CN108345654A - A kind of image Hash search method based on semi-supervised ladder network - Google Patents
A kind of image Hash search method based on semi-supervised ladder network Download PDFInfo
- Publication number
- CN108345654A CN108345654A CN201810062346.2A CN201810062346A CN108345654A CN 108345654 A CN108345654 A CN 108345654A CN 201810062346 A CN201810062346 A CN 201810062346A CN 108345654 A CN108345654 A CN 108345654A
- Authority
- CN
- China
- Prior art keywords
- image
- supervised
- semi
- ladder network
- retrieved
- 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
- G06F16/58—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
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- 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
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Library & Information Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The image Hash search method based on semi-supervised ladder network that the invention discloses a kind of.The image and image to be retrieved chosen in database are pre-processed；Pretreated image is inputted in trained semi-supervised ladder network, feature extraction is carried out；Semi-supervised ladder network includes unsupervised ladder network and convolutional neural networks, convolutional neural networks include sequentially connected convolutional layer, pond layer, full articulamentum and hidden layer, wherein, characteristics of image is carried out Hash expression by hidden layer using Hash activation primitive, obtains condition code；Contrast characteristic's code, determines the affiliated grouping of image to be retrieved；The distance between the feature vector of image to be retrieved and image in the grouping is calculated, output is apart from immediate N images.The present invention solves in image retrieval procedure the problem that accuracy rate is low, and the time-consuming longer and sample mark of algorithm is insufficient.
Description
Technical field
The invention belongs to field of image search, more particularly to a kind of image Hash retrieval based on semi-supervised ladder network
Method.
Background technology
With the enhancing of hardware computing capability, image retrieval is each in search engine, e-commerce, video pedestrian's re-detection etc.
In kind practical problem, more extensive application has been obtained.Since image derives from a wealth of sources in internet, the data of image
Amount is abnormal huge, so best description can not all be given to each image, sample number is equivalent to for Data processing
According to more, but there is the data volume of label relatively fewer.In this case, the description of similar image is conducive to understand image.Cause
This, this respect can be solved the problems, such as to scheme to search figure, as long as there is the image high with this image similarity in existing database,
It can understand this image with the description of the image high with its similarity.People select oneself on e-commerce platform and need
When the thing to be bought, like carrying out comparison price in different shops mostly, then selects the thing price oneself liked most
Low businessman places an order.Therefore, people most like to search for like product on business platform, and with graph search in this searching method
It is the most directly and effective.Image search method is exactly the important tool for solving this series of problems.
Currently, almost all of search company and e-commerce platform, such as：Baidu, Google, Taobao, Jingdone district, sub- horse
Inferior wait all has used various image retrieval algorithms when picture search and like product detect to some extent, and achieves aobvious
The effect of work, image relevant treatment algorithm have become a very important skill of search engine and other computer realms
Art.Image retrieval is used most in internet and e-commerce venture, also produces huge economic benefit.Image retrieval side
Method is obtained for significant progress in terms of theory and practice.Although although image retrieval algorithm has caused in internet arena
Enough attention, but the time efficiency of its searching algorithm and search accuracy need to be improved.
