CN107330074A - The image search method encoded based on deep learning and Hash - Google Patents
The image search method encoded based on deep learning and Hash Download PDFInfo
- Publication number
- CN107330074A CN107330074A CN201710525604.1A CN201710525604A CN107330074A CN 107330074 A CN107330074 A CN 107330074A CN 201710525604 A CN201710525604 A CN 201710525604A CN 107330074 A CN107330074 A CN 107330074A
- Authority
- CN
- China
- Prior art keywords
- image
- retrieval
- hash
- value
- deep learning
- 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
- 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
- G06F16/5866—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using information manually generated, e.g. tags, keywords, comments, manually generated location and time information
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Library & Information Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of model training method encoded based on deep learning and Hash, including will part mark view data as network model training data, the training data is expressed as into class two-value Hash by depth network to encode, wherein, the class two-value Hash coding refers to that value is a kind of simulation two-value Hash coding of successive value；Using the class two-value Hash coding of acquisition as input, one or more task layers of depth network are connected to, while being trained using one or more tasks；The two-value Hash coding for being used to represent the training data for obtaining the characteristic information that retrieval is available for described in is encoded based on class two-value Hash.
Description
Technical field
The present invention relates to technical field of computer vision, more particularly to a kind of image encoded based on deep learning and Hash
Search method.
Background technology
With the development of science and technology, the world today has been enter into the big data epoch, and especially view data resource growth is fast
Speed, therefore large-scale image data is retrieved bring new choose to image retrieval technologies field to meet user's request
War.Compared to traditional text based image retrieval technologies (Text-Based Image Retrieval, TBIR), based on interior
The image retrieval (Content-Based Image Retrieval, CBIR) of appearance is more and more widely paid close attention to by people.
In CBIR technologies, how effectively to describe the feature of image and which kind of mode is quick similitude is carried out using
Retrieval is study hotspot in recent years.Because superiority of the deep neural network on feature learning and Hash coding are in retrieval
In calculating speed and memory space on superiority, occur in that using depth convolutional neural networks, or salted hash Salted, or the two
The image search method being combined.
For example, a kind of image search method based on depth network extraction feature, utilizes the depth convolutional network trained
To image zooming-out feature, the Euclidean distance of characteristics of image and sequence are carried out in the feature and database by calculating query image
Image retrieval.Reference papers " Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor
Lempitsky.Neural Codes for Image Retrieval.ECCV 2014”.The defect of this method is, a side
Face, the feature extracted using this method is the higher real number vector of dimension, therefore storage overhead and amount of calculation are larger, it is impossible to met
The demand of current network database size rapid development；On the other hand, the depth network that this method extracts feature is not for number
It is trained according to storehouse data, retrieval effectiveness depends critically upon similar between the data that database data and training network are used
Degree, if similarity degree is relatively low, accordingly can cause retrieval effectiveness poor.
In the prior art, also a kind of image search method based on many perceptual property retrieval types, using perceptual property it
Between association, train the joint classification device of multiple perceptual properties to carry out the perceptual property that prognostic chart picture has.During retrieval, according to user
Association between given retrieval type and known perceptual property, builds a new retrieval type, according to image in database
Perceptual property and the matching degree of retrieval type are retrieved.Reference papers " Behjat Siddiquie, Rogerio S.Feris,
and Larry S.Davis.Image Ranking and Retrieval based on Multi-Attribute
Queries.CVPR 2011”.The defect of this method is, on the one hand, only trained during due to training in perceptual property data,
This method cannot be used directly for other retrieval tasks, limit its application prospect；On the other hand, when it is desirable that on increase database
Perceptual property when, the model of joint training can not be directly extended in new perceptual property, it is necessary to start all over again
Training, so as to limit the scalability of this method.
In addition, China Patent Publication No. CN105512273A also discloses a kind of figure learnt based on variable length depth Hash
As search method, this method is trained using image triple to depth network, and target is to learn one with allowing network end-to-end
Individual two-value Hash coding so that similar image has similar coding, and the encoding variability of dissimilar image is larger.This method
Defect is, on the one hand due to that can only use a kind of similarity measurement, therefore the two-value Hash coding finally given in training
Single retrieval tasks are only used for, the application of this method is limited；On the other hand, this method has used figure in training
As triple, so that model convergence is slower when causing to train, time consumption for training is big.
