CN110413824A - A kind of search method and device of similar pictures - Google Patents
A kind of search method and device of similar pictures Download PDFInfo
- Publication number
- CN110413824A CN110413824A CN201910534899.8A CN201910534899A CN110413824A CN 110413824 A CN110413824 A CN 110413824A CN 201910534899 A CN201910534899 A CN 201910534899A CN 110413824 A CN110413824 A CN 110413824A
- Authority
- CN
- China
- Prior art keywords
- picture
- similarity
- measured
- target photo
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
- G06F16/538—Presentation of query results
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Abstract
The invention discloses a kind of search method of similar pictures and devices, are related to technical field of image processing, and the Search Requirement deviation to solve the problems, such as search result in the prior art and user is larger and invents.This method specifically includes that the global feature information that Target Photo is extracted according to edge detection method;According to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated;According to MaskRCNN model, the minutia information of Target Photo is extracted;According to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated;According to the overall similarity, the details similarity and preset retrieval weight, the retrieval similarity of the picture to be measured and the Target Photo is calculated;According to the sequence of the numerical value for retrieving similarity from big to small, picture to be measured corresponding with the retrieval similarity is shown.Present invention is mainly applied to during picture retrieval.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to the search method and device of a kind of similar pictures.
Background technique
It is a kind of for the purpose of obtaining information needed to search similar pictures as means using picture as information storage means
New retrieval information mode.In the prior art, the method for retrieving similar pictures includes: to obtain multiple conspicuousness areas of Target Photo
Domain;Extract the convolutional neural networks CNN feature of multiple salient regions;According to the CNN feature of multiple salient regions, mesh is obtained
It marks on a map the feature vector of piece;According to the feature vector of Target Photo, multiple candidates for including from the candidate picture group of Target Photo
It is obtained and the matched similar pictures of Target Photo in picture.
Convolutional neural networks CNN (Convolutional Neural Networks, CNN) is a kind of comprising convolutional calculation
And the feedforward neural network with depth structure.Convolutional neural networks copy the visual perception mechanism construction of biology, can be supervised
Educational inspector practises and unsupervised learning, and the sparsity that the convolution kernel parameter sharing in hidden layer is connected with interlayer makes convolutional Neural net
Network can reveal feature with lesser calculation amount plaid matching and be learnt, and be widely used in computer vision, natural language processing
Equal fields.When retrieving similar pictures, user may require to look up that color is close, style is close, text is close or brand phase
Same commodity.And the method for existing retrieval similar pictures, it is the CNN aspect ratio of the salient region based on whole picture to looking into
It looks for, has ignored the non-limiting feature such as brand trademark and text, the Search Requirement deviation for resulting in search result and user is larger.
Summary of the invention
In view of this, the present invention provides the search method and device of a kind of similar pictures, main purpose is to solve existing
Search result and the larger problem of the Search Requirement deviation of user in technology.
According to the present invention on one side, a kind of search method of similar pictures is provided, comprising:
According to edge detection method, the global feature information of Target Photo is extracted;
According to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated;
According to Mask RCNN model, the minutia information of the Target Photo is extracted;
According to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated;
According to the overall similarity, the details similarity and preset retrieval weight, the picture to be measured and institute are calculated
State the retrieval similarity of Target Photo;
According to the sequence of the numerical value for retrieving similarity from big to small, show corresponding to be measured with the retrieval similarity
Picture.
According to the present invention on the other hand, a kind of retrieval device of similar pictures is provided, comprising:
Extraction module, for extracting the global feature information of Target Photo according to edge detection method;
First computing module, for according to the global feature information, calculating the whole of picture to be measured and the Target Photo
Body similarity;
The extraction module, for extracting the minutia information of the Target Photo according to Mask RCNN model;
First computing module is also used to calculate the picture to be measured and the mesh according to the minutia information
It marks on a map the details similarity of piece;
Second computing module, for calculating according to the overall similarity, the details similarity and preset retrieval weight
The retrieval similarity of the picture to be measured and the Target Photo;
Display module is shown and the retrieval phase for the sequence of the numerical value according to the retrieval similarity from big to small
Like the corresponding picture to be measured of degree.
According to another aspect of the invention, a kind of storage medium is provided, at least one is stored in the storage medium can
It executes instruction, the executable instruction makes processor execute the corresponding operation of search method such as above-mentioned similar pictures.
In accordance with a further aspect of the present invention, a kind of computer equipment is provided, comprising: processor, memory, communication interface
And communication bus, the processor, the memory and the communication interface complete mutual lead to by the communication bus
Letter;
For the memory for storing an at least executable instruction, it is above-mentioned that the executable instruction executes the processor
The corresponding operation of the search method of similar pictures.
By above-mentioned technical proposal, technical solution provided in an embodiment of the present invention is at least had the advantage that
The present invention provides a kind of search method of similar pictures and devices, first according to edge detection method, extract target
The global feature information of picture calculates the overall similarity of picture and Target Photo to be measured, further according to Mask RCNN model, mentions
The minutia information for taking the Target Photo calculates the details of picture and Target Photo to be measured further according to minutia information
Similarity calculates the retrieval of picture and Target Photo to be measured further according to overall similarity, details similarity and preset retrieval weight
Similarity, finally the sequence according to the numerical value of retrieval similarity from big to small, shows picture to be measured corresponding with retrieval similarity.
