CN110059206A - A kind of extensive hashing image search method based on depth representative learning - Google Patents
A kind of extensive hashing image search method based on depth representative learning Download PDFInfo
- Publication number
- CN110059206A CN110059206A CN201910249642.8A CN201910249642A CN110059206A CN 110059206 A CN110059206 A CN 110059206A CN 201910249642 A CN201910249642 A CN 201910249642A CN 110059206 A CN110059206 A CN 110059206A
- Authority
- CN
- China
- Prior art keywords
- loss function
- label
- layer
- data
- hash
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Library & Information Science (AREA)
- Image Analysis (AREA)
Abstract
A kind of extensive hashing image search method based on depth representative learning, the post-processing of Optimization Learning and search result including data prediction, based on semi-supervised convolutional neural networks layer building, based on pairs of Hash loss function design step loss function.The present invention carries out the extraction of characteristics of image and the study of hash function using depth convolutional neural networks layer and full articulamentum, devise the assembling loss function for intersecting entropy loss item, triple loss item and pseudo label loss item and constituting of tape label, it is optimized using the stochastic gradient descent method with momentum, computational efficiency with higher finally realizes the unified image retrieval performance of accuracy and speed.The present invention carries out picture search problem to carry out High Efficiency Modeling under the basis based on image data structure and label under the status of picture search field fast development, the model accuracy and more optimized inquiry velocity effectively improved.
Description
Technical field
The invention belongs to field of image search, are related to a kind of large-scale image search method.
Background technique
With the rapid development of the technologies such as big data technology, development of Mobile Internet technology, technology of Internet of things, image, video etc.
Acquisition, convergence and the storage of multimedia resource are more and more convenient.In practical applications, 80% or so data be all with document,
The unstructured data of the forms such as image, video, audio storage, and same such unstructured data all presses index every year
Increase about 3/5ths.In unstructured data, image be again it is most important, occupy enormous proportions, but at the same time
Important information is contained.Computer vision is primarily upon the research of image data, therefore also naturally enough becomes machine learning
One of with the most popular research point of artificial intelligence technology.
Application problem in face of various complexity and the amount of images growth rate that is exceedingly fast promote traditional images analysis to handle
Technology encounters very big challenge, especially huge resource, superelevation dimension, mass memory, slowly in terms of challenge.
To solve the problems, such as this some column, a variety of approximate algorithms with low time complexity are suggested, wherein to be based on hash algorithm
The algorithm of representative obtains a large amount of concerns, because it has many advantages, such as constant enquiry time and a small amount of compression binary code storage.But
A large amount of image data is all not no label, and especially in fields such as picture search, the image data of only only a few is had
Label information, this is also to consume what a large amount of manpower and material resources obtained.How to merge has flag data and a large amount of nothings using a small amount of
The image search method of flag data is the method with Great significance and actual demand.Design targetedly efficient needle
Derivation algorithm to large-scale image search is vital.
Hash algorithm is typical nearest neighbor search method, is mainly used for solving memory space recited above greatly and retrieves
The problem of time length.In hash algorithm, common target is one mapping function W of study, and sample is expressed as a string of fixations
The binary-coding of length, encoding usually using {+1, -1 } or {+1,0 } indicates bit each, so that semantically similar
Sample can obtain similar coding, we are usually using Hamming distance from the similitude between measurement binary-coding.
In initial work, local sensitivity hash algorithm (LSH) is a kind of popular algorithm, the key of LSH algorithm
Part is to divide space using the hyperplane being randomly generated, and it is every to determine to fall in the position of hyperplane according to sample
One value, LSH require similar sample to fall among the same Hash bucket with greater probability value, without similar sample with
Minimum probability value is fallen among the same Hash bucket, and by using bit multiple, final LSH can reach satisfactory inspection
Suo Xiaoguo, LSH method have stringent theoretical proof to guarantee its performance.But in order to obtain shorter more compact coding
And better retrieval performance, researcher consider the label information of image from different objective function is constructed, and choose different excellent
Change method must further improve the performance of Hash retrieval using linearly or nonlinearly model etc..
The appearance of deep neural network causes huge repercussion in computer vision field, and researcher is unnecessary to go again
The complicated inefficient expression feature of engineer, but neural network oneself is allowed to go to learn the characterization of image.The present invention is exactly to combine
Deep neural network designs a kind of semi-supervised Hash searching algorithm, the Hash searching algorithm of the remote ultra-traditional of retrieval rate.