A large amount of data in image data base are utilized in image search method based on deep learning, by neural network mould
Type extracts characteristics of image, is then stored in image data base.When image retrieval, image to be retrieved is extracted by neural network
Go out characteristics of image, then compared with the characteristics of image in database, finally exports N similar images.Do not have in deep learning
Before having rise, content-based image retrieval is mainly histograms of oriented gradients (the Histogram of by image
Oriented Gradient, HOG) feature, the features such as color histogram are compared, and dimension is relatively low, but effect is not
It is highly desirable.Due to the promotion of hardware computing capability, deep learning is quickly developed, it is demonstrated experimentally that deep learning is extracted
The characteristics of image gone out can largely retain the primitive character of image, be carried to the feature of image in the fields such as image recognition
It has been more than previous algorithm to take.Therefore, people are to carry out feature extraction, but nerve net to image using deep learning in recent years
The characteristic dimension that network extracts is higher, is unfavorable for carrying out the detection of the distance between feature in image retrieval, substantially increases figure
As the time of retrieval.For example, in neural network model AlexNet, the feature of extraction is up to 4096 dimensions, that is to say, that every photograph
The feature that piece extracts all be one 4096 dimension column vector, it is assumed that the data in image data base have it is tens of thousands of if, every time examine
During rope, calculates the distance between characteristics of image and need prodigious calculation amount, be thus extremely difficult in a relatively short period of time
Similar image is exported, the still more quantity nowadays in internet in image data base is up to million, ten million.Therefore needing will be refreshing
The characteristics of image obtained through network carries out dimensionality reduction or grouping, i.e., will inherently similar image assign in same group, schemes every time
When as retrieval, it is only necessary to first find affiliated group of image, the calculating of characteristic distance then be carried out in group, thus largely
The speed of algorithm is accelerated, can be reached in real time to user's output pixel image.
Data in image data base are not to have label, are all greatly and traditional god there is no label
All it is to be trained model parameter on the basis of having the data of label through network model, so needing to build a kind of semi-supervised
Neural network model has the data of label and major part to be instructed to neural network parameter without label data using fraction
Practice, passes through the feature for learning fraction data and most of data distribution without label data.It is main during training
If the most of potential data distribution without label data of training, fraction have the sample of label to play directive function, that is, refer to
Lead model learning to image feature.
In recent years, domestic and international researchers have been carried out certain in terms of being had label data based on increase and improving efficiency of algorithm
Research, data enhancing and Feature Dimension Reduction mainly are carried out to image data using image processing techniques, data enhancing is i.e. by instead
Turn, the operations increase such as Arbitrary Rotation has label data amount, improve the accuracy of model training to a certain extent, but
It doesn't solve the problem fundamentally.Feature Dimension Reduction mainly drops characteristics of image by traditional algorithms such as principal component analysis
Dimension, effect is not fine.The less side with raising efficiency of algorithm of label data is more efficiently solved therefore, it is necessary to a kind of
Method.
Invention content
In order to solve the technical issues of above-mentioned background technology proposes, the present invention is intended to provide a kind of being based on semi-supervised ladder net
The image Hash search method of network, solves that accuracy rate in image retrieval procedure is low, and it is insufficient that algorithm takes longer and sample mark
The problem of.
In order to achieve the above technical purposes, the technical scheme is that：
A kind of image Hash search method based on semi-supervised ladder network, includes the following steps：
(1) image and image to be retrieved chosen in database are pre-processed；
(2) pretreated image is inputted in trained semi-supervised ladder network, carries out feature extraction；Described half supervises
It includes unsupervised ladder network and convolutional neural networks to superintend and direct ladder network, and the convolutional neural networks include sequentially connected convolution
Layer, pond layer, full articulamentum and hidden layer, wherein characteristics of image is carried out Hash expression by hidden layer using Hash activation primitive, is obtained
Condition code；
(3) contrast characteristic's code, determines the affiliated grouping of image to be retrieved；
(4) the distance between the feature vector for calculating image to be retrieved and image in the grouping, exports apart from closest
N images, the most like N as retrieved images.
Further, in step (1), the pretreatment includes going mean value and whitening operation, realizes data normalization.
Further, in semi-supervised ladder network, the convolutional layer of convolutional neural networks is accessed after unsupervised ladder network,
Different neurons in convolutional layer carry out convolution operation to image, extract the feature of image different angle, then the pond for passing through n*n
Layer carries out maxpool operations to image, i.e., selects maximum value therein in the range of input feature vector n*n and be transmitted to as feature
Next layer of corresponding position to reduce the size of characteristic pattern, and reduces parameter amount.
Further, in the full articulamentum of convolutional neural networks, using neuron by different neuron institute in convolutional layer
The feature integration of extraction forms a feature vector, i.e. characteristic pattern to together；The activation primitive that full articulamentum uses linearly is corrected
Unit is as follows：
f(xj)=max (0, xj), xj=wtxi+b
Wherein, xiRepresent the input of this layer, f (xj) represent the output valve of this layer, wtFor the t layers of connection weight with t+1 layers
Value, b is deviant.