Compared with prior art, different preset retrievals is arranged by using for overall similarity and details similarity in the embodiment of the present invention
Weight, to calculate the retrieval similarity of picture and Target Photo to be measured.By increasing details similarity to the shadow of retrieval similarity
It rings, the details conflict of similar object can be distinguished, improve the discrimination of similar pictures, to improve recognition effect, reduce retrieval
And the deviation of user search demand as a result.
The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention,
And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects of the present invention, feature and advantage can
It is clearer and more comprehensible, the followings are specific embodiments of the present invention.
Detailed description of the invention
By reading the following detailed description of the preferred embodiment, various other advantages and benefits are common for this field
Technical staff will become clear.The drawings are only for the purpose of illustrating a preferred embodiment, and is not considered as to the present invention
Limitation.And throughout the drawings, the same reference numbers will be used to refer to the same parts.In the accompanying drawings:
Fig. 1 shows a kind of search method flow chart of similar pictures provided in an embodiment of the present invention;
Fig. 2 shows the search method flow charts of another similar pictures provided in an embodiment of the present invention;
Fig. 3 shows a kind of retrieval device composition block diagram of similar pictures provided in an embodiment of the present invention;
Fig. 4 shows the retrieval device composition block diagram of another similar pictures provided in an embodiment of the present invention;
Fig. 5 shows a kind of structural schematic diagram of computer equipment provided in an embodiment of the present invention.
Specific embodiment
Exemplary embodiments of the present disclosure are described in more detail below with reference to accompanying drawings.Although showing the disclosure in attached drawing
Exemplary embodiment, it being understood, however, that may be realized in various forms the disclosure without should be by embodiments set forth here
It is limited.On the contrary, these embodiments are provided to facilitate a more thoroughly understanding of the present invention, and can be by the scope of the present disclosure
It is fully disclosed to those skilled in the art.
The embodiment of the invention provides a kind of search methods of similar pictures, as shown in Figure 1, this method comprises:
101, according to edge detection method, the global feature information of Target Photo is extracted.
Target Photo refers to the commodity picture that user has inquired, and the object of the invention is to the similar diagrams of searched targets figure
Piece.Global feature information refers to the Global Information of object in targeted graphical, including color and profile.Assuming that Target Photo content
For the white housing of an A brand, then white housing is global feature information.
Data volume can significantly be reduced using edge detection method, and reject incoherent information, while reserved graph
As important structure attribute, be conducive to extract global feature information.Edge detection method can be divided into two classes: first derivative is searched
Method and second dervative zero pass through method.It is detected based on the method for lookup by finding the maximum and minimum value in image first derivative
Boundary, usually by boundary alignment in the maximum direction of gradient.The method passed through based on zero is by finding image second order derivative zero
It passes through to find boundary, usually Laplacian zero crossing or the zero crossing of nonlinear difference expression.Illustratively, target
Image content is the white housing of an A brand, extracts the shape of housing and the color of housing.
102, according to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated.
Picture to be measured, referring to may all picture similar with Target Photo, it may be possible to all pictures in system, it may
It is the picture concerned retrieved according to the verbal description of Target Photo, it is also possible to according to the attribute retrieval of object in Target Photo
The relational graph arrived.The image credit and quantity for treating mapping piece in embodiments of the present invention are without limitation.With acquisition target
The method of the global feature information of picture is identical, obtains the global feature information of picture to be measured, then calculates picture and mesh to be measured
It marks on a map the overall similarity of piece.Overall similarity can according to Euclidean distance, manhatton distance, Minkowski distance, it is remaining
The methods of string similarity or Pearson correlation coefficients calculate.
103, according to Mask RCNN model, the minutia information of the Target Photo is extracted.
Target Photo refers to the commodity picture that user has inquired, and the object of the invention is to the similar diagrams of searched targets figure
Piece.Minutia information refers to the detail section in Target Photo, and referring to relative to whole picture proportion is smaller but has
The image section of abundant information.Assuming that Target Photo content is the white housing of an A brand, then A brand is minutia letter
Breath, A brand refer to the trade mark shape of the brand or the shape and color of trade mark.
If extracting minutia information using Mask RCNN algorithm, the algorithm model of Mask RCNN algorithm is needed
It to be trained by marking out the training picture of Target Photo minutia information.Illustratively, Target Photo content is one
The white housing of part A brand marks out the shape of trade mark and the shape of general housing in training image training, is made by training
Trade mark on housing and outer set can be distinguished by obtaining Target Photo.
104, according to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated.
It is identical as the minutia information approach of Target Photo is obtained, the minutia information of picture to be measured is obtained, then
Calculate the details similarity of picture and Target Photo to be measured.Details similarity can be according to Euclidean distance, manhatton distance, bright
Can the methods of Paderewski distance, cosine similarity or Pearson correlation coefficients calculate.
105, according to the overall similarity, the details similarity and preset retrieval weight, the picture to be measured is calculated
With the retrieval similarity of the Target Photo.
Preset retrieval weight refers to overall similarity and details the similarity ratio shared when calculating retrieval similarity.
Assuming that preset retrieval weight is 1:2, the corresponding weight of overall similarity is 1, and the corresponding weight of details similarity is 2, retrieves phase
Be overall similarity multiplied by 1 like degree, with details similarity multiplied by 2 and.When calculating retrieval similarity, preset inspection can be set
Suo Quanchong and be 1, it is assumed that preset retrieval weight be 1:2, the corresponding weight of overall similarity be 1/3, details similarity pair
The weight answered is 2/3, and retrieval similarity is overall similarity multiplied by 1/3, with details similarity multiplied by 2/3 and.
