CN108875076A - A kind of quick trademark image retrieval method based on Attention mechanism and convolutional neural networks - Google Patents
A kind of quick trademark image retrieval method based on Attention mechanism and convolutional neural networks Download PDFInfo
- Publication number
- CN108875076A CN108875076A CN201810750096.1A CN201810750096A CN108875076A CN 108875076 A CN108875076 A CN 108875076A CN 201810750096 A CN201810750096 A CN 201810750096A CN 108875076 A CN108875076 A CN 108875076A
- Authority
- CN
- China
- Prior art keywords
- attention
- mac
- network
- feature
- model
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The quick trademark image retrieval method based on Attention mechanism and convolutional neural networks that the invention discloses a kind of, including Caffe deep learning Open Framework is built, open source VGG16 network model is trained；Based on the design of VGG16 network model include the Attention network of two layers of convolutional layer, and adds Attention network in trained VGG16 network model；The VGG16 network model for being added to Attention network is trained using the training set in FlickrLogos-32 data set；Attention-MAC trade mark Feature Selection Model is generated based on the trained VGG16 network model for being added to Attention network；Trademark image to be checked is retrieved based on Attention-MAC trade mark Feature Selection Model, and generates search result.The present invention avoids the parameter using full articulamentum redundancy, achievees the purpose that reduced model, improves the speed of training and retrieval, reduces false detection rate.
Description
Technical field
The present invention relates to field of computer technology, are based on Attention mechanism and convolutional Neural net more particularly to one kind
The quick trademark image retrieval method of network.
Background technique
In recent years, with the rapid development of economy, the registration amount of all kinds of trade marks is rising year by year.Therefore, trade mark retrieval pair
Have great importance in trade mark registration, management and protection.For trade mark registration side, can be found in time by trade mark retrieval
The obstacle of trade mark registration application, such as trade mark to be registered whether by other people rush to register, whether with other existing trade marks in certain block region
There are similar etc..For trademark office, by trade mark retrieval can intelligence obtain have quotient in trade mark to be registered and database
Target similarity degree reduces the workload using artificial contrast, improves work efficiency.
Traditional trade mark detection includes based on the artificial database images retrieval for marking and comparing, based on quotient with identifying system
Mark coding retrieval, the image retrieval based on image binaryzation feature and the retrieval based on keyword etc. of graphical element.Tradition
It to the mark of image and compares in method and needs to take a substantial amount of time and manpower, increase the cost of trademark image retrieval.With
Gradualling mature for convolutional neural networks, trademark image is retrieved using convolutional neural networks and classify as trade mark detect
With the new approaches of identifying system.Model is trained using the training image in data set first, adjusts each network layer
Weight is tested with the image in test set later, and the model for finally completing test is applied to system.Although such method
Intelligence degree is higher, but structure is complicated for traditional area convolutional neural networks, complete to connect layer parameter redundancy, cause training process compared with
Slowly, it needs to consume a lot of time and resources；And in such method, for data set, it is desirable that it marks classification and location information
Completely, however for trademark office, not fully, it is dfficult to apply to training for the image labeling in trademark database.
Summary of the invention
In view of this, the present invention provides a kind of quick trademark image based on Attention mechanism and convolutional neural networks
As search method, by combining Attention mechanism with convolutional neural networks, and remove full articulamentum on this basis,
Using the character representation of the output representative image of convolutional neural networks middle layer, a kind of quick trade mark detection method is provided.
In order to realize above-mentioned purpose of the invention, the present invention provides one kind to be based on Attention mechanism and convolutional Neural
The quick trademark image retrieval method of network, described method includes following steps：
S1, Caffe deep learning Open Framework is built, open source VGG16 network model is trained；
S2, the Attention network comprising two layers of convolutional layer is designed based on VGG16 network model, and trained
Attention network is added in VGG16 network model；
S3, using the training set in FlickrLogos-32 data set to the VGG16 network for being added to Attention network
Model is trained；
S4, Attention-MAC quotient is generated based on the trained VGG16 network model for being added to Attention network
Mark Feature Selection Model；
S5, trademark image to be checked is retrieved based on Attention-MAC trade mark Feature Selection Model, and generates inspection
Hitch fruit.
Preferably, the step S1 includes the following steps：
S1-1, Caffe deep learning Open Framework is built, VGG16 network model is carried out using ImageNet data set
Pre-training；
S1-2, the VGG16 network model that pre-training is obtained using the training set in FlickrLogos-32 data set into
The training of row transfer learning.
Preferably, the step S2 includes the following steps：
S2-1, the model parameter design based on VGG16 network model include the Attention network of two layers of convolutional layer；
S2-2, between the last layer output and full articulamentum of the pond layer of the trained VGG16 network model,
Add designed Attention network.