Summary of the invention
In order to overcome the lower deficiency of accuracy rate of existing large-scale image search method, the present invention provides a kind of accurate
The higher extensive hashing image search method based on depth representative learning of rate.
The technical solution adopted by the present invention to solve the technical problems is:
A kind of extensive hashing image search method based on depth representative learning, comprising the following steps:
Step 1: data prediction；
Step 2: deep neural network constructs, using pretreated labeling in data prediction step or without label
Data carry out deep neural network model building；
Step 3: Hash loss function designs, the Hash loss function is by softmax cross entropy loss function, ternary
Group loss function, pseudo label loss function three parts composition.
Step 4: deep neural network training and model optimization derive letter for the Hash loss function of step 3
Several gradients carries out the update of parameter until convergence to entire depth neural network model using the gradient descent method with momentum；
Step 5, the deep neural network model after input picture to training, is calculated binary system Hash codes to data
Image in library calculates Hamming distances, returns to candidate result of the most like image as picture search.
Preferably, the step 5 further includes, the candidate result is the result of coarseness；Select layer mind second from the bottom
Output through network is further calculated the Euclidean distance of these characterizations in candidate result, is arranged again with this as characterization
Sequence, the final result are exactly fine-grained search result.
Preferably, the data prediction step includes but is not limited to the standardization of image normalizing, drop in the step 1
It makes an uproar, deblurring, image data augmentation preliminary treatment, further includes whether having class declaration according to data to be divided into tape label data
With without label data, for there is the data of label, it is also necessary to construct triple structure.
Preferably, the output of neural network the last layer includes two parts in the step 3: Hash codes and classification
Value；The output of neural network the last layer includes two parts: Hash codes and class label；The class label intersects as softmax
The input of entropy loss function is converted into the class probability between [0,1] by softmax function, then passes through cross entropy loss function
Obtain penalty values to the end；The Hash codes are the inputs of triple loss function and pseudo label loss function.
The softmax cross entropy loss function:
JS(xi)=- yilogP(xi)
The wherein index of i representative sample, x represent data sample, and y represents label, P (xi) represent i-th of sample and belong to y's
Probability value；
The triple loss function:
Wherein, i is sample index, and x represents data sample,Indicate xLSample with same label,Represent and
xLSample with different labels, h (*) represent the corresponding output of network, mtFor threshold parameter；
The pseudo label loss function:
Wherein, i, j are similarly sample index, for the data of not tape label, assign the maximum label of probability value to its work
For pseudo label, mpFor threshold parameter, the maximum distance of two different classes of samples is represented, A (i, j)=1 indicates that sample i, j belong to
Same class, A (i, j)=0 indicate that sample i, j belong to inhomogeneity, and h (*) represents the corresponding output of network.The Hash loss function
Are as follows:
Wherein JS(xi) it is softmax cross entropy loss function,It is triple loss function,It is pseudo label loss function, p is pseudo label, and α, β and γ are weight parameters.
Preferably, for softmax cross entropy loss function, passing through deep learning lower portion in the step 4
Automatic derivation；For triple loss function and pseudo label loss function, derivation is re-started using back-propagation algorithm.
Preferably, in an iterative process, being trained optimization to tape label data first, first not in the step 4
Using pseudo label loss function；When loss drops to given threshold, unlabeled exemplars data are added, are lost in conjunction with pseudo label
Function is iterated optimization to Hash loss function until convergence.
Preferably, the deep neural network uses basic neural network framework in the step 2, pass through stacking
Convolutional layer, pond layer and full articulamentum form core network.In the neural network framework, 2 layers are added to connect entirely using 5 layers of convolutional layer
The mode of layer is connect, and distinguishes Hash coding layer and classification layer simultaneously in the last layer；Network parameter is set as, first layer convolution
Layer: convolution kernel size is 11x11, stride 4, port number 64, activation primitive Relu；Pool layers of Max: core size is
3x3, stride 2；Second layer convolutional layer: convolution kernel size is 5x5, and stride 1, port number 192, activation primitive is
Relu.Pool layers of Max: core size is 3x3, stride 2；Third layer convolutional layer: convolution kernel size be 3x3, stride 1,
Port number is 384, activation primitive Relu；4th layer of convolutional layer: convolution kernel size is 5x5, stride 1, and port number is
256, activation primitive Relu；Layer 5 convolutional layer: convolution kernel size is 3x3, stride 1, port number 256, activation letter
Number is Relu；The full articulamentum neuron nodal point number of first layer is 2048；The full articulamentum neuron nodal point number of the second layer is Hash codes
The sum of digit and classification number.