106, the sequence according to the numerical value of the retrieval similarity from big to small, shows corresponding with the retrieval similarity
Picture to be measured.
The numerical value for retrieving similarity is sorted from large to small, according to its put in order lookup with retrieval similarity it is corresponding to
Then mapping piece shows picture to be measured.When showing, according to the picture number size of show area setting while shown.It is opening up
When showing, can putting in order according to picture to be measured, choose the picture number that can show simultaneously of show area, show picture to be measured.
The present invention provides a kind of search methods of similar pictures, first according to edge detection method, extract Target Photo
Global feature information calculates the overall similarity of picture and Target Photo to be measured, further according to Mask RCNN model, described in extraction
The minutia information of Target Photo calculates the details similarity of picture and Target Photo to be measured further according to minutia information,
Further according to overall similarity, details similarity and preset retrieval weight, the retrieval similarity of picture and Target Photo to be measured is calculated,
The finally sequence according to the numerical value of retrieval similarity from big to small, shows picture to be measured corresponding with retrieval similarity.With it is existing
Technology is compared, and different preset retrieval weights is arranged by using for overall similarity and details similarity in the embodiment of the present invention,
To calculate the retrieval similarity of picture and Target Photo to be measured.By increasing influence of the details similarity to retrieval similarity, energy
The details conflict for enough distinguishing similar object, improves the discrimination of similar pictures, to improve recognition effect, reduce search result with
The deviation of user search demand.
The embodiment of the invention provides the search methods of another similar pictures, as shown in Fig. 2, this method comprises:
201, according to edge detection method, the global feature information of Target Photo is extracted.
Target Photo refers to user with the commodity picture inquired, and the object of the invention is to the similar diagrams of searched targets figure
Piece.Global feature information refers to the Global Information of object in targeted graphical, including color and profile.Assuming that Target Photo content
For the white housing of an A brand, then white housing is global feature information.
Existing clothes, shoes and hats, electronic product and other items include usually that both sides is similar when searching like product,
First is that minutia is identical, second is that global feature is similar.Global feature information includes target object color and object edge profile.
The global feature information for obtaining Target Photo, specifically includes: using edge detection method, identifying and extracts in the Target Photo
The object edge profile of target object;Extract the target object face of the target object in the object edge profile
Color.Image Edge-Detection significantly reduces data volume, and eliminates it is considered that incoherent information, remains image
Important structure attribute.It is extracted by edge detection algorithm and identifies object edge profile.Object edge profile is Target Photo
The profile of middle target object, so the color in object edge profile, is the actual color of target object.
The target object color for extracting target object in object edge profile, specifically includes: according to the first preset division grain
Degree, is divided into multiple grid pictures for the Target Photo;In the object edge profile described in the Target Photo, inquiry is every
The mesh color of a grid picture, the mesh color include single color and secondary colour;It is mixed for calculating the mesh color
The secondary colour ratio that the total quantity of the picture number and the grid picture that close color is compared;If the secondary colour ratio is greater than the
One preset proportion then repartitions the Target Photo according to the second preset granularity of division;If the secondary colour ratio is little
In first preset proportion, then the single chromatic graph piece of each color for the grid picture that the mesh color is single color is recorded
Quantity;If the single color ratio example of the total quantity of the single color picture number and the grid picture is greater than the second preset ratio
Example, it is determined that the corresponding mesh color of the single color picture number is target object color;If the list of each color
Difference between picture number of the same colour is less than third preset quantity, it is determined that the mesh color of the grid picture is target object
Color.Determining target object color, it may be possible to some individual color, it is also possible to multiple color combination.
202, according to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated.
Picture to be measured, referring to may all picture similar with Target Photo, it may be possible to all pictures in system, it may
It is the picture concerned retrieved according to the verbal description of Target Photo, it is also possible to according to the attribute retrieval of object in Target Photo
The relational graph arrived.The image credit and quantity for treating mapping piece in embodiments of the present invention are without limitation.This step is specific
It include: the global feature information for obtaining the picture to be measured, the global feature information of the picture to be measured includes object under test face
Color and edge contour to be measured;According to the global feature information of the picture to be measured, the target object color and the target side
Edge profile calculates the object color similarity and edge contour similarity of the Target Photo Yu the picture to be measured;According to pre-
Whole specific gravity is set, the overall similarity of the Target Photo Yu the picture to be measured is calculated.Preset entirety specific gravity, refers to object
Body color similarity and the edge contour similarity specific gravity shared when calculating overall similarity.
203, according to Mask RCNN model, the minutia information of the Target Photo is extracted.
Target Photo refers to user with the commodity picture inquired, and the object of the invention is to the similar diagrams of searched targets figure
Piece.Minutia information refers to the detail section in Target Photo, and referring to relative to whole picture proportion is smaller but has
The image section of abundant information.Assuming that Target Photo content is the white housing of an A brand, then A brand is minutia letter
Breath.
Existing clothes, shoes and hats, electronic product and other items include usually that both sides is similar when searching like product,
First is that minutia is identical, second is that global feature is similar.Minutia information includes work mark, pattern trademark, stamp, embroidery
Equal special graphs, usually or colour contrast similar with the integral color of targeted graphical is larger, is distinguished using this thin as obtaining
Save the basis of characteristic information.The minutia information for extracting Target Photo, specifically includes: extracting the mesh color is secondary colour
Grid picture;The target object color in the grid picture that the mesh color is secondary colour is filtered out, grid search-engine figure is obtained
Piece;Judge whether the picture profile in the grid search-engine picture is closed outline;If it is judged that be it is yes, then by the net
Character or graph outline in lattice feature image are determined as minutia information;If it is judged that be it is no, then merge described in
Mesh color is the grid picture and its adjacent grid picture of secondary colour, and filters out the target object color, reacquires
The grid search-engine picture.Merge the grid picture and its adjacent grid picture that mesh color is secondary colour, refers to grid
Color is the grid picture merging adjacent with its surrounding centered on the grid picture of secondary colour.
