in the following forms, it may be assumed that complete hardware embodiment, completely
Software implementation (including firmware, microcode etc.) or hardware and software in terms of combine embodiment, may be collectively referred to as here
Circuit, " module " or " system ".
The electronic equipment 300 of this embodiment according to the present invention is described referring to Fig. 3.The electronics that Fig. 3 is shown is set
Standby 300 be only an example, should not function to the embodiment of the present invention and use scope bring any restrictions.
As shown in figure 3, electronic equipment 300 is showed in the form of universal computing device.The component of electronic equipment 300 can wrap
It includes but is not limited to: at least one above-mentioned processing unit 310, at least one above-mentioned storage unit 320, the different system components of connection
The bus 330 of (including storage unit 320 and processing unit 310), display unit 340.
Wherein, the storage unit is stored with program code, and said program code can be held by the processing unit 310
Row, so that various according to the present invention described in the execution of the processing unit 310 above-mentioned " illustrative methods " part of this specification
The step of exemplary embodiment.For example, the processing unit 310 can execute step S110 as shown in fig. 1 to step
S150。
Storage unit 320 may include the readable medium of volatile memory cell form, such as Random Access Storage Unit
(RAM) 3201 and/or cache memory unit 3202, it can further include read-only memory unit (ROM) 3203.
Storage unit 320 can also include program/utility with one group of (at least one) program module 3205
3204, such program module 3205 includes but is not limited to: operating system, one or more application program, other program moulds
It may include the realization of network environment in block and program data, each of these examples or certain combination.
Bus 330 can be to indicate one of a few class bus structures or a variety of, including storage unit bus or storage
Cell controller, peripheral bus, graphics acceleration port, processing unit use any bus structures in a variety of bus structures
Local bus.
Electronic equipment 300 can also be with one or more external equipments 370 (such as keyboard, sensing equipment, bluetooth equipment
Deng) communication, can also be enabled a user to one or more equipment interact with the electronic equipment 300 communicate, and/or with make
Any equipment (such as the router, modulation /demodulation that the electronic equipment 300 can be communicated with one or more of the other calculating equipment
Device etc.) communication.This communication can be carried out by input/output (I/O) interface 350.Also, electronic equipment 300 can be with
By network adapter 360 and one or more network (such as local area network (LAN), wide area network (WAN) and/or public network,
Such as internet) communication.As shown, network adapter 360 is communicated by bus 330 with other modules of electronic equipment 300.
It should be understood that although not shown in the drawings, other hardware and/or software module can not used in conjunction with electronic equipment 300, including but not
Be limited to: microcode, device driver, redundant processing unit, external disk drive array, RAID system, tape drive and
Data backup storage system etc..
By the description of above embodiment, those skilled in the art is it can be readily appreciated that example embodiment described herein
It can also be realized in such a way that software is in conjunction with necessary hardware by software realization.Therefore, implemented according to the disclosure
The technical solution of example can be embodied in the form of software products, which can store in a non-volatile memories
In medium (can be CD-ROM, USB flash disk, mobile hard disk etc.) or on network, including some instructions are so that a calculating equipment (can
To be personal computer, server, terminal installation or network equipment etc.) it executes according to the method for the embodiment of the present disclosure.
In an exemplary embodiment of the disclosure, a kind of computer readable storage medium is additionally provided, energy is stored thereon with
Enough realize the program product of this specification above method.In some possible embodiments, various aspects of the invention can be with
It is embodied as a kind of form of program product comprising program code, it is described when described program product is run on the terminal device
Program code is for executing the terminal device described in above-mentioned " illustrative methods " part of this specification according to the present invention
The step of various exemplary embodiments.
Refering to what is shown in Fig. 4, the program product 400 for realizing the above method of embodiment according to the present invention is described,
It can using portable compact disc read only memory (CD-ROM) and including program code, and can in terminal device, such as
It is run on PC.However, program product of the invention is without being limited thereto, in this document, readable storage medium storing program for executing, which can be, appoints
What include or the tangible medium of storage program that the program can be commanded execution system, device or device use or and its
It is used in combination.