Preferably, the step S3 includes the following steps：
S3-1, the fixed network weight for being added to characteristic extraction part in the VGG16 network model of Attention network；
S3-2, using the training set in FlickrLogos-32 data set to the VGG16 net for being added to Attention network
Attention network in network model is trained.
Preferably, the step S4 includes the following steps：
S4-1, remove full articulamentum in the trained VGG16 network model for being added to Attention network；
S4-2, pond layer is added after Attention network, generate Attention-MAC trade mark Feature Selection Model.
Preferably, the step S5 includes the following steps：
S5-1, the characteristics of image table that Attention-MAC trade mark Feature Selection Model is exported by main method for feature analysis
Show carry out dimensionality reduction, obtains Attention-MAC feature vector；
S5-2, using the trademark image in FlickrLogos-32 data set as input picture, be sequentially inputted to
In Attention-MAC trade mark Feature Selection Model, the Attention-MAC feature vector of branding data collection image is generated, and
Construct the feature database of branding data collection image；
S5-3, using trademark image to be retrieved as input picture, be input to Attention-MAC trade mark Feature Selection Model
In, generate the Attention-MAC feature vector of trademark image to be retrieved；
S5-4, the Attention-MAC feature vector and FlickrLogos-32 data set for calculating trademark image to be retrieved
The cosine similarity of the Attention-MAC feature vector of image obtains branding data collection image based on the first of cosine similarity
Begin to sort；
S5-5, FlickrLogos-32 data images are reset by expanding query, obtains branding data collection figure
As finally sorting with the similarity of trademark image to be retrieved and reporting trade mark search result.
Preferably, the step S1-2 includes the following steps：
S1-2-1, the VGG16 network model that pre-training is obtained using the training set in FlickrLogos-32 data set
Network weight be finely adjusted；
S1-2-2, in the training of transfer learning, pre-training is obtained using standard cross entropy loss function VGG16 net
Network model carries out classification based training.
Preferably, the step S5-1 includes the following steps：
S5-1-1, L is carried out to the character representation of Attention-MAC trade mark Feature Selection Model output2Regularization；
S5-1-2, by main method for feature analysis, to treated, character representation carries out the feature after feature selecting obtains dimensionality reduction
Vector；
S5-1-3, the feature vector after dimensionality reduction is subjected to L again2Regularization obtain Attention-MAC feature to
Amount.
Preferably, the pond layer added in the step S4-2 is using regional average value pond method to Attention-MAC
The character representation of trade mark Feature Selection Model output is handled, wherein regional average value pond method specifically includes as follows
Step：
C1, input picture pass through trained Attention-MAC trade mark Feature Selection Model, will output one group of W × H ×
The space matrix of K；
C2, the set x={ x that this group of three-dimensional matrice is regarded as to one group of two dimensional character response matrixi, wherein i=1,
2...k, k is the overall channel number of this group output two dimensional character figure, xiRepresent the two dimensional character response square of ith feature channel output
Battle array；
C3, Ω is enabled to represent the two dimensional character response matrix W that ith feature channel exportsi×HiIn all possible position,
xi(p) x is representediThe response of upper position p, it is assumed that：
Continuity based on feature is calculated with xi(pj) centered on 3 × 3 regions characteristic mean as xiPondization it is defeated
Out, i.e.,
Wherein, plThe all pixels point that 3 × 3 regions include thus, n=1,2...9, due to position xi(pj) thus 3 × 3
The center in region, therefore as l=5, there is xi(p5)=xi(pj), by each layer of output characteristic response matrix Wi×HiIt calculates
Regional average value pond, the k dimensional feature for obtaining image indicate：
Feature vector fΩIt is the output of Attention-MAC trade mark Feature Selection Model.
In conclusion the invention discloses a kind of quick trademark image based on Attention mechanism and convolutional neural networks
As search method, Caffe deep learning Open Framework is built first, and open source VGG16 network model is trained；Then it is based on
The design of VGG16 network model includes the Attention network of two layers of convolutional layer, and adds in trained VGG16 network model
Add Attention network；Then using the training set in FlickrLogos-32 data set to being added to Attention network
VGG16 network model is trained；And then it is generated based on the trained VGG16 network model for being added to Attention network
Attention-MAC trade mark Feature Selection Model；Finally based on Attention-MAC trade mark Feature Selection Model to quotient to be checked
Logo image is retrieved, and generates search result.The present invention combines Attention mechanism with convolutional neural networks, proposes
Attention-MAC trade mark Feature Selection Model do not include any full articulamentum, but to middle layer convolution output feature
Figure is handled, and the character representation of original image is obtained, and is avoided the parameter using full articulamentum redundancy, has been reached reduced model
Purpose, while the speed of training and retrieval is improved, reduce false detection rate.
