CN110347861A - A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask - Google Patents
A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask Download PDFInfo
- Publication number
- CN110347861A CN110347861A CN201910632985.2A CN201910632985A CN110347861A CN 110347861 A CN110347861 A CN 110347861A CN 201910632985 A CN201910632985 A CN 201910632985A CN 110347861 A CN110347861 A CN 110347861A
- Authority
- CN
- China
- Prior art keywords
- image
- channel
- exposure mask
- vector
- feature
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Library & Information 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)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
(Mask-based Prominent Feature Accumulation, MPFA) the unsupervised image indexing system for the cumulative notable feature based on exposure mask that the invention discloses a kind of, belongs to field of information processing.The feature of this method includes: the characteristic pattern of the VGG network of first abstract image input pre-training, generates exposure mask figure.Exposure mask figure is used to handle to extract notable feature each channel of characteristic pattern, each channel notable feature that then adds up obtains summation vector；The summation vector of figure number of active lanes is characterized for dimension again, the maximum top n channel of accumulated value is taken after sequence, obtains the final expression vector of image after being weighted the operation such as polymerization to selected channel；The expression vector of all images of vector sum database, which carries out nearest neighbor search, finally finally to be indicated to image to be retrieved using cosine similarity, is obtained and the most similar image searching result of image to be retrieved.The present invention is by combining exposure mask and notable feature accumulation method, especially for the image comprising complex background, to have great practical value so that unsupervised image retrieval effect gets a promotion.
Description
Technical field
The present invention relates to field of information processing, in particular to the cumulative notable feature (Mask-based based on exposure mask
Prominent Feature Accumulation, MPFA) unsupervised image indexing system.
Background technique
Image retrieval is a vital task of computer vision field, its target is when given one is forgiven specific reality
Example query image, the shooting visual angle of database images, illumination or circumstance of occlusion under different conditions, extract suitable image
It indicates vector, carries out nearest neighbor search in the database and find the image comprising same instance, to solve the problems, such as to scheme to search figure.
Core in unsupervised image retrieval research is how to utilize the CNN model of pre-training not by supervision messages such as artificial marks
Extract feature finally indicates as image.The quality of Feature Selection will have a direct impact on the result of retrieval.
Most of unsupervised image retrieval algorithm is generally using the method being finely adjusted to pre-training CNN network, to new
Data set is trained in the form of triple.But if requiring to be finely adjusted network after addition new images every time,
Efficiency and expansibility can all reduce.And it is each to carry out extraction feature figure using exposure mask figure in order to solve the problem above-mentioned by the present invention
The notable feature in channel, the sequence of the laggard row of channels importance of the feature in each channel of adding up, choose representative channel into
Row weighting polymerization is finally indicated as image, preferable retrieval effectiveness has been obtained, especially for the image comprising complex background.
Summary of the invention
In order to solve existing technical problem, the present invention provides the unsupervised images of the cumulative notable feature based on exposure mask
Searching system.Scheme is as follows:
Step 1, abstract image input the characteristic pattern of the VGG network of pre-training, generate exposure mask figure.Using exposure mask figure to spy
Each channel of sign figure is handled to extract notable feature, and each channel notable feature that then adds up obtains summation vector.
Step 2 is characterized dimension the summation vector of figure number of active lanes, takes the maximum top n of accumulated value logical after sequence
Road obtains the final expression vector of image after being weighted the operation such as polymerization to selected channel.
Step 3, using cosine similarity to image to be retrieved finally indicate the expressions of all images of vector sum database to
Amount carries out nearest neighbor search, obtains and the most similar image searching result of image to be retrieved.
Detailed description of the invention
Fig. 1 is the net of (MPFA) unsupervised image indexing system of the cumulative notable feature provided by the invention based on exposure mask
Network structure chart
Fig. 2 is the VGG structure chart provided by the invention for extracting depth characteristic
Specific embodiment
It next will be for a more detailed description to implementation method of the invention.
Fig. 1 is the network structure of the unsupervised image indexing system of the cumulative notable feature provided by the invention based on exposure mask
Figure, including:
Step S1: the extraction of notable feature and cumulative
Step S2: the selection in representative channel with polymerize
Step S3: nearest neighbor search
Fig. 2 gives VGG structure chart.
Each step will be specifically described below:
Step S1: the extraction of notable feature and cumulative.Notable feature is contained in the depth convolution feature extracted due to VGG
And redundancy feature, the present invention are extracted notable feature using maximum exposure mask and summation exposure mask, are shown using exposure mask figure to each channel
Feature is write to add up.While notable feature obtains retaining, redundancy feature is also inhibited.