In order to improve the speed and accuracy of extracting minutia information, Target Photo minutia information is extracted, specifically
Include: to match the Target Photo with the object edge profile, screens the target pictorial diagram in the Target Photo
Piece;The target material object picture is inputted into the Mask RCNN model, extracts the two-value exposure mask figure of the target material object picture
Picture;Mark the image outline in the two-value mask image;By the two-value exposure mask after the target material object picture and the label
Image is matched, and the minutia picture in target material object picture corresponding with described image profile is screened;By the grid
Character or graph outline in feature image are determined as minutia information.Accelerated by Mask RCNN model to target figure
The segmentation of piece, in the case of realizing lower Time & Space Complexity, the accurate image outline extracted in target material object picture, with
Obtain accurate minutia information.
204, according to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated.
This step specifically includes: obtaining the minutia information of the picture to be measured;According to the details of the picture to be measured
It is similar to the details of the picture to be measured to calculate the Target Photo for the minutia information of characteristic information and the Target Photo
Degree.It is identical as the minutia information approach of Target Photo is obtained, obtain the minutia information of picture to be measured.
205, according to the overall similarity, the details similarity and preset retrieval weight, the picture to be measured is calculated
With the retrieval similarity of the Target Photo.
Preset retrieval weight refers to overall similarity and details the similarity ratio shared when calculating retrieval similarity.
Assuming that preset retrieval weight is 1:2, then the corresponding weight of overall similarity is 1, and the corresponding weight of details similarity is 2, retrieval
Similarity is overall similarity multiplied by 1, with details similarity multiplied by 2 and.When calculating retrieval similarity, can be set preset
Retrieve weight and be 1, it is assumed that preset retrieval weight be 1:2, then the corresponding weight of overall similarity be 1/3, details is similar
Spending corresponding weight is 2/3, and retrieval similarity is overall similarity multiplied by 1/3, with details similarity multiplied by 2/3 and.
206, the sequence according to the numerical value of the retrieval similarity from big to small, shows corresponding with the retrieval similarity
Picture to be measured.
The numerical value for retrieving similarity is sorted from large to small, according to its put in order lookup with retrieval similarity it is corresponding to
Then mapping piece shows picture to be measured.When showing, according to the picture number size of show area setting while shown.It is opening up
When showing, can putting in order according to picture to be measured, choose the picture number that can show simultaneously of show area, show picture to be measured.
If 207, user's operation meets prerequisite, according to presetting rule, the preset retrieval weight is corrected.
Prerequisite is that user does not choose the corresponding picture to be measured of the retrieval similarity maximum value for the first time, or is arranged to work as and use
The picture to be detected chosen for the first time after the similar pictures of different target picture of the family by lookup is not the highest picture of similarity
Number be greater than predetermined times.When correcting preset retrieval weight, the picture to be measured that can compare user's selection is similar to retrieval
The overall similarity of the corresponding picture to be measured of maximum value and the size of details similarity are spent, determines overall similarity or details phase
It is larger to customer impact like spending, then increase to the biggish weight of user response, preset retrieval weight is corrected with this.
In order to guarantee that the displaying sequence of picture to be detected more meets user demand, guarantee that updated weight can react use
The true idea at family is chosen to be checked for the first time after the similar pictures of the different target picture as user by lookup can also be arranged
When mapping piece is not that the number of the highest picture of similarity is greater than predetermined times, preset retrieval weight is just corrected.
The present invention provides a kind of search methods of similar pictures, first according to edge detection method, extract Target Photo
Global feature information calculates the overall similarity of picture and Target Photo to be measured, further according to Mask RCNN model, described in extraction
The minutia information of Target Photo calculates the details similarity of picture and Target Photo to be measured further according to minutia information,
Further according to overall similarity, details similarity and preset retrieval weight, the retrieval similarity of picture and Target Photo to be measured is calculated,
The finally sequence according to the numerical value of retrieval similarity from big to small, shows picture to be measured corresponding with retrieval similarity.If with
The corresponding picture to be measured of the retrieval similarity maximum value is not chosen at family for the first time, then according to presetting rule, adjusts the preset inspection
Suo Quanchong, or the picture to be detected chosen for the first time after the similar pictures of different target picture as user by lookup are arranged is not
When the number of the highest picture of similarity is greater than predetermined times, then preset retrieval weight can be changed.Compared with prior art, this hair
Different preset retrieval weights is arranged by using for overall similarity and details similarity in bright embodiment, to calculate picture to be measured
With the retrieval similarity of Target Photo.By increasing influence of the details similarity to retrieval similarity, homologue can be distinguished
The details conflict of body, improves the discrimination of similar pictures, to improve recognition effect, reduces search result and user search demand
Deviation.