Additional aspect and advantage of the invention will be set forth in part in the description, and will partially become from the following description
Obviously, or practice through the invention is recognized.
Detailed description of the invention
Above-mentioned and/or additional aspect of the invention and advantage will become from the description of the embodiment in conjunction with the following figures
Obviously and it is readily appreciated that, wherein：
Fig. 1 is a kind of based on Attention mechanism and convolutional Neural net in a kind of preferred embodiment disclosed by the invention
The basic flow chart of the quick trademark image retrieval method of network；
Fig. 2 is that the network structure of the VGG16 network model of pre-training in a kind of preferred embodiment disclosed by the invention is shown
It is intended to；
Fig. 3 is after adding Attention network proposed by the invention in a kind of preferred embodiment disclosed by the invention
Network structure schematic diagram；
Fig. 4 is Attention-MAC trademark image proposed by the invention in a kind of preferred embodiment disclosed by the invention
As the network structure of retrieval model is illustrated；
Fig. 5 is that Attention-MAC trademark image retrieval model is utilized in a kind of preferred embodiment disclosed by the invention
Carry out the retrieval flow figure of trade mark detection.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
In the description of the present invention, it is to be understood that, term " longitudinal direction ", " transverse direction ", "upper", "lower", "front", "rear",
The orientation or positional relationship of the instructions such as "left", "right", "vertical", "horizontal", "top", "bottom" "inner", "outside" is based on attached drawing institute
The orientation or positional relationship shown, is merely for convenience of description of the present invention and simplification of the description, rather than the dress of indication or suggestion meaning
It sets or element must have a particular orientation, be constructed and operated in a specific orientation, therefore should not be understood as to limit of the invention
System.
In the description of the present invention, unless otherwise specified and limited, it should be noted that term " installation ", " connected ",
" connection " shall be understood in a broad sense, for example, it may be mechanical connection or electrical connection, the connection being also possible to inside two elements can
, can also indirectly connected through an intermediary, for the ordinary skill in the art to be to be connected directly, it can basis
Concrete condition understands the concrete meaning of above-mentioned term.
The present invention provides a kind of quick trademark image retrieval side based on Attention mechanism and convolutional neural networks
Method, as shown in Figure 1, including the following steps：
S1, Caffe deep learning Open Framework is built, open source VGG16 network model is trained；
S2, the Attention network comprising two layers of convolutional layer is designed based on VGG16 network model, and trained
Attention network is added in VGG16 network model；
S3, using the training set in FlickrLogos-32 data set to the VGG16 network for being added to Attention network
Model is trained；
S4, Attention-MAC quotient is generated based on the trained VGG16 network model for being added to Attention network
Mark Feature Selection Model；
S5, trademark image to be checked is retrieved based on Attention-MAC trade mark Feature Selection Model, and generates inspection
Hitch fruit.
Preferably, the step S1 includes the following steps：
S1-1, Caffe deep learning Open Framework is built, VGG16 network model is carried out using ImageNet data set
Pre-training；
S1-2, the VGG16 network model that pre-training is obtained using the training set in FlickrLogos-32 data set into
The training of row transfer learning.
Preferably, the step S2 includes the following steps：
S2-1, the model parameter design based on VGG16 network model include the Attention network of two layers of convolutional layer；
S2-2, between the last layer output and full articulamentum of the pond layer of the trained VGG16 network model,
Add designed Attention network.
Preferably, the step S3 includes the following steps：
S3-1, the fixed network weight for being added to characteristic extraction part in the VGG16 network model of Attention network；
S3-2, using the training set in FlickrLogos-32 data set to the VGG16 net for being added to Attention network
Attention network in network model is trained.
Preferably, the step S4 includes the following steps：
S4-1, remove full articulamentum in the trained VGG16 network model for being added to Attention network；
S4-2, pond layer is added after Attention network, generate Attention-MAC trade mark Feature Selection Model.
Preferably, the step S5 includes the following steps：
S5-1, the characteristics of image table that Attention-MAC trade mark Feature Selection Model is exported by main method for feature analysis
Show carry out dimensionality reduction, obtains Attention-MAC feature vector；
S5-2, using the trademark image in FlickrLogos-32 data set as input picture, be sequentially inputted to
In Attention-MAC trade mark Feature Selection Model, the Attention-MAC feature vector of branding data collection image is generated, and
Construct the feature database of branding data collection image；
S5-3, using trademark image to be retrieved as input picture, be input to Attention-MAC trade mark Feature Selection Model
In, generate the Attention-MAC feature vector of trademark image to be retrieved；
S5-4, the Attention-MAC feature vector and FlickrLogos-32 data set for calculating trademark image to be retrieved
The cosine similarity of the Attention-MAC feature vector of image obtains branding data collection image based on the first of cosine similarity
Begin to sort；
S5-5, FlickrLogos-32 data images are reset by expanding query, obtains branding data collection figure
As finally sorting with the similarity of trademark image to be retrieved and reporting trade mark search result.