For the data set D comprising J images, input is each image I in D；I is inputted into VGG network, from the 5th
Pond layer extracts W × H × C characteristic pattern, and W, H are that the width in each channel of characteristic pattern and height, C are characterized the channel that figure includes respectively
Number, F(i)It is expressed as i-th of channel in characteristic pattern, the feature locations in each channel are indicated with { (x, y) }；
Notable feature is extracted using maximum exposure mask and summation exposure mask.For maximum exposure mask, each channel of keeping characteristics figure is most
The position of big activation value, to capture most representative structure in all channels:
For exposure mask of summing, it is retained in the position for the feature being activated in more multichannel, t is the feature of each position each
Channel corresponds to the median of the sum of activation value:
For M ∈ { MMAX,MSUM, by the way that the value of the position retained in M is set as 1, the value of unreserved position is set as k,
The available exposure mask figure M to the notable feature in each channel in extraction feature figuref:
Wherein, k represents the influence degree for removing redundancy feature.When k is negative, inhibit the redundancy feature of background, it is prominent aobvious
Write feature.
Add up respectively to the notable feature in each channel of characteristic pattern after mask process, obtain be with number of active lanes
The summation vector I of dimensionn=[s1,…,sc]。siIt is bigger, illustrate to remain more notable features under the channel:
Step S2: the selection in representative channel with polymerize.It is characterized the summation vector of figure number of active lanes for dimension, sorts
After take the maximum top n channel of accumulated value, to selected channel be weighted polymerization etc. operation after obtain image finally indicate to
Amount.
It is all by exposure mask figures treated summation vector I after L2 is normalized for the images of the J in databasen
It is added, obtains and vector V:
Wherein, v1,v2... it respectively represents corresponding each channel and carries out the sum after notable feature adds up.
The maximum top n channel of accumulated value is taken after carrying out descending sort to element in V.For each image, for selecting
Channel F(i), weighted sum pond is done first::
Wherein wiFor normalized weight, uiFor the activation value at selected channel (x, y) position.
Wherein α, β are the parameter of power normalization and power scaling respectively.
The corresponding C dimensional feature in N number of channel of selection is indicated into splicing, obtaining the global of N × C dimension indicates vectorL2 regularization, principal component analysis (PCA) dimensionality reduction and whitening processing are carried out to ψ, obtain image fixed dimension
Final expression vector ψf。
Step S3: nearest neighbor search.Since the final expression of image to be retrieved and database images has same dimension, this
Invention selection cosine similarity retrieves arest neighbors.
To all images of database, PiFor the final expression vector of i-th image；For image to be retrieved, Q is to choose together
The final expression vector obtained behind the N number of channel of sample.Calculating image to be retrieved indicates that vector Q and database image indicate vector Pi's
Cosine similarity:
Finally in test, search result of the highest image of similarity as image to be retrieved is chosen.
Above in conjunction with attached drawing to the unsupervised image retrieval system of the cumulative notable feature (MPFA) based on exposure mask proposed
The specific embodiment of system and each module is expounded.Pass through the description of embodiment of above, the general technology of fields
Personnel can be understood that the present invention can add the mode of necessary general hardware platform to realize by software, certainly may be used
With by hardware realization, but the former is more preferably embodiment.Based on this understanding, technical solution of the present invention substantially or
Person says that the part that contributes to existing technology can embody in form of a computer software product, which is stored in one
In a storage medium, including some instructions are used so that one or more computer equipment executes described in each embodiment of the present invention
Method.
According to the thought of the present invention, there will be changes in the specific implementation manner and application range.In conclusion this
Description should not be construed as limiting the invention.
Invention described above embodiment does not constitute the restriction to invention protection scope.It is any of the invention
Made modifications, equivalent substitutions and improvements etc., should all be included in the protection scope of the present invention within spirit and principle.
Claims (4)
1. a kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask, which is characterized in that described image retrieval
System includes with flowering structure and step:
(1) extraction of notable feature and cumulative: abstract image inputs the characteristic pattern of the VGG network of pre-training, generates exposure mask figure.It adopts
Each channel of characteristic pattern is handled to extract notable feature with exposure mask figure, each channel notable feature that then adds up obtain summing to
Amount.
(2) selection in representative channel with polymerize: for the summation vector of step (1), maximum top n channel is taken after sequence,
The final expression vector of image is obtained after being weighted the operation such as polymerization to selected channel.
(3) nearest neighbor search: using cosine similarity to the final table of expression vector sum image to be retrieved of all images of database
Show that vector carries out nearest neighbor search, obtains the most similar image searching result.
2. the method as described in claim 1, which is characterized in that the step (1) specifically includes:
(1.1) for the data set D comprising T images, input is each image in D；
(1.2) input is by CNN, from the 5th pond layer extraction feature figure；
(1.3) position of each channel maximum activation value of keeping characteristics figure generates maximum exposure mask, is retained in more multichannel and is swashed
The position of feature living generates summation exposure mask；
(1.4) value is set separately with unreserved position to the reservation in maximum exposure mask and summation exposure mask, obtains exposure mask figure, makes
With the notable feature in each channel in exposure mask figure extraction feature figure；
(1.5) it adds up, obtains with number of active lanes respectively to the notable feature in each channel of characteristic pattern after mask process
For the summation vector of dimension.