Therefore, need at present a kind of quick effective and extend flexible image search method.
The content of the invention
It is an object of the invention to provide a kind of image search method encoded based on deep learning and Hash, this method can
Overcome the defect of above-mentioned prior art.
There is provided a kind of model training method encoded based on deep learning and Hash, bag according to an aspect of the present invention
Include following steps：
Step 1), using the view data of part mark as the training data of network model, will be described by depth network
Training data is expressed as class two-value Hash coding, wherein, the class two-value Hash coding refers to that value is a kind of mould of successive value
Intend two-value Hash coding；
Step 2), using the step 1) the class two-value Hash coding that obtains, as inputting, is connected to one of depth network
Or multiple tasks layer, while being trained using one or more tasks；
Step 3), based on the step 1) class two-value Hash coding obtain with the characteristic information for being available for retrieval
For representing that the two-value Hash of the training data is encoded.
It is preferred that, the step 2) one or more task layers refer to can as image retrieval task task layer.
It is preferred that, described image retrieval tasks refer to carry out image retrieval for the semantic classes of image.
It is preferred that, for the semantic classes of described image, classification task or the metric learning based on image pair can be used
Task is trained.
It is preferred that, described image retrieval tasks are the perceptual property progress image retrievals for image.
It is preferred that, for the perceptual property, one group of perceptual property grader can be trained.
It is preferred that, using the network model, can to mark not exclusively or the view data that does not mark to enter row label pre-
Survey, so that the attribute tags of all images in completion described image data.
Encoded according to another aspect of the present invention there is provided a kind of based on deep learning and Hash for any of the above-described
The method of image retrieval, including：
When carrying out semantic classes retrieval tasks according to a query image, the query graph is obtained using the network model
The two-value Hash coding of picture；Pass through two of all images in the two-value Hash coding and image data base by the query image
Value Hash coding compares, and obtains and is used as retrieval result with the query image semantic classes identical image；Or
During according to the perceptual property information of one or more query images as retrieval tasks, the network model root is utilized
According to the corresponding perceptual property information of all images in the two-value Hash coding restoring data storehouse of image, obtaining has the vision
The image of attribute information is used as retrieval result；Or
Appointed according to the semantic classes and one or more perceptual property information specified of a query image as retrieval
During business, the corresponding vision of all images in restoring data storehouse is encoded according to the two-value Hash of image first with the network model
Attribute information, and database is screened using the perceptual property information of the reduction；Secondly obtained using the network model
Two-value Hash to the query image is encoded；By the way that the two-value Hash of the query image is encoded, filtered out with described
The two-value Hash coding of image in image data base compares, and obtains and the query image semantic classes identical and has
The image for the perceptual property specified is as retrieval result.
According to another aspect of the present invention there is provided a kind of image indexing system, including memory, processor and it is stored in
On reservoir and the computer program that can run on a processor, wherein, perform above-mentioned figure during the processor operation described program
As the step of search method.
According to another aspect of the present invention there is provided a kind of computer-readable recording medium, including it is stored in described readable deposit
Computer program on storage media, wherein, described program is performed such as the step of above-mentioned image search method.
Relative to prior art, the present invention achieves following advantageous effects：The present invention based on deep learning and Kazakhstan
The image search method of uncommon coding, with existing to extract the image search method that real number vector is characterized compared with, significantly drop
Amount of calculation when low searching system between the demand and image of memory space to being compared to each other, retrieval effectiveness is good, and model training is fast,
The scale that current network database goes from strength to strength can be better met；Meanwhile, the image search method that provides of the present invention can be with
For multiple different retrieval tasks, have a extensive future, scalability is good.
Brief description of the drawings
Fig. 1 is that the overall procedure framework for the image search method encoded based on deep learning and Hash that the present invention is provided is shown
It is intended to；
Fig. 2 is the coordinate indexing application schematic diagram of the present invention.