Preferably, the step S1-2 includes the following steps：
S1-2-1, the VGG16 network model that pre-training is obtained using the training set in FlickrLogos-32 data set
Network weight be finely adjusted；
S1-2-2, in the training of transfer learning, pre-training is obtained using standard cross entropy loss function VGG16 net
Network model carries out classification based training.
Preferably, the step S5-1 includes the following steps：
S5-1-1, L is carried out to the character representation of Attention-MAC trade mark Feature Selection Model output2Regularization；
S5-1-2, by main method for feature analysis, to treated, character representation carries out the feature after feature selecting obtains dimensionality reduction
Vector；
S5-1-3, the feature vector after dimensionality reduction is subjected to L again2Regularization obtain Attention-MAC feature to
Amount.
In conclusion the invention discloses a kind of quick trademark image based on Attention mechanism and convolutional neural networks
As search method, Caffe deep learning Open Framework is built first, and open source VGG16 network model is trained；Then it is based on
The design of VGG16 network model includes the Attention network of two layers of convolutional layer, and adds in trained VGG16 network model
Add Attention network；Then using the training set in FlickrLogos-32 data set to being added to Attention network
VGG16 network model is trained；And then it is generated based on the trained VGG16 network model for being added to Attention network
Attention-MAC trade mark Feature Selection Model；Finally based on Attention-MAC trade mark Feature Selection Model to quotient to be checked
Logo image is retrieved, and generates search result.The present invention combines Attention mechanism with convolutional neural networks, proposes
Attention-MAC trade mark Feature Selection Model do not include any full articulamentum, but to middle layer convolution output feature
Figure is handled, and the character representation of original image is obtained, and is avoided the parameter using full articulamentum redundancy, has been reached reduced model
Purpose, while the speed of training and retrieval is improved, reduce false detection rate.
Fig. 2 is that the network structure of the VGG16 network model of pre-training in a kind of preferred embodiment disclosed by the invention is shown
It is intended to, the VGG network model of selection includes 13 layers of convolutional layer and 3 layers of full articulamentum；The knot of training on ImageNet data set
Fruit is 1000 kinds of class categories of output.
Below with reference to Fig. 3, illustrate training for Attention-MAC trade mark Feature Selection Model proposed by the invention
Journey specifically comprises the following steps：
Sa, Caffe deep learning Open Framework is built, VGG16 network model is carried out using ImageNet data set pre-
Training；
Sb, the VGG16 network model that pre-training is obtained using the training image collection in FlickrLogos-32 data set
It is finely adjusted, re -training network weight, completes transfer learning of the model under new data set, further increase and extract feature
Accuracy；
Sc, trained VGG16 network model is added between the output of the last layer pond layer and full articulamentum
Attention network proposed by the present invention；Fixed character extracts subnetwork weight, uses FlickrLogos-32 data set
Training image is trained Attention network.
Sd, trained model is removed to full articulamentum, retains other parts；Pond is added after Attention network
Layer, obtains Attention-MAC trade mark Feature Selection Model proposed by the present invention.
Above-mentioned steps are described in more detail below：
Transfer learning described in step Sb comprises the steps of:
The selection of A1, data set.FlickerLogos-32 data set is divided into three parts P1, P2, P3, respectively trains
Collection, verifying collection and test set.Wherein each part includes 32 different trademark class, and P1 includes 320 images, and P2 includes
3960 images, P3 include 3960 images, and the image of each part is non-intersecting.
The adjustment of A2, loss function.In the training of transfer learning, model is carried out using standard cross entropy loss function
Classification based training：
Wherein, y represents the classification of model output in training；y*The true classification of representative image, (such as a 0-1 vector
Fruit belongs to a certain classification, then representing the position of the category is 1,0) remaining is is；1 is complete 1 vector.
The design and training of Attention network described in step Sc comprise the steps of：
The model parameter of B1, the VGG16 selected according to the present invention, the output of convolutional layer Conv5_3 will be defeated behind pond
This output is known as Origin_feature_map by W × H × K space matrix out.
As the Attention network proposed in B2, the present invention, (Attention_ is referred to as comprising level 2 volume lamination
Conv_1 and Attention_Conv_2), use Softplus function as activation primitive.
The convolution kernel size that B3, wherein Attention_Conv_1 are used is 1 × 1, the port number of number and space matrix
It is mutually all k；For the convolution kernel size that Attention_Conv_2 is used for 1 × 1, number is 1.Training is initial at random when starting
Change the parameter θ of convolution kernel.