Further, in the hidden layer of convolutional neural networks, for the image in database, each neuron utilizes
Sigmoid function pair characteristics of image is normalized between 0 to 1；Then the average value p for calculating whole characteristic pattern, sets p to threshold
Characteristics of image is divided into two classifications of A, B by value, is averaged again in A, B classification respectively and threshold classification, to will each god
Four classifications are divided into according to average value through the characteristics of image in member, according to ascending point of each characteristics of image numerical value generic
It is not set as 0,1,2,3；Then the feature of all neurons in hidden layer is formed into a column vector according to sequence from top to bottom,
This column vector is condition code, the calculation amount that condition code is for reducing detection similitude.
Further, in step (3), the Hamming distance between image to be retrieved and condition code is calculated, Hamming distance is most
That close group is grouping belonging to image to be detected.
Further, in step (4), the characteristic pattern and image in grouping belonging to image to be retrieved that calculate image to be retrieved
Characteristic pattern between Euclidean distance, Euclidean distance nearest N images are most like N images.
The advantageous effect brought using above-mentioned technical proposal：
The present invention is based on image retrieval and machine learning algorithm, in conjunction with image data, in data preprocessing phase by data
Image characteristics extraction is carried out, is grouped according to characteristics of image code, it is high according to feature vector comparing result output similarity in group
Image, intrinsic dimensionality reduces recall precision greatly in condition code part solves the problems, such as image retrieval procedure.The present invention carries
A kind of new semi-supervised deep learning network model is gone out, can solve the problems, such as that label data is less, can be used for search engine, intelligence
In the fields such as energy medical treatment, e-commerce.
Description of the drawings
Fig. 1 is the implementing procedure figure of the present invention.
Specific implementation mode
Below with reference to attached drawing, technical scheme of the present invention is described in detail.
As shown in Figure 1, the implementation steps of the present invention are as follows.
Step 1, start, detecting system starts；
Step 2, the image data in database is determined；
Step 3, image data to be retrieved is determined；
Step 4, pre-processing image data, since the deep learning utilized extracts feature, so pretreatment is relatively easy, it is main
To include going the pretreatment operations such as mean value, albefaction；
Step 5, pretreated image is input to the semi-supervised rank that ladder network and convolutional neural networks (CNN) combine
In terraced network and obtain the higher characteristic pattern of dimension；
Step 6, the characteristic pattern obtained in step 5 is expressed as only including the condition code of 0,1,2,3 four value using Hash,
Image data in database is divided into different groups by this condition code, and every group of data have similitude largely；
Step 7, the image data in database is stored with its condition code that Hash is expressed in step 6 to database；
Step 8, the image data in database is deposited with the characteristic pattern extracted in its in steps of 5 semi-supervised ladder network
Store up database；
Step 9, the Hamming distance in image to be detected and database between the condition code of data is calculated, Hamming distance is nearest
That group be grouping where image to be detected；
Step 10, affiliated group of the image calculated according to step 9, in the characteristic pattern and database that calculate image to be detected
Euclidean distance in affiliated group of testing image between the characteristic pattern of image is calculated apart from nearest N (N similar images of output)
Image data；
Step 11, according to the recognition detection of step 10 as a result, N opens similar image in output database；
Step 12, detection terminates.
Embodiment is merely illustrative of the invention's technical idea, and cannot limit protection scope of the present invention with this, it is every according to
Technological thought proposed by the present invention, any change done on the basis of technical solution, each falls within the scope of the present invention.