Embodiment
In order that the purpose of the present invention, technical scheme and advantage are more clearly understood, below in conjunction with accompanying drawing, to according to this
The image search method encoded based on deep learning and Hash provided in the embodiment of invention is illustrated.
Deep learning comes from artificial neural network, and in field of image search, deep learning is capable of the spy of combination image bottom
Levy and to form higher and represent that such as attribute classification feature is represented with the distributed nature for finding view data；Hash coding is one
The algorithm with quick search ability and low memory cost is planted, can be by image using Hash coding in field of image search
Hold and be expressed as the Hash sequence of two-value, and represent with the sequence feature of image.
Carefully studied through inventor, it is proposed that a kind of image search method of use deep neural network, can be end-to-end
Ground study two-value Hash encodes the character representation as image.This method is instructed by using the view data of part mark
Practice, using multiple loss functions for different retrieval tasks, different information is embedded into two-value Hash coding, so that
The two-value Hash coding obtained finally can be used for multiple different retrieval tasks.
In one embodiment of the invention there is provided a kind of image search method encoded based on deep learning and Hash,
This method mainly includes data preparation, model training and image retrieval.
Fig. 1 shows the overall procedure frame for the image search method encoded based on deep learning and Hash that the present invention is provided
Frame schematic diagram, as shown in figure 1, the image search method encoded based on deep learning and Hash of the present invention is comprised the following steps：
S10. data prepare
In order that representing as image is encoded with two-value Hash, in deep neural network front end, it is necessary to using substantial amounts of
View data resource is trained to model.In the training stage, present invention employs the view data resource with part mark,
Mark herein refers to the feature tag of image, for example, label represents the objects such as the cat contained in image, dog, automobile, or it is right
Shape, the color of elephant, material etc., these feature tags both can image have in itself, such as upper figure in images share website
What the label (tag) of picture or later stage obtained by mark.In order to make it easy to understand, below with semantic classes (such as cat, dog,
Desk etc.) and perceptual property (such as red, circular, spot) two kinds of marks exemplified by illustrate.
For different types of semantic classes, if the semantic classes of an image is not known, or semantic classes mark
It is unknown, then corresponding semantic classes be labeled as being labeled as in unknown, as shown in Figure 1 training data chart "" entry；If
Semantic classes is marked, it is known that then the classification according to belonging to it is labeled, and is labeled as in training data chart as shown in Figure 1
The entry of " great white shark, fire balloon ", wherein above-mentioned classification can be labeled as, for example, for n classification, being more than using one
Positive integer equal to 1 and less than or equal to n indicates the classification belonging to it；
For different types of perceptual property, if an image has a kind of perceptual property, by being somebody's turn to do for above-mentioned image
Plant perceptual property and be labeled as positive example, if being labeled as negative example without if, if whether uncertain have, or this kind of vision belongs to
The mark of property is not provided, then is labeled as the corresponding mark in unknown, as shown in Figure 1 training data chart, the first row image
Black, striped and tail be positive example, circular and wooden to bear example, the perceptual property of the second row image is unknown etc..Therefore can
The perceptual property of every image is labeled as to the vector of multidimensional, for example, for m kind perceptual properties, every image is respectively provided with a m
The perceptual property label-vector of dimension.
In data preparation, the view data resource that there is part to mark that the present invention is used refers to, as long as an image
At least one in the semantic classes mark and perceptual property mark that are had is not " unknown ", you can the image is applied into training
Model.By using such mode, substantial amounts of available image data resource on the one hand can be made full use of, alleviates the mistake of model
Fitting problems；On the other hand the application scenarios of this method can be extended, so that for a variety of different retrieval tasks, such as simultaneously
For semantic classes retrieval, perceptual property retrieval or semantic classes and perceptual property retrieval-by-unification etc..
S20. model training
After the data for completing step S10 prepare, the training image data being ready to complete can be passed through in depth network N
A series of nonlinear operations, such as depth convolutional neural networks (CNN) convolution, Chi Hua, full connection, one can be obtained
Value is the multidimensional image character representation f of real number.