B4, the parameter θ of Attention network is trained using standard cross entropy loss function and backpropagation.Instruction
Experienced purpose is the significance level by Attention e-learning each feature into Origin_feature_map.Therefore,
Define the output function φ (f of Attention networki；θ), for indicating feature fiScore on origin_feature_map
(i.e. weight), wherein fi∈Rk, i=1,2...k.
After the completion of B5, Attention network training, function phi (f will be passed through in step B4i；The score φ arrived θ)iAs
Weight, corresponding feature vector f in Origin_feature_mapiIt is multiplied to get the W × H × K adjusted to one group by weight
Space matrix, by this output be known as attention_feature_map.
Pond layer described in step Sd proposes a kind of method in new regional average value pond in the present invention.Specifically
Steps are as follows for calculating：
C1, input picture pass through trained Attention-MAC trade mark Feature Selection Model, will output one group of W × H ×
The space matrix of K.
C2, the set x={ x that this group of three-dimensional matrice is regarded as to one group of two dimensional character response matrixi, wherein i=1,
2...k, k is the overall channel number of this group output two dimensional character figure, xiRepresent the two dimensional character response square of ith feature channel output
Battle array.
C3, Ω is enabled to represent the two dimensional character response matrix W that ith feature channel exportsi×HiIn all possible position,
xi(p) x is representediThe response of upper position p, it is assumed that：
Continuity based on feature is calculated with xi(pj) centered on 3 × 3 regions characteristic mean as xiPondization it is defeated
Out, i.e.,
Wherein, plThe all pixels point that 3 × 3 regions include thus, n=1,2...9, due to position xi(pj) thus 3 × 3
The center in region, therefore as l=5, there is xi(p5)=xi(pj), by each layer of output characteristic response matrix Wi×HiIt calculates
Regional average value pond, the k dimensional feature for obtaining image indicate：
Feature vector fΩIt is the output of Attention-MAC trade mark Feature Selection Model.
Below with reference to Fig. 4, illustrate the calculating process of Attention-MAC feature vector proposed in the present invention.
For the image feature representation of Attention-MAC trade mark Feature Selection Model output, pass through main signature analysis side
Method carries out dimensionality reduction to output character representation, obtains Attention-MAC feature vector proposed by the invention.Circular
It is as follows：
At D1, the k dimensional feature expression exported for Attention-MAC trade mark Feature Selection Model in step Sd
Reason.Over-fitting in order to prevent, first to fΩCarry out L2Regularization.
D2, the k dimensional feature after processed is indicated, using the method for main signature analysis, feature selecting is carried out, for k
A feature, main method for feature analysis, which calculates, finds out the wherein l feature with high correlation or high interactive information, and data are tieed up from k
Drop to l dimension.
D3, the l dimensional vector that main method for feature analysis is calculated carry out L again2Regularization.
The image feature vector that D4, above-mentioned steps obtain, Attention-MAC feature vector as of the present invention.
Below with reference to Fig. 5, illustrate proposed by the invention Attention-MAC trademark image retrieval model to be utilized to carry out quotient
Mark the retrieval flow of detection.
E1, using the trademark image in FlickrLogos-32 data set as input picture, sequentially input Attention-
MAC trademark image retrieval model, obtain indicate FlickrLogos-32 data images l tie up Attention-MAC feature to
Measure fi, i=1,2...n, n is the number of trade mark in FlickrLogos-32 data set.
E2, using trademark image to be retrieved as input picture, be input to Attention-MAC trade mark Feature Selection Model
In, obtain the Attention-MAC feature vector q of trademark image to be retrievedA。
E3, successively calculated using cosine similarity image to be retrieved in FlickrLogos-32 data set image it is similar
Degree.
Utilize similarity siThe similarity for representing i-th image and image to be retrieved in FlickrLogos-32 data set obtains
Point, according to score siIt is initially sorted to FlickrLogos-32 data images.
In E4, the initial sequence of selection, preceding 5 images are come, by its Attention-MAC feature vector and quotient to be retrieved
Mean value is calculated after the Attention-MAC feature vector summation of logo image,
The q that will be obtainedreAs new query vector, preceding 100 images in initial sequence are selected, q is usedreCalculate phase
Like degree, preceding 100 images are reset according to new score, obtain ranking results to the end.
It should be noted that system structure shown in Fig. 1-Fig. 5 of the present invention or method flow are of the invention some
Preferred embodiment is shown here and simply facilitates the understanding present invention and be not considered as limiting the invention, of the invention
Under thought guidance, it is within the scope of the present invention to implement the structure or method obtained according to the technique and scheme of the present invention,
Therefore not to repeat here.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
Centainly refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any
One or more embodiment or examples in can be combined in any suitable manner.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that：Not
A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this
The range of invention is defined by the claims and their equivalents.
