CN104504007B - The acquisition methods and system of a kind of image similarity - Google Patents
The acquisition methods and system of a kind of image similarity Download PDFInfo
- Publication number
- CN104504007B CN104504007B CN201410755506.3A CN201410755506A CN104504007B CN 104504007 B CN104504007 B CN 104504007B CN 201410755506 A CN201410755506 A CN 201410755506A CN 104504007 B CN104504007 B CN 104504007B
- Authority
- CN
- China
- Prior art keywords
- image
- value
- similarity
- pixel
- characteristic values
- 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.)
- Active
Links
Classifications
-
- 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
- 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
- G06F16/5854—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 shape and object relationship
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)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (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)
Abstract
The invention discloses a kind of acquisition methods of image similarity and system, is related to picture search technical field, and technical key point includes：Target image is extracted respectively and by than the GIST characteristic values of image, LBP characteristic values, HSV characteristic values and neural network characteristics value；Calculate each characteristic value of target image and the similarity by each characteristic value than image；Calculation formula：Similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ neural network characteristics similarity × a4 is merged, obtains target image with being merged similarity than image；Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4=1.
Description
Technical field
The present invention relates to picture search technical field, more particularly to a kind of image similarity acquisition methods based on content.
Background technology
Picture search based on content is by image procossing, feature extraction, using the search image of certain Similarity algorithm
Technology.This technology and the maximum difference of the semantic label of traditional artificial addition are that the former searches out the answer come more
It is objective, it is not necessary to which that intentional spending a lot of time does semantic label with energy for every image in database.
The picture search principle for being currently based on content is the color, texture, object geometrical relationship included in itself using image
Stated etc. underlying image feature, features above is extracted by distinctive algorithm, then goes to retrieve using different criterion functions
Image library, the sequence of most like image is extracted.
The specific method of the current main flow of picture search based on content is such, using machine learning method, this kind of side
Method is divided into the object and scene of the boundless universe as Ganlei is other, then artificial extraction characteristics of image, using supervised learning algorithm
Practise grader.The disadvantage one of this kind of method is impossible real world all to be classified, and the degree of accuracy is not also high；Second, no matter
Consider from computer resource and on the training time, expense is all very big.
The content of the invention
The technical problems to be solved by the invention are：For above-mentioned problem, there is provided a kind of image similarity obtains
Method and system, extract characteristic value with reference to machine and merge similarity based method with the characteristic value acquisition manually extracted, for based on content
Picture search the retrieval foundation with higher degree of recalling and the degree of accuracy is provided.
A kind of acquisition methods of image similarity provided by the invention, comprise the following steps：
Step 1：GIST characteristic values, LBP characteristic values, HSV characteristic values and the neural network characteristics of target image are extracted respectively
Value；Extract respectively by than the GIST characteristic values of image, LBP characteristic values, HSV characteristic values and neural network characteristics value；
Step 2：The GIST characteristic values of target image are calculated with obtaining GIST by the similarity of the GIST characteristic values than image
Characteristic similarity；The LBP characteristic values of target image are calculated with obtaining LBP feature phases by the similarity of the LBP characteristic values than image
Like degree；The HSV characteristic values of target image are calculated with obtaining HSV characteristic similarities by the similarity of the HSV characteristic values than image；Meter
The neural network characteristics value of target image is calculated with obtaining neural network characteristics by the similarity of the neural network characteristics value than image
Similarity；
Step 3：According to formula：
Merge similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ god
Through network characterization similarity × a4
Target image is calculated with being merged similarity than image；
Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4
=1.
Further, extract target image or comprised the following steps by the method for the GIST characteristic values than image：
Select 5 yardsticks and the GABOR cores in 8 directions；
Image and above-mentioned 40 GABOR cores are taken turns doing into the image after convolution obtains 40 processing；
Image after each processing is divided into b × b fritter, and the gray average for calculating each fritter obtains 40 × b × b
Individual gray average, the vector of 40 × b × b gray average composition are the GIST characteristic values of 40 × b × b dimensions；
Wherein b value is 3 or 4.