Then a nonlinear operation is carried out again to the image feature representation f of the real number.Due in the training process of network
Need to carry out the back delivery operations of gradient, therefore the activation primitive that can lead may be used herein to simulate two-value Hash coding, without
It is to be trained after jump function that direct use can not be led does two-value quantization to image feature representation f.Letter is being activated using S-shaped
During number, feature f every one-dimensional value can be compressed in a limited scope, for example, compressing during using sigmoid functions
To 0 to 1 scope, -1 to 1 scope is compressed to during using hyperbolic tangent function；
In another embodiment of the present invention, regular terms can also be used to carry out binaryzation constraint, example to character representation f
The value of output is such as constrained as close possible to ± 1.
After the completion of aforesaid operations, one can be obtained, such as dimension is k, class two-value Hash coding C0, wherein C0's
Dimension k is equal to the code length that final two-value Hash is encoded.
The value of known two-value Hash coding is strict two-value, such as 0 and 1, or -1 and 1.And above-mentioned class two-value is breathed out
Uncommon coding refers to that value is successive value, a kind of simulation two-value Hash coding of such as 0 to 1 real number or -1 to 1 real number
Characteristics of image is encoded.
Obtaining above-mentioned class two-value Hash coding C0Afterwards, by C0As input, be connected to different can appoint as image retrieval
The task layer of business, while being trained using multiple tasks.Below will using two kinds of retrieval tasks of semantic classes and perceptual property as
Example is illustrated：
For semantic classes, classification task or metric learning task based on image pair can be used to train.Using
During classification task, it is assumed that task layer corresponds respectively to n kind semantic classes comprising n node.When known to the class label of image,
The loss function for classification can be utilized, such as using softmax loss functions or hinge loss functions, to measure classification
Order of accuarcy；When image category is unknown, then ignore corresponding sample in classification task.Using metric learning task
When, by constraining the sample of identical semantic classes there is similar class two-value Hash to encode C0, different types of encoding samples are poor
It is different larger, so as to learn to be adapted to the class two-value Hash coding of retrieval tasks；
, can be by training one group of perceptual property grader come corresponding perceptual property embodying information for perceptual property
In two-value Hash coding.For example, task layer corresponds respectively to m kind perceptual properties comprising m node.When some vision of image
When known to attribute, then lost with the attribute forecast of weighting, such as sigmoid intersects entropy loss, hinge losses, measure this
Prediction order of accuarcy of the sample in the perceptual property；When some perceptual property mark of image is unknown, then regarded accordingly
Feel and ignore corresponding sample on attributive classification device.By using the loss of weighting, align negative sample and apply different weights, can be with
Alleviate the prediction deviation problem brought because positive and negative sample proportion is unbalanced to a certain extent.After the completion of costing bio disturbance, derivation
Corresponding Grad is calculated, the parameter of network model is updated by back-propagation algorithm.After successive ignition updates, net is completed
The training of network model.
The class two-value Hash for extracting database images using depth network N encodes C0Afterwards, it is necessary to using a threshold value to it
Quantified, obtain real two-value Hash coding C.Threshold value herein can be fixedly installed, and such as 0.5 or 0, also may be used
To be obtained by study.Meanwhile, the parameter A of above-mentioned perceptual property grader, for example, size is k*m matrix, is preserved
For follow-up perceptual property retrieval tasks.
S30. image retrieval
When user needs to carry out image retrieval, the class two-value Hash of query image can be calculated by depth network N first
Encode C0, then by using the threshold value consistent with the database images that set in above-mentioned steps S20 or study is obtained, will look into
Ask the class two-value Hash coding C of image0It is quantified as two-value Hash and encodes C to carry out various retrieval tasks, below will be with semanteme
Illustrated exemplified by classification retrieval tasks, perceptual property retrieval tasks, and semantic classes and perceptual property retrieval-by-unification.