Claims (9)
1. a kind of quick trademark image retrieval method based on Attention mechanism and convolutional neural networks, which is characterized in that institute
The method of stating includes the following steps：
S1, Caffe deep learning Open Framework is built, open source VGG16 network model is trained；
S2, the Attention network comprising two layers of convolutional layer is designed based on VGG16 network model, and in trained VGG16 net
Attention network is added in network model；
S3, using the training set in FlickrLogos-32 data set to the VGG16 network model for being added to Attention network
It is trained；
S4, Attention-MAC trade mark spy is generated based on the trained VGG16 network model for being added to Attention network
Sign extracts model；
S5, trademark image to be checked is retrieved based on Attention-MAC trade mark Feature Selection Model, and generates retrieval knot
Fruit.
2. the quick trademark image retrieval side according to claim 1 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S1 includes the following steps：
S1-1, Caffe deep learning Open Framework is built, VGG16 network model is instructed in advance using ImageNet data set
Practice；
S1-2, the VGG16 network model that pre-training obtains is moved using the training set in FlickrLogos-32 data set
Move the training of study.
3. the quick trademark image retrieval side according to claim 1 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S2 includes the following steps：
S2-1, the model parameter design based on VGG16 network model include the Attention network of two layers of convolutional layer；
S2-2, between the last layer output and full articulamentum of the pond layer of the trained VGG16 network model, addition
Designed Attention network.
4. the quick trademark image retrieval side according to claim 1 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S3 includes the following steps：
S3-1, the fixed network weight for being added to characteristic extraction part in the VGG16 network model of Attention network；
S3-2, using the training set in FlickrLogos-32 data set to the VGG16 network mould for being added to Attention network
Attention network in type is trained.
5. the quick trademark image retrieval side according to claim 1 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S4 includes the following steps：
S4-1, remove full articulamentum in the trained VGG16 network model for being added to Attention network；
S4-2, pond layer is added after Attention network, generate Attention-MAC trade mark Feature Selection Model.
6. the quick trademark image retrieval side according to claim 1 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S5 includes the following steps：
S5-1, the image feature representation that Attention-MAC trade mark Feature Selection Model is exported by main method for feature analysis into
Row dimensionality reduction obtains Attention-MAC feature vector；
S5-2, using the trademark image in FlickrLogos-32 data set as input picture, be sequentially inputted to Attention-
In MAC trade mark Feature Selection Model, the Attention-MAC feature vector of branding data collection image is generated, and constructs trade mark number
According to the feature database of collection image；
S5-3, using trademark image to be retrieved as input picture, be input in Attention-MAC trade mark Feature Selection Model,
Generate the Attention-MAC feature vector of trademark image to be retrieved；
S5-4, the Attention-MAC feature vector and FlickrLogos-32 data images for calculating trademark image to be retrieved
Attention-MAC feature vector cosine similarity, obtain initial row of the branding data collection image based on cosine similarity
Sequence；
S5-5, FlickrLogos-32 data images are reset by expanding query, obtain branding data collection image with
The similarity of trademark image to be retrieved finally sorts and reports trade mark search result.
7. the quick trademark image retrieval side according to claim 2 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S1-2 includes the following steps：
The net of S1-2-1, the VGG16 network model that pre-training is obtained using the training set in FlickrLogos-32 data set
Network weight is finely adjusted；
S1-2-2, in the training of transfer learning, pre-training is obtained using standard cross entropy loss function VGG16 network mould
Type carries out classification based training.
8. the quick trademark image retrieval side according to claim 6 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the step S5-1 includes the following steps：
S5-1-1, L is carried out to the character representation of Attention-MAC trade mark Feature Selection Model output2Regularization；
S5-1-2, by main method for feature analysis to treated character representation carry out feature selecting obtain the feature after dimensionality reduction to
Amount；
S5-1-3, the feature vector after dimensionality reduction is subjected to L again2Regularization obtains Attention-MAC feature vector.