Described program product can be using any combination of one or more readable mediums.Readable medium can be readable letter
Number medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can be but be not limited to electricity, magnetic, optical, electromagnetic, infrared ray or
System, device or the device of semiconductor, or any above combination.The more specific example of readable storage medium storing program for executing is (non exhaustive
List) include: electrical connection with one or more conducting wires, portable disc, hard disk, random access memory (RAM), read-only
Memory (ROM), erasable programmable read only memory (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.
Computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal,
In carry 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.Readable signal medium can also be any readable Jie other than readable storage medium storing program for executing
Matter, the readable medium can send, propagate or transmit for by instruction execution system, device or device use or and its
The program of combined use.
The program code for including on readable medium can transmit with any suitable medium, including but not limited to wirelessly, have
Line, optical cable, RF etc. or above-mentioned any appropriate combination.
The program for executing operation of the present invention can be write with any combination of one or more programming languages
Code, described program design language include object oriented program language-Java, C++ etc., further include conventional
Procedural programming language-such as " C " language or similar programming language.Program code can be fully in user
It calculates and executes in equipment, partly executes on a user device, being executed as an independent software package, partially in user's calculating
Upper side point is executed on a remote computing or is executed in remote computing device or server completely.It is being related to far
Journey calculates in the situation of equipment, and remote computing device can pass through the network of any kind, including local area network (LAN) or wide area network
(WAN), it is connected to user calculating equipment, or, it may be connected to external computing device (such as utilize ISP
To be connected by internet).
In addition, above-mentioned attached drawing is only the schematic theory of processing included by method according to an exemplary embodiment of the present invention
It is bright, rather than limit purpose.It can be readily appreciated that the time that above-mentioned processing shown in the drawings did not indicated or limited these processing is suitable
Sequence.In addition, be also easy to understand, these processing, which can be, for example either synchronously or asynchronously to be executed in multiple modules.
Those skilled in the art after considering the specification and implementing the invention disclosed here, will readily occur to its of the disclosure
His embodiment.This application is intended to cover any variations, uses, or adaptations of the disclosure, these modifications, purposes or
Adaptive change follow the general principles of this disclosure and including the undocumented common knowledge in the art of the disclosure or
Conventional techniques.The description and examples are only to be considered as illustrative, and the true scope and spirit of the disclosure are by claim
It points out.
It should be understood that the present disclosure is not limited to the precise structures that have been described above and shown in the drawings, and
And various modifications and changes may be made without departing from the scope thereof.The scope of the present disclosure is only limited by the attached claims.
Claims (9)
1. a kind of trade-mark searching method based on convolutional neural networks and Scale invariant features transform, which is characterized in that the side
Method includes:
Trademark image database step is established, crawls the other trademark image of universal class of the first preset quantity on network by crawler,
Establish trademark image database;
Convolutional neural networks characteristic extraction step, by the other trade mark of the universal class of the first preset quantity in the trademark image database
Image cutting is training set, verifying collection and test set, carries out training, verifying and test based on convolutional neural networks respectively, mentions
Take out convolutional neural networks feature;
Scale invariant feature extraction step determines key point and the point of interest direction of trademark image, and is described by trademark image
Symbol calculates and extracts scale invariant feature;
Convolutional neural networks feature and scale invariant feature head and the tail are spliced, obtain trademark image by fusion feature generation step
Fusion feature;
Fusion feature compares step, trade mark in the fusion feature and the trademark image database by calculating the trademark image
The Euclidean distance of the feature vector of image judges similarity.
2. the method as described in claim 1, which is characterized in that the convolutional neural networks characteristic extraction step further include:
It is the second preset quantity quotient by the other trademark image cutting of the universal class of the first preset quantity in the trademark image database
The training set of logo image, the verifying collection of third preset quantity trademark image, the 4th preset quantity trademark image test set, it is described
The sum of second preset quantity, third preset quantity and the 4th preset quantity are the first preset quantity.