Semantic classes retrieval refers to, when user gives a query image, it is necessary to which retrieval has phase in image data base
With the image of semantic classes, for example, Fig. 2 is the coordinate indexing application schematic diagram of the present invention, as shown in Fig. 2 the first rows, user gives
The image of a fixed automobile in image data base, it is necessary to retrieve all images comprising automobile., can in order to realize the function
The two-value Hash coding of image in the two-value Hash coding and database of query image containing automobile is compared and completed
Semantic classes is retrieved, for example, the two-value Hash coding and two of image in database by calculating the query image containing automobile
It is worth Hash coding Hamming distance, and retrieval result is returned according to the order of distance from small to large；
Perceptual property retrieval refers to that user specifies one or more perceptual properties as retrieval type, it is necessary in view data
Retrieval is with the image for specifying perceptual property in storehouse.For example, as shown in the rows of Fig. 2 second, user give perceptual property " white " and
" metal " in image data base, it is necessary to retrieve all images comprising above-mentioned perceptual property., can in order to realize the function
C and perceptual property grader A inner product is encoded by calculating the two-value Hash of database images, for example, using being not required to largely multiply
The look-up table of method computing realizes that binary set C and A inner product are calculated, to restore the perceptual property information of image, so as to complete
Perceptual property is retrieved, for example, be ranked up using the probability size of perceptual property；
The retrieval-by-unification of semantic classes and perceptual property refers to that user specifies a query image and specified one or more
Perceptual property is, it is necessary to which retrieval and the query image have identical semantic classes in image data base, and are specified with user
Perceptual property image.For example, as shown in Fig. 2 the third lines, the image and perceptual property " red " of the given automobile of user,
Need to retrieve in image data base and be labeled as automobile comprising semantic classes and perceptual property is labeled as all of " red "
Image.In order to realize the function, first with identical method when being retrieved with perceptual property, the perceptual property pair restored is used
All database images are screened, for example, remove the image that perceptual property is less than the threshold value according to certain threshold value；Then utilize
Identical method when being retrieved with semantic classes, compares the two-value Hash coding of query image and the database images filtered out
Compared with and complete retrieval, such as after being ranked up by calculating Hamming distance, screening is returned through by the order of distance from small to large
Image is used as retrieval result.
In another embodiment of the present invention, the network model trained in step S20 can be used, it is endless to marking
View data that is complete or not marking carries out Tag Estimation, so as to the attribute tags of completion view data, and is added into number
According in storehouse.
Although in the above-described embodiments, employing semantic classes and perceptual property to illustrate to compile based on deep learning and Hash
Code image search method, but those of ordinary skill in the art should be understood that in other embodiments can according to different demands,
Semantic classes or the training method for perceptual property are directed to by using described in above-described embodiment, other images are utilized
Feature realizes the image search method that the present invention is provided as retrieval tasks and image labeling, for example, will shoot or make image
Place etc. other have the markup information of certain correlation with image, using to the above-mentioned training pattern similar for semantic classes
Be trained, or when screening the photographs of certain style, using the photographic work of the photography style as positive example, using with
The above-mentioned training pattern similar for perceptual property is trained.
Relative to prior art, the image inspection encoded based on deep learning and Hash provided in embodiments of the present invention
Suo Fangfa, is encoded as graphical representation by using the end-to-end study two-value Hash of deep neural network, considerably reduced
Storage overhead and retrieval amount of calculation, the loss function independent of image triple that this method is used is by different characteristics of image
Information is embedded into two-value Hash coding, improves the convergence rate of network so that final two-value Hash coding can be used for
A variety of different retrieval tasks, while trained single grader for the different attribute of image, it is to avoid attributive classification device
Between too strong dependence, it is ensured that training pattern scalability.
Although the present invention be described by means of preferred embodiments, but the present invention be not limited to it is described here
Embodiment, without departing from the present invention also include made various changes and change.
Claims (10)
1. a kind of model training method encoded based on deep learning and Hash, is comprised the following steps：
Step 1), using the view data of part mark as network model training data, by depth network by the training
Data are expressed as class two-value Hash coding, wherein, the class two-value Hash coding refers to that value is a kind of simulation two of successive value
It is worth Hash coding；
Step 2), using the step 1) the class two-value Hash coding that obtains, as inputting, is connected to one or many of depth network
Individual task layer, while being trained using one or more tasks；
Step 3), based on the step 1) class two-value Hash coding obtain being used for the characteristic information for being available for retrieval
Represent the two-value Hash coding of the training data.