Further, as the realization to method shown in above-mentioned Fig. 1, the embodiment of the invention provides a kind of similar pictures
Device is retrieved, as shown in figure 3, the device includes:
Extraction module 31, for extracting the global feature information of Target Photo according to edge detection method;
First computing module 32, for calculating picture to be measured and the Target Photo according to the global feature information
Overall similarity;
The extraction module 31, for extracting the minutia information of the Target Photo according to Mask RCNN model;
First computing module 32, is also used to according to the minutia information, calculate the picture to be measured with it is described
The details similarity of Target Photo;
Second computing module 33, for according to the overall similarity, the details similarity and preset retrieval weight, meter
Calculate the retrieval similarity of the picture to be measured Yu the Target Photo;
Display module 34 is shown and the retrieval for the sequence of the numerical value according to the retrieval similarity from big to small
The corresponding picture to be measured of similarity.
The present invention provides a kind of retrieval devices of similar pictures to extract Target Photo first according to edge detection method
Global feature information calculates the overall similarity of picture and Target Photo to be measured, further according to Mask RCNN model, described in extraction
The minutia information of Target Photo calculates the details similarity of picture and Target Photo to be measured further according to minutia information,
Further according to overall similarity, details similarity and preset retrieval weight, the retrieval similarity of picture and Target Photo to be measured is calculated,
The finally sequence according to the numerical value of retrieval similarity from big to small, shows picture to be measured corresponding with retrieval similarity.With it is existing
Technology is compared, and different preset retrieval weights is arranged by using for overall similarity and details similarity in the embodiment of the present invention,
To calculate the retrieval similarity of picture and Target Photo to be measured.By increasing influence of the details similarity to retrieval similarity, energy
The details conflict for enough distinguishing similar object, improves the discrimination of similar pictures, to improve recognition effect, reduce search result with
The deviation of user search demand.
Further, as the realization to method shown in above-mentioned Fig. 2, the embodiment of the invention provides another similar pictures
Retrieval device, as shown in figure 4, the device includes:
Extraction module 41, for extracting the global feature information of Target Photo according to edge detection method;
First computing module 42, for calculating picture to be measured and the Target Photo according to the global feature information
Overall similarity;
The extraction module 41, for extracting the minutia information of the Target Photo according to Mask RCNN model;
First computing module 42, is also used to according to the minutia information, calculate the picture to be measured with it is described
The details similarity of Target Photo;
Second computing module 43, for according to the overall similarity, the details similarity and preset retrieval weight, meter
Calculate the retrieval similarity of the picture to be measured Yu the Target Photo;
Display module 44 is shown and the retrieval for the sequence of the numerical value according to the retrieval similarity from big to small
The corresponding picture to be measured of similarity.
Further, the global feature information includes target object color and object edge profile;
The extraction module 41, comprising:
First extraction unit 411 identifies for using edge detection method and extracts the target object in the Target Photo
The object edge profile;
Second extraction unit 412, for extracting the target object of the target object in the object edge profile
Color.
Further, second extraction unit 412, comprising:
Subelement 4121 is divided, for according to the first preset granularity of division, the Target Photo to be divided into multiple grids
Picture;
Subelement 4122 is inquired, for inquiring each net in the object edge profile described in the Target Photo
The mesh color of trrellis diagram piece, the mesh color include single color and secondary colour;
Computation subunit 4123, for calculating the picture number and the grid picture that the mesh color is secondary colour
The secondary colour ratio that total quantity is compared;
The division subelement 4121, if being also used to the secondary colour ratio greater than the first preset proportion, according to the
Two preset granularity of division repartition the Target Photo;
Subelement 4124 is recorded, if being not more than first preset proportion for the secondary colour ratio, records institute
State the single color picture number of each color for the grid picture that mesh color is single color;
Subelement 4125 is determined, if for the single of the single color picture number and the total quantity of the grid picture
Color ratio example is greater than the second preset proportion, it is determined that the corresponding mesh color of the single color picture number is target object color;
The determining subelement 4125, if the difference being also used between the single color picture number of each color is small
In third preset quantity, it is determined that the mesh color of the grid picture is target object color.
Further, the extraction module 41, comprising:
Screening unit 413 screens the target for matching the Target Photo with the object edge profile
Target material object picture in picture;
Extraction unit 414 extracts the target for the target material object picture to be inputted the Mask RCNN model
The two-value mask image of picture in kind;
Marking unit 415, for marking the image outline in the two-value mask image;
The screening unit 413, be also used to by the two-value mask image after the target material object picture and the label into
Row matching, screens the minutia picture in target material object picture corresponding with described image profile;
Determination unit 416, for by the grid search-engine picture character or graph outline be determined as minutia
Information.
Further, first computing module 42, comprising:
Acquiring unit 421, for obtaining the global feature information of the picture to be measured, the global feature of the picture to be measured
Information includes object under test color and edge contour to be measured;
Computing unit 422, for according to the global feature information of the picture to be measured, the target object color and described
Object edge profile calculates the object color similarity and edge contour similarity of the Target Photo Yu the picture to be measured;
The computing unit 422 is also used to calculate the Target Photo and the picture to be measured according to preset whole specific gravity
The overall similarity.
Further, first computing module 42, comprising:
The acquiring unit 421 is also used to obtain the minutia information of the picture to be measured;
The computing unit 422 is also used to according to the minutia information of the picture to be measured and the Target Photo
Minutia information calculates the details similarity of the Target Photo Yu the picture to be measured.
Further, the method also includes:
Correction module 45, for the numerical value sequence from big to small according to the retrieval similarity, show with it is described
After retrieving the corresponding picture to be measured of similarity, if user's operation meets prerequisite, according to presetting rule, described in amendment
Preset retrieval weight, the prerequisite are that user does not choose the corresponding picture to be measured of the retrieval similarity maximum value for the first time,
Or the picture to be detected chosen for the first time after the similar pictures of the different target picture as user by lookup are arranged is not similarity
The number of highest picture is greater than predetermined times.