3. method according to claim 2, which is characterized in that the method also includes:
Second preset quantity trademark image progress data are enhanced to obtain the 5th preset quantity trademark image training set:
Data enhancing includes five kinds of modes: use horizontal mirror image by image left-half and right half part with image it is vertical in
Axis is that center axis is exchanged;Cut out image in a certain range at random in original image;Image level position translation;Above and below image
Position translation;Image random small angle rotation in [0,15] degree range;
5th preset quantity is 5 times of the second preset quantity.
4. the method as described in claim 1, which is characterized in that the convolutional neural networks characteristic extraction step further include:
Convolutional neural networks feature is extracted by the depth residual error network based on ResNet32 and the activation primitive based on PReLU;
The depth residual error network includes 5 layers of convolutional layer, 4 layers of residual error layer, 1 layer of full articulamentum and SoftmaxWithLoss classification
Device layer;
The activation primitive coefficient value based on PReLU is 0.25.
5. the method as described in claim 1, which is characterized in that the scale invariant feature extraction step further include:
Trademark image scale space and critical point detection, by obtaining sigma value using Gaussian operator, using sigma value
LOG filtering image for creating scale space, pass through pixel intensity value and position key point;
Calculate key vertex neighborhood interior derivative amplitude F (x, y) and direction θ (x, y):
Wherein, L is gray level
Distribution weight by gradient magnitude as pixel orientation, the highest pixel orientation of weight are point of interest direction;
Trademark image descriptor is calculated by gradient direction, extracts scale invariant feature.
6. the method as described in claim 1, which is characterized in that described to calculate trademark image descriptor step by gradient direction
Suddenly further include:
Choose 16 × 16 window of the trademark image centered on key point;
By the pixel in window according to Gaussian weighting be divided into 4 × 4 region;
The relative direction of pixel and key point is subjected to Gaussian weighting and is put into the direction histogram containing 8 pixels, is obtained
Scale invariant feature.
7. a kind of retrieve device based on the trade mark of convolutional neural networks and Scale invariant features transform, which is characterized in that the dress
It sets and includes:
Trademark image database module is established, for crawling the other trademark image of universal class of the first preset quantity on network by crawler
Picture establishes trademark image database;
Convolutional neural networks characteristic extracting module, for the universal class of the first preset quantity in the trademark image database is other
Trademark image cutting is training set, verifying collection and test set, carries out training, verifying and survey based on convolutional neural networks respectively
Examination, extracts convolutional neural networks feature;
Scale invariant feature extraction module for determining key point and the point of interest direction of trademark image, and passes through trademark image
Descriptor computation extracts scale invariant feature;
Fusion feature generation module obtains trademark image for splicing convolutional neural networks feature and scale invariant feature head and the tail
The fusion feature of picture;
Fusion feature contrast module, in the fusion feature and the trademark image database by calculating the trademark image
The Euclidean distance of the feature vector of trademark image judges similarity.
8. a kind of electronic equipment, which is characterized in that including
Processor;And
Memory is stored with computer-readable instruction on the memory, and the computer-readable instruction is held by the processor
Method according to any one of claim 1 to 6 is realized when row.