Claims (7)
1. a kind of image Hash search method based on semi-supervised ladder network, which is characterized in that include the following steps：
(1) image and image to be retrieved chosen in database are pre-processed；
(2) pretreated image is inputted in trained semi-supervised ladder network, carries out feature extraction；The semi-supervised rank
Terraced network include unsupervised ladder network and convolutional neural networks, the convolutional neural networks include sequentially connected convolutional layer,
Pond layer, full articulamentum and hidden layer, wherein characteristics of image is carried out Hash expression by hidden layer using Hash activation primitive, obtains spy
Levy code；
(3) contrast characteristic's code, determines the affiliated grouping of image to be retrieved；
(4) the distance between the feature vector of image to be retrieved and image in the grouping is calculated, output is apart from immediate N
Image, the most like N as retrieved images.
2. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that in step
(1) in, the pretreatment includes going mean value and whitening operation, realizes data normalization.
3. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that supervised half
It superintends and directs in ladder network, the convolutional layer of convolutional neural networks, the different neurons pair in convolutional layer is accessed after unsupervised ladder network
Image carries out convolution operation, extracts the feature of image different angle, then carry out maxpool behaviour to image by the pond layer of n*n
Make, i.e., selects maximum value therein in the range of input feature vector n*n and be transmitted to next layer of corresponding position as feature, to contract
The size of small characteristic pattern, and reduce parameter amount.
4. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that in convolution
In the full articulamentum of neural network, using neuron by different neuron is extracted in convolutional layer feature integration to together, group
At a feature vector, i.e. characteristic pattern；The linear amending unit of activation primitive that full articulamentum uses is as follows：
f(xj)=max (0, xj), xj=wtxi+b
Wherein, xiRepresent the input of this layer, f (xj) represent the output valve of this layer, wtFor the t layers of connection weight with t+1 layers, b
For deviant.
5. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that in convolution
In the hidden layer of neural network, for the image in database, each neuron is standardized characteristics of image using sigmoid functions
Between to 0 to 1；Then the average value p for calculating whole characteristic pattern sets p to threshold value and characteristics of image is divided into two classes of A, B
, do not average again in A, B classification respectively and threshold classification, to by the characteristics of image in each neuron according to average value
It is divided into four classifications, 0,1,2,3 is respectively set to according to each characteristics of image numerical value generic is ascending；Then by hidden layer
In all neurons feature according to from top to bottom sequence form a column vector, this column vector is condition code.
6. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that in step
(3) in, the Hamming distance between image to be retrieved and condition code is calculated, that nearest group of Hamming distance is image to be detected
Affiliated grouping.
7. the image Hash search method based on semi-supervised ladder network according to claim 1, which is characterized in that in step
(4) in, calculate image to be retrieved characteristic pattern and image to be retrieved belonging to grouping in image characteristic pattern between Euclidean distance,
Nearest N images of Euclidean distance are that most like N opens images.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810062346.2A CN108345654A (en) | 2018-01-23 | 2018-01-23 | A kind of image Hash search method based on semi-supervised ladder network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810062346.2A CN108345654A (en) | 2018-01-23 | 2018-01-23 | A kind of image Hash search method based on semi-supervised ladder network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108345654A true CN108345654A (en) | 2018-07-31 |
Family
ID=62960850
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810062346.2A Pending CN108345654A (en) | 2018-01-23 | 2018-01-23 | A kind of image Hash search method based on semi-supervised ladder network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108345654A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191255A (en) * | 2018-09-04 | 2019-01-11 | 中山大学 | A kind of commodity alignment schemes based on the detection of unsupervised characteristic point |
| CN109344271A (en) * | 2018-09-30 | 2019-02-15 | 南京物盟信息技术有限公司 | Video portrait records handling method and its system |