Further, extract target image or comprised the following steps by the method for the LBP characteristic values than image：
The calculation window of the pixel of 3 pixels × 3 is set；
Calculation window is added on image, compares and puts whether pixel value is more than central point pixel value around calculation window,
If being set to 1 more than by the pixel put around this, 0 is otherwise set to, the pixel value of point around 8 is arranged in order to obtain one 8
The signless binary number in position, the new pixel value using this unsigned binary number as central point；
Whole image is traveled through using calculation window, obtains the new pixel value of each non-edge pixels point in image, and will be each
The new pixel value of edge pixel point is set to certain value, that is, obtains the image after LBP is handled；
Image after LBP is handled is divided into c × c fritter, and calculates each fritter grey level histogram respectively, it is described
Gray scale in grey level histogram is represented using 64 ranks, just obtains the pixel number being distributed in each fritter on 64 GTGs；By c ×
The pixel number being distributed on 64 GTGs of c fritter, which is arranged in order, finally gives the LBP characteristic values that c × c × 64 are tieed up；
Wherein c value is 3 or 4.
Further, extract target image or comprised the following steps by the method for the neural network characteristics value than image：
Choose three-level neutral net；
The value window of d pixels × d pixels is set；
D × d pixel is repeatedly taken at random in the picture using value window；Nerve is brought into successively with the pixel of taking-up
The input layer of network, is trained to neutral net, until neutral net output layer neuron and neutral net input
Layer neuron value difference not little Yu setting value when, then it is assumed that now the node of neutral net hidden layer is neural network characteristics value；
The input layer number of the neutral net is d × d with output layer neuron number, and input layer is neural
First number is more than hidden layer neuron number；
D is non-zero natural number.
Further, extract target image or comprised the following steps by the method for the HSV characteristic values than image：
H, S, V value of the pixel are calculated according to R, G, B value of each pixel of image；
Image is divided into e × e fritter, and calculates the color histogram of each fritter respectively, in the color histogram
Color using 90 ranks represent, just obtain the pixel number being distributed in each fritter in 90 color ranges；By e × e fritter 90
The pixel number being distributed on GTG, which is arranged in order, finally gives the HSV characteristic values that e × e × 90 are tieed up；
Wherein e value is 2 or 3.
Present invention also offers a kind of acquisition system of image similarity, including：
GIST characteristics extraction modules, for extracting target image and by the GIST characteristic values than image；
LBP characteristics extraction modules, for extracting target image and by the LBP characteristic values than image；
HSV characteristics extraction modules, for extracting target image and by the HSV characteristic values than image；
Neural network characteristics value extraction module, for extracting target image and by the neural network characteristics value than image；
GIST characteristic similarity computing modules, for calculating the GIST characteristic values of target image and by the GIST spies than image
The similarity of value indicative obtains GIST characteristic similarities；
LBP characteristic similarity computing modules, for calculating the LBP characteristic values of target image and by the LBP features than image
The similarity of value obtains LBP characteristic similarities；
HSV characteristic similarity computing modules, for calculating the HSV characteristic values of target image and by the HSV features than image
The similarity of value obtains HSV characteristic similarities；
Neural network characteristics similarity calculation module, for calculating the neural network characteristics value of target image and by than image
The similarity of neural network characteristics value obtain neural network characteristics similarity；
Similarity Fusion Module, for according to formula
Merge similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ god
Through network characterization similarity × a4
Target image is calculated with being merged similarity than image；
Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4
=1.
GIST characteristics extraction modules further comprise：
GABOR core determining modules, for selecting the GABOR cores of 5 yardsticks and 8 directions；
Convolution module, for image and above-mentioned 40 GABOR cores to be taken turns doing into the image after convolution obtains 40 processing；
GIST characteristic value calculating modules, for the image after each processing to be divided into b × b fritter, and calculate each small
The gray average of block, 40 × b × b gray average is obtained, the vector of 40 × b × b gray average composition is 40 × b × b
The GIST characteristic values of dimension；
Wherein b value is 3 or 4.