Beneficial effects of the present invention are mainly manifested in:
1, in the network model building based on depth, the network architecture of similar AlexNet is selected, directly by original graph
As the input as model, advantage is that neural network can acquire the characterization of better image data, has stronger extensive
Performance；
2, ingenious to combine triple loss in the design of Hash loss function, softmax intersects entropy loss and pseudo label damage
Item is lost, the study of model multi-angle can be conducive to, to obtain higher robustness.
3, in the optimization of model and the update of parameter, the derived function of loss function and the training side of model are given
Method can reduce manually flag data in the case where guaranteeing test accuracy rate in conjunction with this semi-supervised mode of learning
The required time, to save human cost.
4, in the step of image retrieval, we have proposed multi-level sort methods from thick to thin, can allow in the time
In the case where, further increase the accuracy rate of retrieval.
Detailed description of the invention
Fig. 1 is the flow chart of the extensive hashing image search method the present invention is based on depth representative learning.
Fig. 2 is the network of the extensive hashing image search method the present invention is based on depth representative learning.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
Referring to Figures 1 and 2, a kind of extensive hashing image search method based on depth representative learning, by image data
Structure and label condition driving effectively modeling proposed a kind of based on depthmeter on the basis of existing picture search algorithm
The extensive hashing image search method of study is levied, its advantages are that of obtaining high accuracy and quick search speed and rule
Mould scalability and less memory space.
The extensive hashing image search method the following steps are included:
Step 1: data prediction
In data prediction part, first have to class declaration whether is had according to the data in data set, for example, class name or
Person one-hot coding, is divided into tape label data and without label data, for there is the data of label, it is also necessary to construct ternary
Group structure, that is, for sample xL, choose another sample for having same labelAnother is chosen with different marks
The sample of labelThe data for forming a triple structure, the input as network.It is also desirable to all data sets
Standardization processing is carried out, the operation for subtracting mean value divided by standard deviation is carried out, controls the numeric distribution of sample.In the situation of sample deficiency
Under, need to carry out data augmentation operation, including rotated, translated to picture, fuzzy noise etc., and the width of image is high
Resize to 224.
Step 2: deep neural network constructs
Deep neural network model is carried out using pretreated labeling in data prediction step or without label data
Building, herein, we still take basic neural network framework, by stacking convolutional layer, pond layer and full articulamentum group
At core network.In network model building, the network structure of AlexNet is copied, adds 2 layers of full articulamentum using 5 layers of convolutional layer
Mode, and the last layer simultaneously distinguish Hash coding layer and classification layer.Specific network parameter is set as, first layer convolution
Layer: convolution kernel size is 11x11, stride 4, port number 64, activation primitive Relu.Pool layers of Max: core size is
3x3, stride 2.Second layer convolutional layer: convolution kernel size is 5x5, and stride 1, port number 192, activation primitive is
Relu.Pool layers of Max: core size is 3x3, stride 2.Third layer convolutional layer: convolution kernel size be 3x3, stride 1,
Port number is 384, activation primitive Relu.4th layer of convolutional layer: convolution kernel size is 5x5, stride 1, and port number is
256, activation primitive Relu.Layer 5 convolutional layer: convolution kernel size is 3x3, stride 1, port number 256, activation letter
Number is Relu.The full articulamentum neuron nodal point number of first layer is 2048.The full articulamentum neuron nodal point number of the second layer is Hash codes
The sum of digit and classification number, such as Hash code bit number are 48, share 10 classes, then the number of nodes of full articulamentum is 58.Specific net
Network structure chart is as shown in Figure 2.
Step 3: Hash loss function designs
In the design of loss function, it is divided into three parts, the softmax cross entropy loss function of tape label, triple
Loss function, pairs of pseudo label loss function.
The output of neural network the last layer includes two parts: Hash codes and class label.
Input of the class label as softmax will be converted into the class probability between [0,1] by softmax function, then be led to
It crosses cross entropy objective function and obtains penalty values to the end, softmax cross entropy loss function is most common Classification Loss function.