| CN109933682A (en) * | 2019-01-11 | 2019-06-25 | 上海交通大学 | A kind of image Hash search method and system based on semanteme in conjunction with content information |
| CN112364831A (en) * | 2020-11-30 | 2021-02-12 | 姜培生 | Face recognition method and online education system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104112018A (en) * | 2014-07-21 | 2014-10-22 | 南京大学 | Large-scale image retrieval method |
| CN104182538A (en) * | 2014-09-01 | 2014-12-03 | 西安电子科技大学 | Semi-supervised hash based image retrieval method |
| US20150100558A1 (en) * | 2013-10-04 | 2015-04-09 | Nokia Corporation | Method, Apparatus and Computer Program Product for Similarity Determination in Multimedia Content |
| CN106777349A (en) * | 2017-01-16 | 2017-05-31 | 广东工业大学 | Face retrieval system and method based on deep learning |
| 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 |
-
2018
- 2018-01-23 CN CN201810062346.2A patent/CN108345654A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150100558A1 (en) * | 2013-10-04 | 2015-04-09 | Nokia Corporation | Method, Apparatus and Computer Program Product for Similarity Determination in Multimedia Content |
| CN104112018A (en) * | 2014-07-21 | 2014-10-22 | 南京大学 | Large-scale image retrieval method |
| CN104182538A (en) * | 2014-09-01 | 2014-12-03 | 西安电子科技大学 | Semi-supervised hash based image retrieval method |
| CN106777349A (en) * | 2017-01-16 | 2017-05-31 | 广东工业大学 | Face retrieval system and method based on deep learning |
| 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 |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191255A (en) * | 2018-09-04 | 2019-01-11 | 中山大学 | A kind of commodity alignment schemes based on the detection of unsupervised characteristic point |
| CN109191255B (en) * | 2018-09-04 | 2022-04-15 | 中山大学 | Commodity alignment method based on unsupervised feature point detection |
| CN109344271A (en) * | 2018-09-30 | 2019-02-15 | 南京物盟信息技术有限公司 | Video portrait records handling method and its system |
| CN109933682A (en) * | 2019-01-11 | 2019-06-25 | 上海交通大学 | A kind of image Hash search method and system based on semanteme in conjunction with content information |
| CN112364831A (en) * | 2020-11-30 | 2021-02-12 | 姜培生 | Face recognition method and online education system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | A deeper look at facial expression dataset bias | |
| Sandoval et al. | Two-stage deep learning approach to the classification of fine-art paintings | |
| Wen et al. | Ensemble of deep neural networks with probability-based fusion for facial expression recognition | |
| Jun et al. | Facial expression recognition based on VGGNet convolutional neural network | |
| Liu et al. | Deep supervised hashing for fast image retrieval | |
| Lin et al. | Deep learning of binary hash codes for fast image retrieval | |
| Xiang et al. | Fabric image retrieval system using hierarchical search based on deep convolutional neural network | |
| Shao et al. | Feature learning for image classification via multiobjective genetic programming | |
| Wu et al. | Online multimodal deep similarity learning with application to image retrieval | |
| Jiang et al. | Target recognition based on CNN with LeakyReLU and PReLU activation functions | |
| CN108345654A (en) | A kind of image Hash search method based on semi-supervised ladder network | |
| H Mohamed et al. | Improved white blood cells classification based on pre-trained deep learning models | |
| CN111209962A (en) | Combined image classification method based on CNN (CNN) feature extraction network) and combined heat map feature regression | |
| Aquino et al. | The effect of data augmentation on the performance of convolutional neural networks | |
| Yang et al. | An ensemble classification algorithm for convolutional neural network based on AdaBoost | |
| Wang et al. | Efficient yolo: A lightweight model for embedded deep learning object detection | |
| Salmam et al. | Fusing multi-stream deep neural networks for facial expression recognition | |
| CN112364705A (en) | Light-weight CNN expression recognition method based on multilevel feature fusion | |
| Xu et al. | Weakly supervised facial expression recognition via transferred DAL-CNN and active incremental learning | |
| Avcı et al. | Convolutional neural network designs for finger-vein-based biometric identification | |
| Agastya et al. | Convolutional neural network for pornographic images classification | |
| Patil et al. | Recognizing Plant species using Digitized leaves-A comparative study | |
| Han et al. | Not all instances contribute equally: Instance-adaptive class representation learning for few-shot visual recognition | |
| Peng et al. | A survey: Image classification models based on convolutional neural networks | |
| Hou et al. | A face detection algorithm based on two information flow block and retinal receptive field block |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180731 |