LBP characteristics extraction modules further comprise：
Calculation window setup module, for setting the calculation window of the pixel of 3 pixels × 3；
LBP processing modules, for calculation window to be added into image, whether big compare point pixel value around calculation window
In central point pixel value, if being set to 1 more than by the pixel put around this, be otherwise set to 0, by the pixel value of point around 8 according to
Secondary arrangement obtains 8 signless binary numbers, the new pixel value using this unsigned binary number as central point；Profit
Whole image is traveled through with calculation window, obtains the new pixel value of each non-edge pixels point in image, and by each edge pixel point
New pixel value be set to certain value, that is, obtain the image after LBP is handled；
LBP characteristic value calculating modules, for the image after LBP is handled to be divided into c × c fritter, and calculate respectively
Each fritter grey level histogram, the gray scale in the grey level histogram just obtain 64 ashes in each fritter using the expression of 64 ranks
The pixel number being distributed on rank；The pixel number being distributed on 64 GTGs of c × c fritter is arranged in order finally give c × c ×
The LBP characteristic values of 64 dimensions；
Wherein c value is 3 or 4.
Neural network characteristics value extraction module further comprises：
Value module, for setting the value window of d pixels × d pixels；It is multiple in the picture using value window
Take d × d pixel at random；
Neural network characteristics value computing module, the input layer of neutral net is brought into successively with the pixel of taking-up,
Neutral net is trained, until the value difference of neutral net output layer neuron and neural network input layer neuron is not less than
During setting value, then it is assumed that now the node of neutral net hidden layer is neural network characteristics value；
The neutral net is three-level neutral net, and input layer number and output layer neuron number are d
× d, and input layer number is more than hidden layer neuron number；D is non-zero natural number.
HSV characteristics extraction modules further comprise：
HSV modular converters, for calculating H, S, V value of the pixel according to R, G, B value of each pixel of image；
HSV characteristic value calculating modules, for image to be divided into e × e fritter, and the color for calculating each fritter respectively is straight
Fang Tu, the color in the color histogram is represented using 90 ranks, just obtains the pixel being distributed in each fritter in 90 color ranges
Points；The pixel number being distributed on 90 GTGs of e × e fritter is arranged in order and finally gives the HSV features that e × e × 90 are tieed up
Value；
Wherein e value is 2 or 3.
In summary, by adopting the above-described technical solution, the beneficial effects of the invention are as follows：
Extraction of the present invention to characteristics of image is except including GIST characteristic values, LBP characteristic values and HSV characteristic values, these are artificial
Outside the characteristic value of extraction, in addition to the neural network characteristics value that machine automatically extracts, overcome the characteristic value only manually extracted
When, feature selecting does not have standard, and the different feature degrees of accuracy is different, can not extract lacking for the characteristic value that image describes in itself
Point.
Present invention also offers a kind of similarity blending algorithm, and multiple characteristic values are merged according to certain weight,
Compared with original image search method, in the case of recall rate identical, the accuracy rate of the inventive method is searched than other for both
The Suo Fangfa high 5-10% of accuracy rate.
Embodiment
All features disclosed in this specification, or disclosed all methods or during the step of, except mutually exclusive
Feature and/or step beyond, can combine in any way.
Any feature disclosed in this specification, unless specifically stated otherwise, can be equivalent by other or with similar purpose
Alternative features are replaced.I.e., unless specifically stated otherwise, each feature is an example in a series of equivalent or similar characteristics
.
The invention provides a kind of acquisition methods of image similarity, comprise the following steps：
Step 1：GIST characteristic values, LBP characteristic values, HSV characteristic values and the neural network characteristics of target image are extracted respectively
Value；Extract respectively by than the GIST characteristic values of image, LBP characteristic values, HSV characteristic values and neural network characteristics value；
Step 2：The GIST characteristic values of target image are calculated with obtaining GIST by the similarity of the GIST characteristic values than image
Characteristic similarity；The LBP characteristic values of target image are calculated with obtaining LBP feature phases by the similarity of the LBP characteristic values than image
Like degree；The HSV characteristic values of target image are calculated with obtaining HSV characteristic similarities by the similarity of the HSV characteristic values than image；Meter
The neural network characteristics value of target image is calculated with obtaining neural network characteristics by the similarity of the neural network characteristics value than image
Similarity；
Step 3：According to formula：
Merge similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ god
Through network characterization similarity × a4
Target image is calculated with being merged similarity than image；
Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4
=1；Preferably, a1=a4=0.3；A2=a3=0.2.