Hash codes part is the input of triple loss function item and pairs of pseudo label loss item, and triple loss is not
It is same as another loss function that softmax intersects entropy loss, it is characterized in that identical two samples of the meaning of one's words is promoted to measure
Under space from as far as possible closely, and different two samples of the meaning of one's words under metric space from must as far as possible far.It is pairs of to be
Tag entry loss item can regard certain a part of triple loss item in the case where the meaning of one's words is identical or different as.
First part is that commonly used softmax intersects entropy loss in image classification, for each sample, our all mistakes
Softmax obtains the probability value P of classification, the calculating target function in a manner of cross entropy.It is as follows:
JS(xi)=- yilog P(xi)
The wherein index of i representative sample, x representative sample, y represent label, P (xi) represent the probability that i-th of sample belongs to y
Value；
Second part is triple loss item, for the triple data of the tape label constructed in step 1, Wo Menshe
The distance that a mode measures them is counted.Each sample x can have output valve h (x), for identical exemplar xLWithThe distance between they are minimized, are calculated with Euclidean distance.Conversely, then maximizing the distance between they.It is as follows:
Wherein, i is sample index, and x represents data sample,Indicate xLSample with same label,Representative and xL
Sample with different labels, h (*) represent the corresponding output of network, mtFor threshold parameter；
Part III is pseudo label loss item, and for unlabelled sample, we can obtain institute by neural network output
Belong to the probability value of each classification, the pseudo label that we assign the classification where maximum value as it constructs pseudo- as second part
The distance between label measurement.It is as follows:
Wherein, i, j are similarly sample index, mpFor threshold parameter, the maximum distance of two different classes of samples, A are represented
(i, j)=1 indicates that sample i, j belong to same class, and A (i, j)=0 indicates that sample i, j belong to inhomogeneity, and it is corresponding that h (*) represents network
Output.
Hash loss function specifically:
Wherein JS(xi) be tape label softmax cross entropy loss function,It is triple damage
Function is lost,It is pseudo label loss function, p is pseudo label.Three loss functions
It can be regulated and controled by weight between, wherein α, β and γ be as weight parameter, we enable α=1 herein, β=1, γ=
0.5。
Step 4: deep neural network training and model optimization
For the loss function of step 3, the gradient of derived function loss function utilizes the gradient descent method pair with momentum
Entire model carries out the update of parameter until convergence.
Wherein, we are trained convergence to the sample of tape label first, that is, loss decline about to 0.2 with
Under, it then is added unlabeled exemplars, alternately be trained.This semi-supervised alternating training method can be very big
The time required for marker samples is reduced in degree, saves cost of human resources.The side SGD with momentum is selected in terms of optimizer
Method, learning rate are set as 0.001, and momentum is set as 0.9.Common gradient declines updating method are as follows:
Wherein WtFor the parameter of t moment neural network, ▽ J is the derivative of loss function J.
Momentum gradient descent method is as follows:
Wherein θ is momentum parameter, replaces former gradient to carry out parameter update using gradient after exponent-weighted average.Because every
The information of gradient before gradient after a exponent-weighted average contains.
For softmax cross entropy loss function, pass through the automatic derivation of deep learning lower portion；Triple is lost
Function and pseudo label loss function, re-start derivation using back-propagation algorithm；
For triple loss function derivation:
Wherein I represents indicator function, when C is True, then for 1 on the contrary be then -1；
For pseudo label loss function derivation:
As A (i, j)=1:
As A (i, j)=0:
Optimization is iterated to objective function until convergence.
Step 5: image retrieval
After the completion of model training, we keep the Hash coding of training data.Give a test query picture, mould
Type can export the binary-coding that it is predicted.The Hash of the binary-coding of prediction and training data is encoded and carries out hamming by we
The measurement of distance returns to candidate result of the most like image as picture search, is waited according to apart from size given threshold
Selected works prepare.
If it is required that higher inquiry precision, can carry out finer sequence, to the inverse of Candidate Set and training sample
Second layer feature carries out Euclidean distance comparison, to resequence, obtains more accurately retrieval effectiveness.It is looked into for example, we give
Ask image Iq, candidate pool P has been obtained after hamming sorts, we are that the higher front layer of dimension is used to connect output entirely as candidate
Collect the detailed characterizations of P, we use VqAnd Vi PRespectively represent query image IqWith the feature vector of candidate pool P, we pass through Europe
Formula distance measure P in i-th of sample and query image similarity S, as follows: where S is smaller, represent two samples it
Between similarity it is higher, we can reorder to candidate pool by S, obtain more preferably query result.
si=| | Vq-Vi P||
Wherein, S is smaller, and the similarity represented between two samples is higher, we can reset candidate pool by S
Sequence obtains more preferably query result.