2. the model training method according to claim 1 encoded based on deep learning and Hash, the step 2) one
Individual or multiple tasks layer refer to can as image retrieval task task layer.
3. the model training method according to claim 2 encoded based on deep learning and Hash, described image retrieval is appointed
Business refers to carry out image retrieval for the semantic classes of image.
4. the model training method according to claim 3 encoded based on deep learning and Hash, for described image
Semantic classes, can use classification task or metric learning task based on image pair to train.
5. the model training method according to claim 2 encoded based on deep learning and Hash, described image retrieval is appointed
Business is the perceptual property progress image retrieval for image.
6. the model training method encoded based on deep learning and Hash according to claim 5, for the vision
Attribute, can train one group of perceptual property grader.
7. the model training method encoded based on deep learning and Hash according to any one of claim 1 to 6, utilizes institute
Network model is stated, Tag Estimation can be carried out to the view data for marking endless all or none mark, so that completion described image number
The attribute tags of all images in.
8. a kind of be used for the method for the image retrieval encoded based on deep learning and Hash such as any one of claim 1 to 7,
Including：
When carrying out semantic classes retrieval tasks according to a query image, the query image is obtained using the network model
Two-value Hash is encoded；By the way that the two-value of the two-value Hash coding and all images in image data base of the query image is breathed out
Uncommon coding compares, and obtains and is used as retrieval result with the query image semantic classes identical image；Or
During according to the perceptual property information of one or more query images as retrieval tasks, using the network model according to figure
The corresponding perceptual property information of all images in the two-value Hash coding restoring data storehouse of picture, obtaining has the perceptual property
The image of information is used as retrieval result；Or
During according to the semantic classes and one or more perceptual property information specified of a query image as retrieval tasks,
The corresponding vision that all images in restoring data storehouse are encoded according to the two-value Hash of image first with the network model belongs to
Property information, and database is screened using the perceptual property information of the reduction；Secondly obtained using the network model
The two-value Hash coding of the query image；By the way that the two-value Hash of the query image is encoded, with the figure filtered out
As the two-value Hash coding of the image in database compares, obtain with the query image semantic classes identical and with phase
The image for the perceptual property answered is as retrieval result.
9. a kind of image indexing system, including memory, processor and storage can be run on a memory and on a processor
Computer program, wherein, perform step as claimed in claim 8 during the processor operation described program.
10. a kind of computer-readable recording medium, including the computer program being stored on the readable storage medium storing program for executing, wherein,
Described program performs step as claimed in claim 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710525604.1A CN107330074B (en) | 2017-06-30 | 2017-06-30 | Image retrieval method based on deep learning and Hash coding |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710525604.1A CN107330074B (en) | 2017-06-30 | 2017-06-30 | Image retrieval method based on deep learning and Hash coding |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107330074A true CN107330074A (en) | 2017-11-07 |
| CN107330074B CN107330074B (en) | 2020-05-26 |
Family
ID=60199596
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710525604.1A Active CN107330074B (en) | 2017-06-30 | 2017-06-30 | Image retrieval method based on deep learning and Hash coding |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107330074B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN108280451A (en) * | 2018-01-19 | 2018-07-13 | 北京市商汤科技开发有限公司 | Semantic segmentation and network training method and device, equipment, medium, program |