9. a kind of computer readable storage medium, is stored thereon with computer program, the computer program is executed by processor
Shi Shixian is according to claim 1 to any one of 6 the methods.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910259887.9A CN109978067A (en) | 2019-04-02 | 2019-04-02 | A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910259887.9A CN109978067A (en) | 2019-04-02 | 2019-04-02 | A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109978067A true CN109978067A (en) | 2019-07-05 |
Family
ID=67082279
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910259887.9A Pending CN109978067A (en) | 2019-04-02 | 2019-04-02 | A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109978067A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113051983A (en) * | 2019-12-28 | 2021-06-29 | 中移(成都)信息通信科技有限公司 | Method for training field crop disease recognition model and field crop disease recognition |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108108499A (en) * | 2018-02-07 | 2018-06-01 | 腾讯科技(深圳)有限公司 | Face retrieval method, apparatus, storage medium and equipment |
CN108763262A (en) * | 2018-04-03 | 2018-11-06 | 南昌奇眸科技有限公司 | A kind of brand logo search method |
CN108805216A (en) * | 2018-06-19 | 2018-11-13 | 合肥工业大学 | Face image processing process based on depth Fusion Features |
CN109035315A (en) * | 2018-08-28 | 2018-12-18 | 武汉大学 | Merge the remote sensing image registration method and system of SIFT feature and CNN feature |
WO2019043406A1 (en) * | 2017-08-31 | 2019-03-07 | Calipsa Limited | Anomaly detection from video data from surveillance cameras |
-
2019
- 2019-04-02 CN CN201910259887.9A patent/CN109978067A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019043406A1 (en) * | 2017-08-31 | 2019-03-07 | Calipsa Limited | Anomaly detection from video data from surveillance cameras |
CN108108499A (en) * | 2018-02-07 | 2018-06-01 | 腾讯科技(深圳)有限公司 | Face retrieval method, apparatus, storage medium and equipment |
CN108763262A (en) * | 2018-04-03 | 2018-11-06 | 南昌奇眸科技有限公司 | A kind of brand logo search method |
CN108805216A (en) * | 2018-06-19 | 2018-11-13 | 合肥工业大学 | Face image processing process based on depth Fusion Features |
CN109035315A (en) * | 2018-08-28 | 2018-12-18 | 武汉大学 | Merge the remote sensing image registration method and system of SIFT feature and CNN feature |
Non-Patent Citations (1)
Title |
---|
汪善义: "改进SIFT算法及其在医学CT图像检索与配准中的应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113051983A (en) * | 2019-12-28 | 2021-06-29 | 中移(成都)信息通信科技有限公司 | Method for training field crop disease recognition model and field crop disease recognition |
CN113051983B (en) * | 2019-12-28 | 2022-08-23 | 中移(成都)信息通信科技有限公司 | Method for training field crop disease recognition model and field crop disease recognition |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20220083801A1 (en) | Methods and apparatuses for vehicle appearance feature recognition, methods and apparatuses for vehicle retrieval, storage medium, and electronic devices | |
US20200004777A1 (en) | Image Retrieval with Deep Local Feature Descriptors and Attention-Based Keypoint Descriptors | |
US9349076B1 (en) | Template-based target object detection in an image | |
CN111680678B (en) | Target area identification method, device, equipment and readable storage medium | |
CN112052186B (en) | Target detection method, device, equipment and storage medium | |
CN107545241A (en) | Neural network model is trained and biopsy method, device and storage medium | |
CN111241989A (en) | Image recognition method and device and electronic equipment | |
CN111597884A (en) | Facial action unit identification method and device, electronic equipment and storage medium | |
CN108229324A (en) | Gesture method for tracing and device, electronic equipment, computer storage media | |
CN111488873B (en) | Character level scene text detection method and device based on weak supervision learning | |
CN110222780A (en) | Object detecting method, device, equipment and storage medium | |
CN110059794A (en) | Man-machine recognition methods and device, electronic equipment, storage medium | |
CN113239227B (en) | Image data structuring method, device, electronic equipment and computer readable medium | |
CN109918513A (en) | Image processing method, device, server and storage medium | |
CN116152863B (en) | Personnel information identification method and device, electronic equipment and storage medium | |
CN114093022A (en) | Activity detection device, activity detection system, and activity detection method | |
CN109711441A (en) | Image classification method, device, storage medium and electronic equipment | |
Islam et al. | Convolutional auto-encoder and independent component analysis based automatic place recognition for moving robot in invariant season condition | |
CN111444802A (en) | Face recognition method and device and intelligent terminal | |
CN109978067A (en) | A kind of trade-mark searching method and device based on convolutional neural networks and Scale invariant features transform | |
CN116152576B (en) | Image processing method, device, equipment and storage medium | |
CN110111299A (en) | Rust staining recognition methods and device | |
CN113487374A (en) | Block E-commerce platform transaction system based on 5G network | |
Kiaei et al. | Human face detection based on skin color with principal component analysis (PCA) algorithm | |
US20230057167A1 (en) | Motif-based image classification |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190705 |
|
RJ01 | Rejection of invention patent application after publication |