GIST features are a kind of didactic features of biology, and itself simulates biological vision system.GIST features are by more
The multi-direction GABOR wave filter groups of yardstick are to the shape information that is obtained after image filtering.GABOR wave filter groups make in image procossing
With extensive, main purpose is the profile details for the different scale different directions for extracting image, in the present invention, scale selection 5,
Set direction 8.
In a specific embodiment, extract target image or following step is included by the method for the GIST characteristic values than image
Suddenly：
Select 5 yardsticks and the GABOR cores in 8 directions；
Image and above-mentioned 40 GABOR cores are taken turns doing into the image after convolution obtains 40 processing.In view of present invention side
Method is mainly used on smart mobile phone, and GABOR wave filters are windowed FFT in itself, in order to accelerate calculating speed, at this
Image is first transformed to frequency domain in one specific embodiment of invention, after taking turns doing product with GABOR cores, when result is switched back to
Domain obtains profile information.
Image after each processing is divided into b × b fritter, and the gray average for calculating each fritter obtains 40 × b × b
Individual gray average, the vector of 40 × b × b gray average composition are the GIST characteristic values of 40 × b × b dimensions；
Wherein b value is 3 or 4.
Wherein, extract target image or comprised the following steps by the method for the LBP characteristic values than image：
The calculation window of the pixel of 3 pixels × 3 is set；
Calculation window is added on image, compares and puts whether pixel value is more than central point pixel value around calculation window,
If being set to 1 more than by the pixel put around this, 0 is otherwise set to, the pixel value of point around 8 is arranged in order to obtain one 8
The signless binary number in position, the new pixel value using this unsigned binary number as central point.Take the pixel of point around 8
The order of value can be taken clockwise as starting point using the point in the calculation window upper left corner, can also be suitable to be put around other as starting point
Hour hands take counterclockwise, but whether which type of order valued combinations 8 bits are obtained with, as long as ensureing traversal view picture
Value sequence consensus in each calculation window of image.
Whole image is traveled through using calculation window, obtains the new pixel value of each non-edge pixels point in image, and will be each
The new pixel value of edge pixel point is set to certain value, that is, obtains the image after LBP is handled.Wherein, edge pixel point refers to figure
As the pixel on four edges, non-edge pixels point refers to the pixel in image in addition to the pixel on four edges.
Image after LBP is handled is divided into c × c fritter, and calculates each fritter grey level histogram respectively, it is described
Gray scale in grey level histogram is represented using 64 ranks, just obtains the pixel number being distributed in each fritter on 64 GTGs；By c ×
The pixel number being distributed on 64 GTGs of c fritter, which is arranged in order, finally gives the LBP characteristic values that c × c × 64 are tieed up；Wherein c's
Value is 3 or 4.The present invention is without traditional color histogram is used but extracts the 4*4 of LBP images regional area Nogata
Figure, in order to accelerate to calculate, not using 256 GTGs, using 64 GTGs.
Wherein, extract target image or comprised the following steps by the method for the neural network characteristics value than image：
The present invention is using the autoencoder (own coding neutral net) of 3 grades of neutral nets, own coding neutral net purpose
It is one function of study, making output, the neuron of the hidden layer of network is exactly the feature extracted as close as input.
Autoencoder operation principle is that the neuron number of input layer is equal with output layer number, and hidden layer neuron number must
It must be less than other two layers.Only in this way, own coding neutral net could compress redundancy, and extraction can characterize the master of characteristics of image
Want information.
The structure of the neutral net used in one embodiment of the invention is as follows：
Input layer number：8 × 8, hidden layer neuron number：25, output layer neuron number：8×8.Using
Training algorithm be back-propagation algorithm.
The value window of the pixel of 8 pixels × 8 is set.
8 × 8 pixels are repeatedly taken at random in the picture using value window；Neutral net is brought into the pixel of taking-up
Input layer, neutral net is trained, until neutral net output result with input value difference little Yu not set
During definite value, then it is assumed that the characteristic value that now node of neutral net hidden layer needs for us.
Specifically, when the mean square deviation of input value and output result is less than 0.2, then it is assumed that now neutral net hidden layer
The characteristic value that node needs for us.