| CN108427738A (en) * | 2018-03-01 | 2018-08-21 | 中山大学 | A kind of fast image retrieval method based on deep learning |
| CN108549915A (en) * | 2018-04-27 | 2018-09-18 | 成都考拉悠然科技有限公司 | Image hash code training pattern algorithm based on two-value weight and classification learning method |
| CN108829912A (en) * | 2018-04-16 | 2018-11-16 | 浙江工业大学 | A kind of circuit input vector characterization method based on APHash |
| CN108984642A (en) * | 2018-06-22 | 2018-12-11 | 西安工程大学 | A kind of PRINTED FABRIC image search method based on Hash coding |
| CN109002463A (en) * | 2018-06-05 | 2018-12-14 | 国网辽宁省电力有限公司信息通信分公司 | A kind of Method for text detection based on depth measure model |
| CN109241317A (en) * | 2018-09-13 | 2019-01-18 | 北京工商大学 | Based on the pedestrian's Hash search method for measuring loss in deep learning network |
| CN109543772A (en) * | 2018-12-03 | 2019-03-29 | 北京锐安科技有限公司 | Data set automatic matching method, device, equipment and computer readable storage medium |
| CN110135582A (en) * | 2019-05-09 | 2019-08-16 | 北京市商汤科技开发有限公司 | Neural metwork training, image processing method and device, storage medium |
| CN111581420A (en) * | 2020-04-30 | 2020-08-25 | 徐州医科大学 | Medical image real-time retrieval method based on Flink |
| CN111866443A (en) * | 2019-04-25 | 2020-10-30 | 黄河 | Video stream data storage method, device, system and storage medium |
| CN112949722A (en) * | 2021-03-05 | 2021-06-11 | 北京大学深圳研究生院 | Few-sample learning method and system based on image combination feature expression |
| WO2021232752A1 (en) * | 2020-05-22 | 2021-11-25 | 深圳前海微众银行股份有限公司 | Hash encoding method, apparatus and device, and readable storage medium |
| CN114372205A (en) * | 2022-03-22 | 2022-04-19 | 腾讯科技（深圳）有限公司 | Training method, device and equipment of characteristic quantization model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105469096A (en) * | 2015-11-18 | 2016-04-06 | 南京大学 | Feature bag image retrieval method based on Hash binary code |
| CN105512273A (en) * | 2015-12-03 | 2016-04-20 | 中山大学 | Image retrieval method based on variable-length depth hash learning |
| CN106126581A (en) * | 2016-06-20 | 2016-11-16 | 复旦大学 | Cartographical sketching image search method based on degree of depth study |
| CN106407352A (en) * | 2016-09-06 | 2017-02-15 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Traffic image retrieval method based on depth learning |
-
2017
- 2017-06-30 CN CN201710525604.1A patent/CN107330074B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105469096A (en) * | 2015-11-18 | 2016-04-06 | 南京大学 | Feature bag image retrieval method based on Hash binary code |
| CN105512273A (en) * | 2015-12-03 | 2016-04-20 | 中山大学 | Image retrieval method based on variable-length depth hash learning |
| CN106126581A (en) * | 2016-06-20 | 2016-11-16 | 复旦大学 | Cartographical sketching image search method based on degree of depth study |
| CN106407352A (en) * | 2016-09-06 | 2017-02-15 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Traffic image retrieval method based on depth learning |
Non-Patent Citations (1)
| Title |
|---|
| HAOMIAO LIU等: ""Deep Supervised Hashing for Fast Image Retrieval"", 《THE IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107992611B (en) * | 2017-12-15 | 2018-12-28 | 清华大学 | The high dimensional data search method and system of hash method are distributed based on Cauchy |
| 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 |
| CN108280451A (en) * | 2018-01-19 | 2018-07-13 | 北京市商汤科技开发有限公司 | Semantic segmentation and network training method and device, equipment, medium, program |
| CN108427738A (en) * | 2018-03-01 | 2018-08-21 | 中山大学 | A kind of fast image retrieval method based on deep learning |
| CN108829912A (en) * | 2018-04-16 | 2018-11-16 | 浙江工业大学 | A kind of circuit input vector characterization method based on APHash |
| CN108549915B (en) * | 2018-04-27 | 2021-06-15 | 成都考拉悠然科技有限公司 | Image hash code training model algorithm based on binary weight and classification learning method |
| CN108549915A (en) * | 2018-04-27 | 2018-09-18 | 成都考拉悠然科技有限公司 | Image hash code training pattern algorithm based on two-value weight and classification learning method |