The present invention provides a kind of retrieval devices of similar pictures to extract Target Photo first according to edge detection method
Global feature information calculates the overall similarity of picture and Target Photo to be measured, further according to Mask RCNN model, described in extraction
The minutia information of Target Photo calculates the details similarity of picture and Target Photo to be measured further according to minutia information,
Further according to overall similarity, details similarity and preset retrieval weight, the retrieval similarity of picture and Target Photo to be measured is calculated,
The finally sequence according to the numerical value of retrieval similarity from big to small, shows picture to be measured corresponding with retrieval similarity.If with
The corresponding picture to be measured of the retrieval similarity maximum value is not chosen at family for the first time, then can adjust described preset according to presetting rule
Weight is retrieved, or the picture to be detected chosen for the first time after the similar pictures of the different target picture as user by lookup are arranged is not
When being that the number of the highest picture of similarity is greater than predetermined times, then preset retrieval weight can be changed.Compared with prior art, originally
Different preset retrieval weights is arranged by using for overall similarity and details similarity in inventive embodiments, to calculate to mapping
The retrieval similarity of piece and Target Photo.By increasing influence of the details similarity to retrieval similarity, can distinguish similar
The details conflict of object, improves the discrimination of similar pictures, to improve recognition effect, reduces search result and user search demand
Deviation.
A kind of storage medium is provided according to an embodiment of the present invention, and it is executable that the storage medium is stored at least one
The search method of the similar pictures in above-mentioned any means embodiment can be performed in instruction, the computer executable instructions.
Fig. 5 shows a kind of structural schematic diagram of the computer equipment provided according to an embodiment of the present invention, the present invention
Specific embodiment does not limit the specific implementation of computer equipment.
As shown in figure 5, the computer equipment may include: processor (processor) 502, communication interface
(Communications Interface) 504, memory (memory) 506 and communication bus 508.
Wherein: processor 502, communication interface 504 and memory 506 complete mutual lead to by communication bus 508
Letter.
Communication interface 504, for being communicated with the network element of other equipment such as client or other servers etc..
Processor 502 can specifically execute in the search method embodiment of above-mentioned similar pictures for executing program 510
Correlation step.
Specifically, program 510 may include program code, which includes computer operation instruction.
Processor 502 may be central processor CPU or specific integrated circuit ASIC (Application
Specific Integrated Circuit), or be arranged to implement the integrated electricity of one or more of the embodiment of the present invention
Road.The one or more processors that computer equipment includes can be same type of processor, such as one or more CPU;
It can be different types of processor, such as one or more CPU and one or more ASIC.
Memory 506, for storing program 510.Memory 506 may include high speed RAM memory, it is also possible to further include
Nonvolatile memory (non-volatile memory), for example, at least a magnetic disk storage.
Program 510 specifically can be used for so that processor 502 executes following operation:
According to edge detection method, the global feature information of Target Photo is extracted;
According to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated;
According to Mask RCNN model, the minutia information of the Target Photo is extracted;
According to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated;
According to the overall similarity, the details similarity and preset retrieval weight, the picture to be measured and institute are calculated
State the retrieval similarity of Target Photo;
According to the sequence of the numerical value for retrieving similarity from big to small, show corresponding to be measured with the retrieval similarity
Picture.
Obviously, those skilled in the art should be understood that each module of the above invention or each step can be with general
Computing device realize that they can be concentrated on a single computing device, or be distributed in multiple computing devices and formed
Network on, optionally, they can be realized with the program code that computing device can perform, it is thus possible to which they are stored
It is performed by computing device in the storage device, and in some cases, it can be to be different from shown in sequence execution herein
Out or description the step of, perhaps they are fabricated to each integrated circuit modules or by them multiple modules or
Step is fabricated to single integrated circuit module to realize.In this way, the present invention is not limited to any specific hardware and softwares to combine.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all include within protection scope of the present invention.
Claims (10)
1. a kind of search method of similar pictures characterized by comprising
According to edge detection method, the global feature information of Target Photo is extracted;
According to the global feature information, the overall similarity of picture to be measured Yu the Target Photo is calculated;
According to MaskRCNN model, the minutia information of the Target Photo is extracted;
According to the minutia information, the details similarity of the picture to be measured and the Target Photo is calculated;
According to the overall similarity, the details similarity and preset retrieval weight, the picture to be measured and the mesh are calculated
It marks on a map the retrieval similarity of piece;
According to the sequence of the numerical value for retrieving similarity from big to small, show corresponding to mapping with the retrieval similarity
Piece.
2. the method as described in claim 1, which is characterized in that the global feature information includes target object color and target
Edge contour;
It is described according to edge detection method, extract the global feature information of Target Photo, comprising:
Using the edge detection method, the object edge profile of the target object in the Target Photo is identified and extracted;
Extract the target object color of the target object in the object edge profile.
3. method according to claim 2, which is characterized in that described to extract the target object in the object edge profile
Target object color, comprising:
According to the first preset granularity of division, the Target Photo is divided into multiple grid pictures;
In the object edge profile described in the Target Photo, the mesh color of each grid picture, the net are inquired
Lattice color includes single color and secondary colour;
Calculate the secondary colour ratio that the total quantity of picture number and the grid picture that the mesh color is secondary colour is compared;
If the secondary colour ratio is greater than the first preset proportion, the target is repartitioned according to the second preset granularity of division
Picture;
If the secondary colour ratio is not more than first preset proportion, the grid that the mesh color is single color is recorded
The single color picture number of each color of picture;
If the single color ratio example of the total quantity of the single color picture number and the grid picture is greater than the second preset proportion,
Then determine that the corresponding mesh color of the single color picture number is target object color;
If the difference between the single color picture number of each color is less than third preset quantity, it is determined that the grid
The mesh color of picture is target object color.