Wherein, extract target image or comprised the following steps by the method for the HSV characteristic values than image：
H, S, V value of the pixel are calculated according to R, G, B value of each pixel of image.
Image is divided into e × e fritter, and calculates the color histogram of each fritter respectively, in the color histogram
Color using 90 ranks represent, just obtain the pixel number being distributed in each fritter in 90 color ranges；By e × e fritter 90
The pixel number being distributed on GTG, which is arranged in order, finally gives the HSV characteristic values that e × e × 90 are tieed up；Wherein e value is 2 or 3.
Similarity criterion selects：There are many similar functions to may determine that the similarity between two characteristic vectors, for example, it is European
Distance function, cosine function, Hausdorff functions etc..In the present invention, make by repeatedly test, final choice cosine function
For the standard of the similar judgement of image.
Such as by target image with bringing cosine function into by the GIST characteristic values than image, by both phases are calculated
Like degree.By that analogy, other characteristic similarities are obtained.
The invention is not limited in foregoing embodiment.The present invention, which expands to, any in this manual to be disclosed
New feature or any new combination, and disclose any new method or process the step of or any new combination.
Claims (10)
1. a kind of acquisition methods of image similarity, it is characterised in that comprise the following steps：
Step 1：GIST characteristic values, LBP characteristic values, HSV characteristic values and the neural network characteristics value of target image are extracted respectively；Point
Indescribably take by than the GIST characteristic values of image, LBP characteristic values, HSV characteristic values and neural network characteristics value；
Step 2：The GIST characteristic values of target image are calculated with obtaining GIST features by the similarity of the GIST characteristic values than image
Similarity；The LBP characteristic values of target image are calculated with obtaining LBP characteristic similarities by the similarity of the LBP characteristic values than image；
The HSV characteristic values of target image are calculated with obtaining HSV characteristic similarities by the similarity of the HSV characteristic values than image；Calculate mesh
The neural network characteristics value of logo image is similar to obtaining neural network characteristics by the similarity of the neural network characteristics value than image
Degree；
Step 3：According to formula：
Merge similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ nerve nets
Network characteristic similarity × a4
Target image is calculated with being merged similarity than image；
Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4=1.
A kind of 2. acquisition methods of image similarity according to claim 1, it is characterised in that extraction target image or by
Method than the GIST characteristic values of image comprises the following steps：
Select 5 yardsticks and the GABOR cores in 8 directions；
Image and above-mentioned 40 GABOR cores are taken turns doing into the image after convolution obtains 40 processing；
Image after each processing is divided into b × b fritter, and calculates the gray average of each fritter, obtains 40 × b × b
Gray average, the vector of 40 × b × b gray average composition are the GIST characteristic values of 40 × b × b dimensions；
Wherein b value is 3 or 4.
A kind of 3. acquisition methods of image similarity according to claim 1, it is characterised in that extraction target image or by
Method than the LBP characteristic values of image comprises the following steps：
The calculation window of the pixel of 3 pixels × 3 is set；
Calculation window is added on image, compares and puts whether pixel value is more than central point pixel value around calculation window, if greatly
The pixel put around by this is set to 1, is otherwise set to 0, and the pixel value of point around 8 is arranged in order to obtain 8 nothings
The binary number of symbol, the new pixel value using this unsigned binary number as central point；
Whole image is traveled through using calculation window, obtains the new pixel value of each non-edge pixels point in image, and by each edge
The new pixel value of pixel is set to certain value, that is, obtains the image after LBP is handled；
Image after LBP is handled is divided into c × c fritter, and calculates each fritter grey level histogram, the gray scale respectively
Gray scale in histogram is represented using 64 ranks, just obtains the pixel number being distributed in each fritter on 64 GTGs；By c × c
The pixel number being distributed on 64 GTGs of fritter, which is arranged in order, finally gives the LBP characteristic values that c × c × 64 are tieed up；
Wherein c value is 3 or 4.