9. the quick trademark image retrieval side according to claim 5 based on Attention mechanism and convolutional neural networks
Method, which is characterized in that the pond layer added in the step S4-2 is using regional average value pond method to Attention-MAC quotient
The character representation of mark Feature Selection Model output is handled, wherein regional average value pond method specifically includes following step
Suddenly：
C1, input picture pass through trained Attention-MAC trade mark Feature Selection Model, export one group of W × H × K sky
Between matrix；
C2, the set x={ x that this group of three-dimensional matrice is regarded as to one group of two dimensional character response matrixi, wherein i=1,2...k, k are
The overall channel number of this group output two dimensional character figure, xiRepresent the two dimensional character response matrix of ith feature channel output；
C3, Ω is enabled to represent the two dimensional character response matrix W that ith feature channel exportsi×HiIn all possible position, xi(p)
Represent xiThe response of upper position p, it is assumed that：
Continuity based on feature is calculated with xi(pj) centered on 3 × 3 regions characteristic mean as xiPondization output, i.e.,
Wherein, plThe all pixels point that 3 × 3 regions include thus, n=1,2...9, due to position xi(pj) 3 × 3 region thus
Center have x therefore as l=5i(p5)=xi(pj), by each layer of output characteristic response matrix Wi×HiZoning
Mean value pond, the k dimensional feature for obtaining image indicate：
Feature vector fΩIt is the output of Attention-MAC trade mark Feature Selection Model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810750096.1A CN108875076B (en) | 2018-07-10 | 2018-07-10 | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810750096.1A CN108875076B (en) | 2018-07-10 | 2018-07-10 | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108875076A true CN108875076A (en) | 2018-11-23 |
| CN108875076B CN108875076B (en) | 2021-07-20 |
Family
ID=64300452
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810750096.1A Active CN108875076B (en) | 2018-07-10 | 2018-07-10 | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875076B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109697257A (en) * | 2018-12-18 | 2019-04-30 | 天罡网(北京)安全科技有限公司 | It is a kind of based on the network information retrieval method presorted with feature learning anti-noise |
| CN109857897A (en) * | 2019-02-14 | 2019-06-07 | 厦门一品威客网络科技股份有限公司 | A kind of trademark image retrieval method, apparatus, computer equipment and storage medium |
| CN110334226A (en) * | 2019-04-25 | 2019-10-15 | 吉林大学 | The depth image search method of fusion feature Distribution Entropy |
| CN110599459A (en) * | 2019-08-14 | 2019-12-20 | 深圳市勘察研究院有限公司 | Underground pipe network risk assessment cloud system based on deep learning |
| CN111694977A (en) * | 2020-06-12 | 2020-09-22 | 桂林电子科技大学 | Vehicle image retrieval method based on data enhancement |
| CN111694974A (en) * | 2020-06-12 | 2020-09-22 | 桂林电子科技大学 | Depth hash vehicle image retrieval method integrating attention mechanism |
| CN113127661A (en) * | 2021-04-06 | 2021-07-16 | 中国科学院计算技术研究所 | Multi-supervision medical image retrieval method and system based on cyclic query expansion |
| CN113569070A (en) * | 2021-07-24 | 2021-10-29 | 平安科技（深圳）有限公司 | Image detection method and device, electronic equipment and storage medium |
| WO2022036867A1 (en) * | 2020-08-21 | 2022-02-24 | 广东电网有限责任公司清远供电局 | Method for reconstructing three-dimensional model of transformer substation |
| CN113569070B (en) * | 2021-07-24 | 2024-07-02 | 平安科技（深圳）有限公司 | Image detection method and device, electronic equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103226826A (en) * | 2013-03-20 | 2013-07-31 | 西安电子科技大学 | Method for detecting changes of remote sensing image of visual attention model based on local entropy |
| CN108038519A (en) * | 2018-01-30 | 2018-05-15 | 浙江大学 | A kind of uterine neck image processing method and device based on dense feature pyramid network |
| US20180144248A1 (en) * | 2016-11-18 | 2018-05-24 | Salesforce.Com, Inc. | SENTINEL LONG SHORT-TERM MEMORY (Sn-LSTM) |
| WO2018106783A1 (en) * | 2016-12-06 | 2018-06-14 | Siemens Energy, Inc. | Weakly supervised anomaly detection and segmentation in images |
| CN108171141A (en) * | 2017-12-25 | 2018-06-15 | 淮阴工学院 | The video target tracking method of cascade multi-pattern Fusion based on attention model |
| CN108229267A (en) * | 2016-12-29 | 2018-06-29 | 北京市商汤科技开发有限公司 | Object properties detection, neural metwork training, method for detecting area and device |
-
2018
- 2018-07-10 CN CN201810750096.1A patent/CN108875076B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103226826A (en) * | 2013-03-20 | 2013-07-31 | 西安电子科技大学 | Method for detecting changes of remote sensing image of visual attention model based on local entropy |
| US20180144248A1 (en) * | 2016-11-18 | 2018-05-24 | Salesforce.Com, Inc. | SENTINEL LONG SHORT-TERM MEMORY (Sn-LSTM) |
| WO2018106783A1 (en) * | 2016-12-06 | 2018-06-14 | Siemens Energy, Inc. | Weakly supervised anomaly detection and segmentation in images |