3. the method as described in claim 1, which is characterized in that the step (2) specifically includes:
(2.1) the summation vector obtained for step (1) carries out descending sort and takes maximum top n channel；
(2.2) for image each in database, polymerization will be weighted after the corresponding N number of channel splicing of respective characteristic pattern；
(2.3) vector obtained for weighting polymerization, carries out L2 regularization, PCA dimensionality reduction and albefaction, obtains image fixed dimension
It is final to indicate vector.
4. the method as described in claim 1, which is characterized in that the step (3) specifically includes:
(3.1) the expression vector of all images of database is obtained by step (2)；
(3.2) for image to be retrieved, same N number of channel is chosen to obtain finally indicating vector；
(3.3) calculating image to be retrieved indicates that vector sum database images indicate the cosine similarity of vector, carries out figure to be retrieved
The nearest neighbor search of picture chooses the highest image of similarity as search result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910632985.2A CN110347861A (en) | 2019-07-12 | 2019-07-12 | A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910632985.2A CN110347861A (en) | 2019-07-12 | 2019-07-12 | A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110347861A true CN110347861A (en) | 2019-10-18 |
Family
ID=68176273
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910632985.2A Pending CN110347861A (en) | 2019-07-12 | 2019-07-12 | A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110347861A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027634A (en) * | 2019-12-16 | 2020-04-17 | 中山大学 | Regularization method and system based on class activation mapping graph guidance |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102508917A (en) * | 2011-11-21 | 2012-06-20 | 洪涛 | Multi-dimensional object robust high-speed retrieval and positioning method for some feature images |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
| CN109753578A (en) * | 2019-01-25 | 2019-05-14 | 浙江理工大学 | A kind of image search method based on exposure mask selection convolution feature |
| US20190171906A1 (en) * | 2017-12-04 | 2019-06-06 | Adobe Inc. | Digital Image Search Based On Arbitrary Image Features |
-
2019
- 2019-07-12 CN CN201910632985.2A patent/CN110347861A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102508917A (en) * | 2011-11-21 | 2012-06-20 | 洪涛 | Multi-dimensional object robust high-speed retrieval and positioning method for some feature images |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
| US20190171906A1 (en) * | 2017-12-04 | 2019-06-06 | Adobe Inc. | Digital Image Search Based On Arbitrary Image Features |
| CN109753578A (en) * | 2019-01-25 | 2019-05-14 | 浙江理工大学 | A kind of image search method based on exposure mask selection convolution feature |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111027634A (en) * | 2019-12-16 | 2020-04-17 | 中山大学 | Regularization method and system based on class activation mapping graph guidance |
| CN111027634B (en) * | 2019-12-16 | 2023-04-18 | 中山大学 | Regularization method and system based on class activation mapping image guidance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhong et al. | An end-to-end dense-inceptionnet for image copy-move forgery detection | |
| Yan et al. | Deep learning for imbalanced multimedia data classification | |
| Tieu et al. | Boosting image retrieval | |
| Lin et al. | Fast color-spatial feature based image retrieval methods | |
| Zakariya et al. | Combining visual features of an image at different precision value of unsupervised content based image retrieval | |
| CN107562742A (en) | A kind of image processing method and device | |
| CN110163117B (en) | Pedestrian re-identification method based on self-excitation discriminant feature learning | |
| Wang et al. | Multiscale deep alternative neural network for large-scale video classification | |
| Staszewski et al. | A new approach to descriptors generation for image retrieval by analyzing activations of deep neural network layers | |
| CN109325507A (en) | A kind of image classification algorithms and system of combination super-pixel significant characteristics and HOG feature | |
| CN109871905A (en) | A kind of plant leaf identification method based on attention mechanism depth model | |
| CN112597324A (en) | Image hash index construction method, system and equipment based on correlation filtering | |
| Zhang et al. | A lightweight transformer network for hyperspectral image classification | |
| CN111612024A (en) | Feature extraction method and device, electronic equipment and computer-readable storage medium | |
| CN109492610A (en) | A kind of pedestrian recognition methods, device and readable storage medium storing program for executing again | |
| Alzu'Bi et al. | Compact root bilinear cnns for content-based image retrieval | |
| Chen et al. | Image classification based on convolutional denoising sparse autoencoder | |
| Liu et al. | Image retrieval using CNN and low-level feature fusion for crime scene investigation image database | |
| CN114764941A (en) | Expression recognition method and device and electronic equipment | |
| CN110209865A (en) | A kind of object identification and matching process based on deep learning | |
| CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN110347861A (en) | A kind of unsupervised image indexing system of the cumulative notable feature based on exposure mask | |
| Zhang et al. | Semantic hierarchy preserving deep hashing for large-scale image retrieval | |
| Zhou et al. | Superpixel attention guided network for accurate and real-time salient object detection | |
| Rahman et al. | A flexible multi-layer self-organizing map for generic processing of tree-structured data |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191018 |