4. method according to claim 2, which is characterized in that it is described according to MaskRCNN model, extract the Target Photo
Minutia information, comprising:
The Target Photo is matched with the object edge profile, screens the target pictorial diagram in the Target Photo
Piece;
The target material object picture is inputted into the MaskRCNN model, extracts the two-value exposure mask figure of the target material object picture
Picture;
Mark the image outline in the two-value mask image;
The target material object picture is matched with the two-value mask image after the label, screening and described image profile pair
The minutia picture in target material object picture answered;
By in the grid search-engine picture character or graph outline be determined as minutia information.
5. the method as described in claim 1, which is characterized in that it is described according to the global feature information, calculate picture to be measured
With the overall similarity of the Target Photo, comprising:
The global feature information of the picture to be measured is obtained, the global feature information of the picture to be measured includes object under test color
With edge contour to be measured;
According to the global feature information of the picture to be measured, the target object color and the object edge profile, institute is calculated
State the object color similarity and edge contour similarity of Target Photo Yu the picture to be measured;
According to preset whole specific gravity, the overall similarity of the Target Photo Yu the picture to be measured is calculated.
6. the method as described in claim 1, which is characterized in that it is described according to the minutia information, it calculates described to be measured
The details similarity of picture and the Target Photo, comprising:
Obtain the minutia information of the picture to be measured;
According to the minutia information of the minutia information of the picture to be measured and the Target Photo, the target figure is calculated
The details similarity of piece and the picture to be measured.
7. the method as described in claim 1, which is characterized in that the numerical value according to the retrieval similarity is from big to small
Sequentially, after showing picture to be measured corresponding with the retrieval similarity, the method also includes:
If user's operation meets prerequisite, according to presetting rule, the preset retrieval weight, the prerequisite are corrected
The corresponding picture to be measured of the retrieval similarity maximum value, or difference of the setting as user by lookup are not chosen for the first time for user
The picture to be detected chosen for the first time after the similar pictures of Target Photo is not that the number of the highest picture of similarity is greater than preset time
Number.
8. a kind of retrieval device of similar pictures characterized by comprising
Extraction module, for extracting the global feature information of Target Photo according to edge detection method;
First computing module, for calculating whole phase of the picture to be measured with the Target Photo according to the global feature information
Like degree;
The extraction module, for extracting the minutia information of the Target Photo according to MaskRCNN model;
First computing module is also used to calculate the picture to be measured and the target figure according to the minutia information
The details similarity of piece;
Second computing module, for according to the overall similarity, the details similarity and preset retrieval weight, described in calculating
The retrieval similarity of picture to be measured and the Target Photo;
Display module is shown and the retrieval similarity for the sequence of the numerical value according to the retrieval similarity from big to small
Corresponding picture to be measured.
9. a kind of storage medium, it is stored with an at least executable instruction in the storage medium, the executable instruction makes to handle
Device executes the corresponding operation of search method such as similar pictures of any of claims 1-7.
10. a kind of computer equipment, comprising: processor, memory, communication interface and communication bus, the processor described are deposited
Reservoir and the communication interface complete mutual communication by the communication bus;
The memory executes the processor as right is wanted for storing an at least executable instruction, the executable instruction
Ask the corresponding operation of the search method of similar pictures described in any one of 1-7.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534899.8A CN110413824A (en) | 2019-06-20 | 2019-06-20 | A kind of search method and device of similar pictures |
PCT/CN2019/118369 WO2020253063A1 (en) | 2019-06-20 | 2019-11-14 | Method and device for searching for similar images |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534899.8A CN110413824A (en) | 2019-06-20 | 2019-06-20 | A kind of search method and device of similar pictures |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110413824A true CN110413824A (en) | 2019-11-05 |
Family
ID=68359439
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910534899.8A Pending CN110413824A (en) | 2019-06-20 | 2019-06-20 | A kind of search method and device of similar pictures |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110413824A (en) |
WO (1) | WO2020253063A1 (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111079793A (en) * | 2019-11-21 | 2020-04-28 | 北京云聚智慧科技有限公司 | Icon similarity determining method and electronic equipment |