The scheme of the present embodiment, using depth convolutional neural networks layer and full articulamentum carry out characteristics of image extraction and
The study of hash function devises the intersection entropy loss item, triple loss item and pseudo label damage of tape label on this basis
The assembling loss function that item is constituted is lost, is optimized using the stochastic gradient descent method with momentum, calculating with higher
Efficiency, finally, the searching method that the present invention is combined using coarseness and fine granularity, realizes the unified image of accuracy and speed with this
Retrieval performance.The present invention carries out based on image data picture search problem under the status of picture search field fast development
High Efficiency Modeling, the model accuracy effectively improved and more optimized inquiry velocity are carried out under the basis of structure and label.The present invention
The algorithm proposed has better accuracy and robustness than traditional several image retrieval algorithms, and the present invention can be used for
Large-scale image retrieval scene, serves intelligent searching system, such as: search engine to scheme to search figure function, electric business website
Similar commercial articles searching, the individualized content of social platform recommends etc..
Protection content of the invention is not limited to above embodiments.Without departing from the spirit and scope of the invention, originally
Field technical staff it is conceivable that variation and advantage be all included in the present invention, and with appended claims be protect
Protect range.
Claims (10)
1. a kind of extensive hashing image search method based on depth representative learning, which is characterized in that the method includes with
Lower step:
Step 1: pre-processing image data；
Step 2: deep neural network constructs, using pretreated labeling in data prediction step or without label data
Carry out deep neural network model building；
Step 3: Hash loss function designs, the Hash loss function is damaged by softmax cross entropy loss function, triple
Lose function, pseudo label loss function three parts composition.
Step 4: deep neural network training and model optimization, for the Hash loss function of step 3, derived function loss
The gradient of function carries out the update of parameter to entire depth neural network model until receiving using the gradient descent method with momentum
It holds back.
Step 5, the deep neural network model after input picture to training, is calculated binary system Hash codes in database
Image calculate Hamming distances, return to candidate result of the most like image as picture search.
2. a kind of extensive hashing image search method based on depth representative learning as described in claim 1, feature exist
In the step 5 further includes that the candidate result is the result of coarseness；The output of layer neural network second from the bottom is selected to make
For characterization, the Euclidean distance of these characterizations is further calculated in candidate result, is resequenced with this, the knot finally obtained
Fruit is exactly fine-grained search result.
3. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 1 or 2, feature
It is, in the step 1, the data prediction step includes but is not limited to the standardization of image normalizing, noise reduction, deblurring, figure
It further include whether having class declaration according to data to be divided into tape label data and without number of tags as the preliminary treatment of data augmentation
According to for there is the data of label, it is also necessary to construct triple structure.
4. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 3, feature exist
In in the step 3, the output of neural network the last layer includes two parts: Hash codes and class label；The class label is made
For the input of softmax cross entropy loss function, the class probability between [0,1] is converted by softmax function, then passes through friendship
Pitch the penalty values of entropy loss function call to the end；The Hash codes are the defeated of triple loss function and pseudo label loss function
Enter.
5. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 4, feature exist
In the softmax cross entropy loss function:
JS(xi)=- yilogP(xi)
The wherein index of i representative sample, x represent data sample, and y represents label, P (xi) represent the probability that i-th of sample belongs to y
Value；
The triple loss function:
Wherein, i is sample index, and x represents data sample,Indicate xLSample with same label,Representative and xLIt has
The sample of different labels, h (*) represent the corresponding output of network, mtFor threshold parameter；
The pseudo label loss function:
Wherein, i, j are similarly sample index, for the data of not tape label, assign the maximum label of probability value to it as puppet
Label, mpFor threshold parameter, the maximum distance of two different classes of samples is represented, A (i, j)=1 indicates sample i, and j belongs to same
Class, A (i, j)=0 indicate that sample i, j belong to inhomogeneity, and h (*) represents the corresponding output of network.
6. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 5, feature exist
In the Hash loss function are as follows:
Wherein JS(xi) it is softmax cross entropy loss function,It is triple loss function,It is pseudo label loss function, p is pseudo label, and α, β and γ are weight parameters.
7. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 1 or 2, feature
It is, in the step 4, for softmax cross entropy loss function, passes through the automatic derivation of deep learning lower portion；For
Triple loss function and pseudo label loss function, re-start derivation using back-propagation algorithm.
8. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 1 or 2, feature
It is, in the step 4, in an iterative process, optimization is trained to tape label data first, does not first use pseudo label to damage
Lose function；When loss drops to given threshold, unlabeled exemplars data are added, in conjunction with pseudo label loss function, to Hash
Loss function is iterated optimization until convergence.
9. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 1 or 2, feature
It is, in the step 2, the deep neural network uses basic neural network framework, by stacking convolutional layer, Chi Hua
Layer and full articulamentum form core network.
10. a kind of extensive hashing image search method based on depth representative learning as claimed in claim 9, feature exist
In the neural network framework distinguishes Hash in such a way that 5 layers of convolutional layer add 2 layers of full articulamentum, and in the last layer simultaneously
Coding layer and classification layer；Network parameter is set as, first layer convolutional layer: convolution kernel size is 11x11, stride 4, port number
It is 64, activation primitive Relu；Pool layers of Max: core size is 3x3, stride 2；Second layer convolutional layer: convolution kernel size
For 5x5, stride 1, port number 192, pool layers of activation primitive Relu, Max: core size is 3x3, stride 2；
Third layer convolutional layer: convolution kernel size is 3x3, stride 1, port number 384, activation primitive Relu；4th layer of convolution
Layer: convolution kernel size is 5x5, stride 1, port number 256, activation primitive Relu；Layer 5 convolutional layer: convolution kernel is big
Small is 3x3, stride 1, port number 256, activation primitive Relu；The full articulamentum neuron nodal point number of first layer is
2048；The full articulamentum neuron nodal point number of the second layer is the sum of Hash code bit number and classification number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910249642.8A CN110059206A (en) | 2019-03-29 | 2019-03-29 | A kind of extensive hashing image search method based on depth representative learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910249642.8A CN110059206A (en) | 2019-03-29 | 2019-03-29 | A kind of extensive hashing image search method based on depth representative learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110059206A true CN110059206A (en) | 2019-07-26 |
Family
ID=67317886
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910249642.8A Pending CN110059206A (en) | 2019-03-29 | 2019-03-29 | A kind of extensive hashing image search method based on depth representative learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110059206A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751267A (en) * | 2019-09-30 | 2020-02-04 | 京东城市(北京)数字科技有限公司 | Neural network structure searching method, training method, device and storage medium |
| CN110929099A (en) * | 2019-11-28 | 2020-03-27 | 杭州趣维科技有限公司 | Short video frame semantic extraction method and system based on multitask learning |
| CN111274430A (en) * | 2020-01-19 | 2020-06-12 | 易拍全球(北京)科贸有限公司 | Porcelain field image retrieval algorithm based on feature reconstruction supervision |
| CN111274445A (en) * | 2020-01-20 | 2020-06-12 | 山东建筑大学 | Similar video content retrieval method and system based on triple deep learning |
| CN111339342A (en) * | 2019-11-08 | 2020-06-26 | 深圳北航新兴产业技术研究院 | Three-dimensional model retrieval method based on angle ternary center loss |
| CN111460200A (en) * | 2020-03-04 | 2020-07-28 | 西北大学 | Image retrieval method and model based on multitask deep learning and construction method thereof |
| CN111860620A (en) * | 2020-07-02 | 2020-10-30 | 苏州富鑫林光电科技有限公司 | Multilayer hierarchical neural network architecture system for deep learning |
| CN112000827A (en) * | 2020-08-27 | 2020-11-27 | 广州搜料亿网络科技有限公司 | Hardware image retrieval method and system based on deep learning |
| CN112329921A (en) * | 2020-11-11 | 2021-02-05 | 浙江大学 | Diuretic dose reasoning device based on deep characterization learning and reinforcement learning |