A kind of 4. acquisition methods of image similarity according to claim 1, it is characterised in that extraction target image or by
Method than the neural network characteristics value of image comprises the following steps：
Choose three-level neutral net；
The value window of d pixels × d pixels is set；
D × d pixel is repeatedly taken at random in the picture using value window；Neutral net is brought into successively with the pixel of taking-up
Input layer, neutral net is trained, until neutral net output layer neuron and neural network input layer god
Through member value difference not little Yu setting value when, then it is assumed that now the node of neutral net hidden layer is neural network characteristics value；
The input layer number of the neutral net is d × d with output layer neuron number, and input layer is individual
Number is more than hidden layer neuron number；
D is non-zero natural number.
A kind of 5. acquisition methods of image similarity according to claim 1, it is characterised in that extraction target image or by
Method than the HSV characteristic values of image comprises the following steps：
H, S, V value of the pixel are calculated according to R, G, B value of each pixel of image；
Image is divided into e × e fritter, and calculates the color histogram of each fritter respectively, the face in the color histogram
Color is represented using 90 ranks, just obtains the pixel number being distributed in each fritter in 90 color ranges；By 90 GTGs of e × e fritter
The pixel number of upper distribution, which is arranged in order, finally gives the HSV characteristic values that e × e × 90 are tieed up；
Wherein e value is 2 or 3.
A kind of 6. acquisition system of image similarity, it is characterised in that including：
GIST characteristics extraction modules, for extracting target image and by the GIST characteristic values than image；
LBP characteristics extraction modules, for extracting target image and by the LBP characteristic values than image；
HSV characteristics extraction modules, for extracting target image and by the HSV characteristic values than image；
Neural network characteristics value extraction module, for extracting target image and by the neural network characteristics value than image；
GIST characteristic similarity computing modules, for calculating the GIST characteristic values of target image and by the GIST characteristic values than image
Similarity obtain GIST characteristic similarities；
LBP characteristic similarity computing modules, for calculating the LBP characteristic values of target image and by the LBP characteristic values than image
Similarity obtains LBP characteristic similarities；
HSV characteristic similarity computing modules, for calculating the HSV characteristic values of target image and by the HSV characteristic values than image
Similarity obtains HSV characteristic similarities；
Neural network characteristics similarity calculation module, for calculating the neural network characteristics value of target image and by the god than image
Similarity through networking character value obtains neural network characteristics similarity；
Similarity Fusion Module, for according to formula
Merge similarity=GIST characteristic similarities × a1+LBP characteristic similarities × a2+HSV characteristic similarities × a3+ nerve nets
Network characteristic similarity × a4
Target image is calculated with being merged similarity than image；
Wherein 0.1≤a1≤0.5,0.1≤a2≤0.4,0.1≤a3≤0.4,0.1≤a4≤0.5, and a1+a2+a3+a4=1.
A kind of 7. acquisition system of image similarity according to claim 6, it is characterised in that GIST characteristics extraction moulds
Block further comprises：
GABOR core determining modules, for selecting the GABOR cores of 5 yardsticks and 8 directions；
Convolution module, for image and above-mentioned 40 GABOR cores to be taken turns doing into the image after convolution obtains 40 processing；
GIST characteristic value calculating modules, for the image after each processing to be divided into b × b fritter, and calculate each fritter
Gray average, 40 × b × b gray average is obtained, the vector of 40 × b × b gray average composition is what 40 × b × b was tieed up
GIST characteristic values；
Wherein b value is 3 or 4.
A kind of 8. acquisition system of image similarity according to claim 6, it is characterised in that LBP characteristics extraction moulds
Block further comprises：
Calculation window setup module, for setting the calculation window of the pixel of 3 pixels × 3；
LBP processing modules, for calculation window to be added into image, compare and put around calculation window during whether pixel value be more than
Heart point pixel value, if being set to 1 more than by the pixel put around this, is otherwise set to 0, and the pixel value of point around 8 is arranged successively
Row obtain 8 signless binary numbers, the new pixel value using this unsigned binary number as central point；Utilize meter
Window traversal whole image is calculated, obtains the new pixel value of each non-edge pixels point in image, and by the new of each edge pixel point
Pixel value is set to certain value, that is, obtains the image after LBP is handled；
LBP characteristic value calculating modules, for the image after LBP is handled to be divided into c × c fritter, and calculate respectively each
Fritter grey level histogram, the gray scale in the grey level histogram is represented using 64 ranks, is just obtained in each fritter on 64 GTGs
The pixel number of distribution；The pixel number being distributed on 64 GTGs of c × c fritter is arranged in order and finally gives the dimension of c × c × 64
LBP characteristic values；
Wherein c value is 3 or 4.