| CN109002463A (en) * | 2018-06-05 | 2018-12-14 | 国网辽宁省电力有限公司信息通信分公司 | A kind of Method for text detection based on depth measure model |
| CN108984642A (en) * | 2018-06-22 | 2018-12-11 | 西安工程大学 | A kind of PRINTED FABRIC image search method based on Hash coding |
| CN108984642B (en) * | 2018-06-22 | 2021-07-27 | 西安工程大学 | Printed fabric image retrieval method based on Hash coding |
| CN109241317A (en) * | 2018-09-13 | 2019-01-18 | 北京工商大学 | Based on the pedestrian's Hash search method for measuring loss in deep learning network |
| CN109543772A (en) * | 2018-12-03 | 2019-03-29 | 北京锐安科技有限公司 | Data set automatic matching method, device, equipment and computer readable storage medium |
| CN111866443A (en) * | 2019-04-25 | 2020-10-30 | 黄河 | Video stream data storage method, device, system and storage medium |
| CN110135582A (en) * | 2019-05-09 | 2019-08-16 | 北京市商汤科技开发有限公司 | Neural metwork training, image processing method and device, storage medium |
| CN111581420A (en) * | 2020-04-30 | 2020-08-25 | 徐州医科大学 | Medical image real-time retrieval method based on Flink |
| WO2021232752A1 (en) * | 2020-05-22 | 2021-11-25 | 深圳前海微众银行股份有限公司 | Hash encoding method, apparatus and device, and readable storage medium |
| CN112949722A (en) * | 2021-03-05 | 2021-06-11 | 北京大学深圳研究生院 | Few-sample learning method and system based on image combination feature expression |
| CN112949722B (en) * | 2021-03-05 | 2023-05-05 | 北京大学深圳研究生院 | Image combination feature expression-based few-sample learning method and system |
| CN114372205A (en) * | 2022-03-22 | 2022-04-19 | 腾讯科技（深圳）有限公司 | Training method, device and equipment of characteristic quantization model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107330074B (en) | 2020-05-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107330074A (en) | The image search method encoded based on deep learning and Hash | |
| CN109165306B (en) | Image retrieval method based on multitask Hash learning | |
| Warburg et al. | Mapillary street-level sequences: A dataset for lifelong place recognition | |
| CN110188227B (en) | Hash image retrieval method based on deep learning and low-rank matrix optimization | |
| Jiang et al. | Learning discriminative latent attributes for zero-shot classification | |
| Yu et al. | Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop | |
| Ogawa et al. | Object detection for comics using manga109 annotations | |
| CN109241317B (en) | Pedestrian Hash retrieval method based on measurement loss in deep learning network | |
| CN108427738B (en) | Rapid image retrieval method based on deep learning | |
| CN109902672B (en) | Image labeling method and device, storage medium and computer equipment | |
| US11640714B2 (en) | Video panoptic segmentation | |
| CN107729312B (en) | Multi-granularity word segmentation method and system based on sequence labeling modeling | |
| CN103390063B (en) | A kind of based on ant group algorithm with the search method of related feedback images of probability hypergraph | |
| CN110059807A (en) | Image processing method, device and storage medium | |
| CN109522436A (en) | Similar image lookup method and device | |
| CN103324954B (en) | Image classification method based on tree structure and system using same | |
| CN113177141B (en) | Multi-label video hash retrieval method and device based on semantic embedded soft similarity | |
| CN114358188A (en) | Feature extraction model processing method, feature extraction model processing device, sample retrieval method, sample retrieval device and computer equipment | |
| Lagunas et al. | Transfer learning for illustration classification | |
| Chen et al. | Recognizing the style of visual arts via adaptive cross-layer correlation | |
| CN114329181A (en) | Question recommendation method and device and electronic equipment | |
| CN103942214B (en) | Natural image classification method and device on basis of multi-modal matrix filling | |
| CN115457332A (en) | Image multi-label classification method based on graph convolution neural network and class activation mapping | |
| CN110909174B (en) | Knowledge graph-based method for improving entity link in simple question answering | |
| Rahman et al. | Deep multiple instance learning for zero-shot image tagging |
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 |