| CN108229267A (en) * | 2016-12-29 | 2018-06-29 | 北京市商汤科技开发有限公司 | Object properties detection, neural metwork training, method for detecting area and device |
| CN108171141A (en) * | 2017-12-25 | 2018-06-15 | 淮阴工学院 | The video target tracking method of cascade multi-pattern Fusion based on attention model |
| CN108038519A (en) * | 2018-01-30 | 2018-05-15 | 浙江大学 | A kind of uterine neck image processing method and device based on dense feature pyramid network |
Non-Patent Citations (2)
| Title |
|---|
| LE WANG;JINLIANG ZANG;QILIN ZHANG;ZHENXING NIU;GANG HUA: ""Action Recognition by an Attention-Aware Temporal Weighted Convolutional Neural Network"", 《SENSORS》 * |
| 李振东，钟勇，陈蔓，曹冬平: ""基于深度特征的快速人脸图像检索方法"", 《光学学报》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109697257A (en) * | 2018-12-18 | 2019-04-30 | 天罡网(北京)安全科技有限公司 | It is a kind of based on the network information retrieval method presorted with feature learning anti-noise |
| CN109857897B (en) * | 2019-02-14 | 2021-06-29 | 厦门一品威客网络科技股份有限公司 | Trademark image retrieval method and device, computer equipment and storage medium |
| CN109857897A (en) * | 2019-02-14 | 2019-06-07 | 厦门一品威客网络科技股份有限公司 | A kind of trademark image retrieval method, apparatus, computer equipment and storage medium |
| CN110334226A (en) * | 2019-04-25 | 2019-10-15 | 吉林大学 | The depth image search method of fusion feature Distribution Entropy |
| CN110334226B (en) * | 2019-04-25 | 2022-04-05 | 吉林大学 | Depth image retrieval method fusing feature distribution entropy |
| CN110599459A (en) * | 2019-08-14 | 2019-12-20 | 深圳市勘察研究院有限公司 | Underground pipe network risk assessment cloud system based on deep learning |
| CN111694974A (en) * | 2020-06-12 | 2020-09-22 | 桂林电子科技大学 | Depth hash vehicle image retrieval method integrating attention mechanism |
| CN111694977A (en) * | 2020-06-12 | 2020-09-22 | 桂林电子科技大学 | Vehicle image retrieval method based on data enhancement |
| WO2022036867A1 (en) * | 2020-08-21 | 2022-02-24 | 广东电网有限责任公司清远供电局 | Method for reconstructing three-dimensional model of transformer substation |
| CN113127661A (en) * | 2021-04-06 | 2021-07-16 | 中国科学院计算技术研究所 | Multi-supervision medical image retrieval method and system based on cyclic query expansion |
| CN113127661B (en) * | 2021-04-06 | 2023-09-12 | 中国科学院计算技术研究所 | Multi-supervision medical image retrieval method and system based on cyclic query expansion |
| CN113569070A (en) * | 2021-07-24 | 2021-10-29 | 平安科技（深圳）有限公司 | Image detection method and device, electronic equipment and storage medium |
| CN113569070B (en) * | 2021-07-24 | 2024-07-02 | 平安科技（深圳）有限公司 | Image detection method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108875076B (en) | 2021-07-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108875076A (en) | A kind of quick trademark image retrieval method based on Attention mechanism and convolutional neural networks | |
| CN106504233B (en) | Unmanned plane inspection image electric power widget recognition methods and system based on Faster R-CNN | |
| Zhu et al. | Traffic-sign detection and classification in the wild | |
| CN104850845B (en) | A kind of traffic sign recognition method based on asymmetric convolutional neural networks | |
| CN107358257B (en) | Under a kind of big data scene can incremental learning image classification training method | |
| Hu et al. | Recognition of pornographic web pages by classifying texts and images | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| CN102054178B (en) | A kind of image of Chinese Painting recognition methods based on local semantic concept | |
| CN105426919B (en) | The image classification method of non-supervisory feature learning is instructed based on conspicuousness | |
| CN104462494B (en) | A kind of remote sensing image retrieval method and system based on unsupervised feature learning | |
| CN103810299B (en) | Image retrieval method on basis of multi-feature fusion | |
| CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
| CN109948425A (en) | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device | |
| CN107256246A (en) | PRINTED FABRIC image search method based on convolutional neural networks | |
| CN101944174B (en) | Identification method of characters of licence plate | |
| CN107622272A (en) | A kind of image classification method and device | |
| CN104572940B (en) | A kind of image automatic annotation method based on deep learning and canonical correlation analysis | |
| CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN106557579A (en) | A kind of vehicle model searching system and method based on convolutional neural networks | |
| Li et al. | A generative/discriminative learning algorithm for image classification | |
| CN109753959B (en) | Road traffic sign detection method based on self-adaptive multi-scale feature fusion | |
| CN109034035A (en) | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features | |
| CN111612051A (en) | Weak supervision target detection method based on graph convolution neural network | |
| CN109583507A (en) | A kind of pig body recognition methods based on depth convolutional neural networks |
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 |