| CN112528066A (en) * | 2020-12-18 | 2021-03-19 | 广东工业大学 | Trademark retrieval method and system based on attention mechanism, computer equipment and storage medium |
| CN115062180A (en) * | 2022-08-15 | 2022-09-16 | 阿里巴巴（中国）有限公司 | Object query method, electronic device and storage medium |
| CN115292535A (en) * | 2022-08-24 | 2022-11-04 | 合肥市正茂科技有限公司 | Hierarchical vehicle image retrieval method based on depth polarization Hash |
| CN117076454A (en) * | 2023-08-21 | 2023-11-17 | 广州地铁集团有限公司 | Engineering quality acceptance form data structured storage method and system |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103226585A (en) * | 2013-04-10 | 2013-07-31 | 大连理工大学 | Self-adaptation Hash rearrangement method for image retrieval |
| CN105512289A (en) * | 2015-12-07 | 2016-04-20 | 郑州金惠计算机***工程有限公司 | Image retrieval method based on deep learning and Hash |
| CN106649688A (en) * | 2016-12-16 | 2017-05-10 | 深圳市华尊科技股份有限公司 | Image retrieval method and terminal |
| CN106980641A (en) * | 2017-02-09 | 2017-07-25 | 上海交通大学 | The quick picture retrieval system of unsupervised Hash and method based on convolutional neural networks |
| CN107273478A (en) * | 2017-06-09 | 2017-10-20 | 华东师范大学 | A kind of semi-supervised hashing image searching method based on Group Lasso |
| CN107992611A (en) * | 2017-12-15 | 2018-05-04 | 清华大学 | The high dimensional data search method and system of hash method are distributed based on Cauchy |
| CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
| CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
| CN109871461A (en) * | 2019-02-13 | 2019-06-11 | 华南理工大学 | The large-scale image sub-block search method to be reordered based on depth Hash network and sub-block |
-
2019
- 2019-03-29 CN CN201910249642.8A patent/CN110059206A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103226585A (en) * | 2013-04-10 | 2013-07-31 | 大连理工大学 | Self-adaptation Hash rearrangement method for image retrieval |
| CN105512289A (en) * | 2015-12-07 | 2016-04-20 | 郑州金惠计算机***工程有限公司 | Image retrieval method based on deep learning and Hash |
| CN106649688A (en) * | 2016-12-16 | 2017-05-10 | 深圳市华尊科技股份有限公司 | Image retrieval method and terminal |
| CN106980641A (en) * | 2017-02-09 | 2017-07-25 | 上海交通大学 | The quick picture retrieval system of unsupervised Hash and method based on convolutional neural networks |
| CN107273478A (en) * | 2017-06-09 | 2017-10-20 | 华东师范大学 | A kind of semi-supervised hashing image searching method based on Group Lasso |
| CN107992611A (en) * | 2017-12-15 | 2018-05-04 | 清华大学 | The high dimensional data search method and system of hash method are distributed based on Cauchy |
| CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
| CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
| CN109871461A (en) * | 2019-02-13 | 2019-06-11 | 华南理工大学 | The large-scale image sub-block search method to be reordered based on depth Hash network and sub-block |
Non-Patent Citations (4)
| Title |
|---|
| JIAN ZHANG 等: "SSDH: Semi-supervised Deep Hashing for Large Scale Image Retrieval", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》 * |
| YUEFU ZHOU 等: "Deep hashing with triplet quantization loss", 《2017 IEEE VISUAL COMMUNICATIONS AND IMAGE PROCESSING (VCIP)》 * |
| 彭天强 等: "基于深度卷积神经网络和二进制哈希学习的图像检索方法", 《电子与信息学报》 * |
| 闭蓉: "基于深度哈希的图片检索***的设计与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751267B (en) * | 2019-09-30 | 2021-03-30 | 京东城市(北京)数字科技有限公司 | Neural network structure searching method, training method, device and storage medium |
| CN110751267A (en) * | 2019-09-30 | 2020-02-04 | 京东城市(北京)数字科技有限公司 | Neural network structure searching method, training method, device and storage medium |
| CN111339342A (en) * | 2019-11-08 | 2020-06-26 | 深圳北航新兴产业技术研究院 | Three-dimensional model retrieval method based on angle ternary center loss |
| CN111339342B (en) * | 2019-11-08 | 2023-05-05 | 深圳北航新兴产业技术研究院 | Three-dimensional model retrieval method based on angle ternary center loss |
| CN110929099A (en) * | 2019-11-28 | 2020-03-27 | 杭州趣维科技有限公司 | Short video frame semantic extraction method and system based on multitask learning |
| CN111274430A (en) * | 2020-01-19 | 2020-06-12 | 易拍全球(北京)科贸有限公司 | Porcelain field image retrieval algorithm based on feature reconstruction supervision |
| CN111274445A (en) * | 2020-01-20 | 2020-06-12 | 山东建筑大学 | Similar video content retrieval method and system based on triple deep learning |