9. the acquisition system of a kind of image similarity according to claim 6, it is characterised in that neural network characteristics value carries
Modulus block further comprises：
Value module, for setting the value window of d pixels × d pixels；It is repeatedly random in the picture using value window
Take d × d pixel；
Neural network characteristics value computing module, the input layer of neutral net is brought into successively with the pixel of taking-up, to god
It is trained through network, until the value difference of neutral net output layer neuron and neural network input layer neuron little Yu not set
During value, then it is assumed that now the node of neutral net hidden layer is neural network characteristics value；
The neutral net is three-level neutral net, and input layer number and output layer neuron number are d × d,
And input layer number is more than hidden layer neuron number；
D is non-zero natural number.
A kind of 10. acquisition system of image similarity according to claim 6, it is characterised in that HSV characteristics extraction moulds
Block further comprises：
HSV modular converters, for calculating H, S, V value of the pixel according to R, G, B value of each pixel of image；
HSV characteristic value calculating modules, for image to be divided into e × e fritter, and the color histogram of each fritter are calculated respectively
Scheme, the color in the color histogram is represented using 90 ranks, just obtains the pixel being distributed in each fritter in 90 color ranges
Number；The pixel number being distributed on 90 GTGs of e × e fritter is arranged in order and finally gives the HSV characteristic values that e × e × 90 are tieed up；
Wherein e value is 2 or 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410755506.3A CN104504007B (en) | 2014-12-10 | 2014-12-10 | The acquisition methods and system of a kind of image similarity |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410755506.3A CN104504007B (en) | 2014-12-10 | 2014-12-10 | The acquisition methods and system of a kind of image similarity |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104504007A CN104504007A (en) | 2015-04-08 |
| CN104504007B true CN104504007B (en) | 2018-01-30 |
Family
ID=52945405
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410755506.3A Active CN104504007B (en) | 2014-12-10 | 2014-12-10 | The acquisition methods and system of a kind of image similarity |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104504007B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106021603A (en) * | 2016-06-20 | 2016-10-12 | 昆明理工大学 | Garment image retrieval method based on segmentation and feature matching |
| CN106469192B (en) * | 2016-08-30 | 2021-07-30 | 北京奇艺世纪科技有限公司 | Text relevance determining method and device |
| CN108764325B (en) * | 2018-05-23 | 2022-07-08 | 腾讯科技（深圳）有限公司 | Image recognition method and device, computer equipment and storage medium |
| CN109740515B (en) * | 2018-12-29 | 2021-08-17 | 科大讯飞股份有限公司 | Evaluation method and device |
| CN110188217A (en) | 2019-05-29 | 2019-08-30 | 京东方科技集团股份有限公司 | Image duplicate checking method, apparatus, equipment and computer-readable storage media |
| CN110275975A (en) * | 2019-06-26 | 2019-09-24 | 北京深醒科技有限公司 | A kind of method for quickly retrieving of similar pictures |
| CN111259966A (en) * | 2020-01-17 | 2020-06-09 | 青梧桐有限责任公司 | Method and system for identifying homonymous cell with multi-feature fusion |
| CN114511717A (en) * | 2022-02-17 | 2022-05-17 | 广发银行股份有限公司 | Method, system, device and medium for background similarity identification |
| CN115830352B (en) * | 2023-02-20 | 2023-06-02 | 深圳中微电科技有限公司 | Image similarity comparison method, device and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2007045878A1 (en) * | 2005-10-20 | 2007-04-26 | Ge Healthcare Uk Limited | Method of processing an image |
| CN101763440A (en) * | 2010-03-26 | 2010-06-30 | 上海交通大学 | Method for filtering searched images |
| CN102883179A (en) * | 2011-07-12 | 2013-01-16 | 中国科学院计算技术研究所 | Objective evaluation method of video quality |
| CN103207986A (en) * | 2013-02-06 | 2013-07-17 | 江苏清大维森科技有限责任公司 | Face recognition method based on local binary pattern-histogram Fourier (LBP-HF) |