| CN111460200A (en) * | 2020-03-04 | 2020-07-28 | 西北大学 | Image retrieval method and model based on multitask deep learning and construction method thereof |
| CN111860620A (en) * | 2020-07-02 | 2020-10-30 | 苏州富鑫林光电科技有限公司 | Multilayer hierarchical neural network architecture system for deep learning |
| CN112000827A (en) * | 2020-08-27 | 2020-11-27 | 广州搜料亿网络科技有限公司 | Hardware image retrieval method and system based on deep learning |
| CN112329921A (en) * | 2020-11-11 | 2021-02-05 | 浙江大学 | Diuretic dose reasoning device based on deep characterization learning and reinforcement learning |
| CN112329921B (en) * | 2020-11-11 | 2023-11-14 | 浙江大学 | Diuretic dose reasoning equipment based on deep characterization learning and reinforcement learning |
| CN112528066A (en) * | 2020-12-18 | 2021-03-19 | 广东工业大学 | Trademark retrieval method and system based on attention mechanism, computer equipment and storage medium |
| CN112528066B (en) * | 2020-12-18 | 2023-08-04 | 广东工业大学 | Trademark retrieval method, system, computer device and storage medium based on attention mechanism |
| CN115062180A (en) * | 2022-08-15 | 2022-09-16 | 阿里巴巴（中国）有限公司 | Object query method, electronic device and storage medium |
| CN115292535A (en) * | 2022-08-24 | 2022-11-04 | 合肥市正茂科技有限公司 | Hierarchical vehicle image retrieval method based on depth polarization Hash |
| CN117076454A (en) * | 2023-08-21 | 2023-11-17 | 广州地铁集团有限公司 | Engineering quality acceptance form data structured storage method and system |
| CN117076454B (en) * | 2023-08-21 | 2024-03-12 | 广州地铁集团有限公司 | Engineering quality acceptance form data structured storage method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059206A (en) | A kind of extensive hashing image search method based on depth representative learning | |
| Wang et al. | Enhancing sketch-based image retrieval by cnn semantic re-ranking | |
| CN106407352B (en) | Traffic image search method based on deep learning | |
| Cai et al. | Medical image retrieval based on convolutional neural network and supervised hashing | |
| CN109299342B (en) | Cross-modal retrieval method based on cycle generation type countermeasure network | |
| Ma et al. | Multi-level correlation adversarial hashing for cross-modal retrieval | |
| CN106227851B (en) | The image search method of depth of seam division search based on depth convolutional neural networks | |
| CN108920720B (en) | Large-scale image retrieval method based on depth hash and GPU acceleration | |
| CN105912611B (en) | A kind of fast image retrieval method based on CNN | |
| CN108280187B (en) | Hierarchical image retrieval method based on depth features of convolutional neural network | |
| CN111198959A (en) | Two-stage image retrieval method based on convolutional neural network | |
| Xia et al. | Exploiting deep features for remote sensing image retrieval: A systematic investigation | |
| Huang et al. | Unconstrained multimodal multi-label learning | |
| CN109241317A (en) | Based on the pedestrian's Hash search method for measuring loss in deep learning network | |
| CN108427740B (en) | Image emotion classification and retrieval algorithm based on depth metric learning | |
| CN110457514A (en) | A kind of multi-tag image search method based on depth Hash | |
| CN103617609B (en) | Based on k-means non-linearity manifold cluster and the representative point choosing method of graph theory | |
| Yang et al. | Local label descriptor for example based semantic image labeling | |
| CN109933682A (en) | A kind of image Hash search method and system based on semanteme in conjunction with content information | |
| CN114882351B (en) | Multi-target detection and tracking method based on improved YOLO-V5s | |
| CN112035689A (en) | Zero sample image hash retrieval method based on vision-to-semantic network | |
| Xu et al. | Weakly supervised facial expression recognition via transferred DAL-CNN and active incremental learning | |
| CN113076490B (en) | Case-related microblog object-level emotion classification method based on mixed node graph | |
| Zhong et al. | Snake: Shape-aware neural 3d keypoint field | |
| CN112149556B (en) | Face attribute identification method based on deep mutual learning and knowledge transfer |
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: 20190726 |
|
| RJ01 | Rejection of invention patent application after publication |