| CN104008551A (en) * | 2014-06-10 | 2014-08-27 | 华南农业大学 | Liberobacter asiaticum detection method based on visible light images |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2010122916A1 (en) * | 2009-04-22 | 2010-10-28 | 株式会社 日立メディコ | Magnetic resonance imaging device and method for displaying running direction of fibrous tissue |
-
2014
- 2014-12-10 CN CN201410755506.3A patent/CN104504007B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2007045878A1 (en) * | 2005-10-20 | 2007-04-26 | Ge Healthcare Uk Limited | Method of processing an image |
| CN101763440A (en) * | 2010-03-26 | 2010-06-30 | 上海交通大学 | Method for filtering searched images |
| CN102883179A (en) * | 2011-07-12 | 2013-01-16 | 中国科学院计算技术研究所 | Objective evaluation method of video quality |
| CN103207986A (en) * | 2013-02-06 | 2013-07-17 | 江苏清大维森科技有限责任公司 | Face recognition method based on local binary pattern-histogram Fourier (LBP-HF) |
| CN104008551A (en) * | 2014-06-10 | 2014-08-27 | 华南农业大学 | Liberobacter asiaticum detection method based on visible light images |
Non-Patent Citations (1)
| Title |
|---|
| 综合神经网络并基于多特征融合的人脸表情识别算法研究;范丽娟;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140215;第I138-818页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104504007A (en) | 2015-04-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104504007B (en) | The acquisition methods and system of a kind of image similarity | |
| Tang et al. | Pixel convolutional neural network for multi-focus image fusion | |
| CN110110617B (en) | Medical image segmentation method and device, electronic equipment and storage medium | |
| Li et al. | Building-a-nets: Robust building extraction from high-resolution remote sensing images with adversarial networks | |
| CN104318570B (en) | A kind of self adaptation camouflage painting method for designing based on background | |
| Torbati et al. | An efficient neural network based method for medical image segmentation | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN107680106A (en) | A kind of conspicuousness object detection method based on Faster R CNN | |
| CN108009637B (en) | Station caption segmentation method of pixel-level station caption identification network based on cross-layer feature extraction | |
| CN103207910B (en) | Image retrieval method based on hierarchical features and genetic programming relevance feedback | |
| CN107507188B (en) | Method and device for extracting image information based on machine learning | |
| CN108960260B (en) | Classification model generation method, medical image classification method and medical image classification device | |
| CN107944459A (en) | A kind of RGB D object identification methods | |
| CN111062441A (en) | Scene classification method and device based on self-supervision mechanism and regional suggestion network | |
| CN112712546A (en) | Target tracking method based on twin neural network | |
| Wang et al. | Spatiotemporal subpixel mapping of time-series images | |
| CN106778791A (en) | A kind of timber visual identity method based on multiple perceptron | |
| CN109409388B (en) | Dual-mode deep learning descriptor construction method based on graphic primitives | |
| Arbelle et al. | Dual-task ConvLSTM-UNet for instance segmentation of weakly annotated microscopy videos | |
| CN103049340A (en) | Image super-resolution reconstruction method of visual vocabularies and based on texture context constraint | |
| CN112862871A (en) | Image fusion method and device | |
| CN113111716A (en) | Remote sensing image semi-automatic labeling method and device based on deep learning | |
| CN114612709A (en) | Multi-scale target detection method guided by image pyramid characteristics | |
| JP6345520B2 (en) | Image search apparatus, image search program, and image search method | |
| CN104778683A (en) | Multi-modal image segmenting method based on functional mapping |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP02 | Change in the address of a patent holder |
Address after: 610041, Chengdu, Chengdu, Sichuan, China (Sichuan) free trade test area, Chengdu high tech Zone, middle section of Tianfu Avenue, No. 1268, 1 Patentee after: Chengdu PinGuo Digital Entertainment Ltd. Address before: 610041 C12-16 software building, Tianfu Software Park, hi tech Zone, Sichuan, Chengdu Patentee before: Chengdu PinGuo Digital Entertainment Ltd. |
|
| CP02 | Change in the address of a patent holder |