CN111581415A (en) * | 2020-03-18 | 2020-08-25 | 时时同云科技(成都)有限责任公司 | Method for determining similar objects, and method and equipment for determining object similarity |
CN111738362A (en) * | 2020-08-03 | 2020-10-02 | 成都睿沿科技有限公司 | Object recognition method and device, storage medium and electronic equipment |
WO2020253063A1 (en) * | 2019-06-20 | 2020-12-24 | 平安科技(深圳)有限公司 | Method and device for searching for similar images |
CN112464002A (en) * | 2020-12-02 | 2021-03-09 | 北京粉笔蓝天科技有限公司 | Method, apparatus, storage medium, and device for graph reasoning topic image retrieval |
CN112633397A (en) * | 2020-12-29 | 2021-04-09 | 江苏惟妙纺织科技有限公司 | Embroidery customization method and system |
CN112651953A (en) * | 2020-12-31 | 2021-04-13 | 平安国际智慧城市科技股份有限公司 | Image similarity calculation method and device, computer equipment and storage medium |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113743420B (en) * | 2021-08-26 | 2023-12-05 | 北京邮电大学 | Web AR image recognition method and system based on cloud edge end cooperation |
CN113626633A (en) * | 2021-09-01 | 2021-11-09 | 北京云蝶智学科技有限公司 | Picture retrieval method and device |
CN115795288B (en) * | 2022-11-16 | 2023-09-08 | 钰深(北京)科技有限公司 | Garment layout duplication checking method and device, electronic equipment and storage medium |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101639858A (en) * | 2009-08-21 | 2010-02-03 | 深圳创维数字技术股份有限公司 | Image search method based on target area matching |
CN101853299A (en) * | 2010-05-31 | 2010-10-06 | 杭州淘淘搜科技有限公司 | Image searching result ordering method based on perceptual cognition |
KR20130059058A (en) * | 2011-11-28 | 2013-06-05 | 건국대학교 산학협력단 | Preprocessing apparatus for query image and searching image in content based image retrieval using sketch query and methof therefor |
CN103593458A (en) * | 2013-11-21 | 2014-02-19 | 电子科技大学 | Mass image searching system based on color features and inverted indexes |
CN103714181A (en) * | 2014-01-08 | 2014-04-09 | 天津大学 | Stratification specific figure search method |
CN104462502A (en) * | 2014-12-19 | 2015-03-25 | 中国科学院深圳先进技术研究院 | Image retrieval method based on feature fusion |
CN106021575A (en) * | 2016-05-31 | 2016-10-12 | 北京奇艺世纪科技有限公司 | Retrieval method and device for same commodities in video |
CN108898610A (en) * | 2018-07-20 | 2018-11-27 | 电子科技大学 | A kind of object contour extraction method based on mask-RCNN |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2007206920A (en) * | 2006-02-01 | 2007-08-16 | Sony Corp | Image processor and image processing method, retrieving device and method, program and recording medium |
CN107403437A (en) * | 2016-05-19 | 2017-11-28 | 上海慧流云计算科技有限公司 | The method, apparatus and robot of robotic tracking's object |
CN109284669A (en) * | 2018-08-01 | 2019-01-29 | 辽宁工业大学 | Pedestrian detection method based on Mask RCNN |
CN110413824A (en) * | 2019-06-20 | 2019-11-05 | 平安科技(深圳)有限公司 | A kind of search method and device of similar pictures |
-
2019
- 2019-06-20 CN CN201910534899.8A patent/CN110413824A/en active Pending
- 2019-11-14 WO PCT/CN2019/118369 patent/WO2020253063A1/en active Application Filing
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101639858A (en) * | 2009-08-21 | 2010-02-03 | 深圳创维数字技术股份有限公司 | Image search method based on target area matching |
CN101853299A (en) * | 2010-05-31 | 2010-10-06 | 杭州淘淘搜科技有限公司 | Image searching result ordering method based on perceptual cognition |
KR20130059058A (en) * | 2011-11-28 | 2013-06-05 | 건국대학교 산학협력단 | Preprocessing apparatus for query image and searching image in content based image retrieval using sketch query and methof therefor |
CN103593458A (en) * | 2013-11-21 | 2014-02-19 | 电子科技大学 | Mass image searching system based on color features and inverted indexes |
CN103714181A (en) * | 2014-01-08 | 2014-04-09 | 天津大学 | Stratification specific figure search method |
CN104462502A (en) * | 2014-12-19 | 2015-03-25 | 中国科学院深圳先进技术研究院 | Image retrieval method based on feature fusion |
CN106021575A (en) * | 2016-05-31 | 2016-10-12 | 北京奇艺世纪科技有限公司 | Retrieval method and device for same commodities in video |
CN108898610A (en) * | 2018-07-20 | 2018-11-27 | 电子科技大学 | A kind of object contour extraction method based on mask-RCNN |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020253063A1 (en) * | 2019-06-20 | 2020-12-24 | 平安科技(深圳)有限公司 | Method and device for searching for similar images |
CN111079793A (en) * | 2019-11-21 | 2020-04-28 | 北京云聚智慧科技有限公司 | Icon similarity determining method and electronic equipment |
CN111581415A (en) * | 2020-03-18 | 2020-08-25 | 时时同云科技(成都)有限责任公司 | Method for determining similar objects, and method and equipment for determining object similarity |
CN111738362A (en) * | 2020-08-03 | 2020-10-02 | 成都睿沿科技有限公司 | Object recognition method and device, storage medium and electronic equipment |
CN112464002A (en) * | 2020-12-02 | 2021-03-09 | 北京粉笔蓝天科技有限公司 | Method, apparatus, storage medium, and device for graph reasoning topic image retrieval |
CN112633397A (en) * | 2020-12-29 | 2021-04-09 | 江苏惟妙纺织科技有限公司 | Embroidery customization method and system |
CN112651953A (en) * | 2020-12-31 | 2021-04-13 | 平安国际智慧城市科技股份有限公司 | Image similarity calculation method and device, computer equipment and storage medium |
CN112651953B (en) * | 2020-12-31 | 2024-03-15 | 平安国际智慧城市科技股份有限公司 | Picture similarity calculation method and device, computer equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
WO2020253063A1 (en) | 2020-12-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110413824A (en) | A kind of search method and device of similar pictures | |
Scharr et al. | Leaf segmentation in plant phenotyping: a collation study | |
Kim et al. | Salient region detection via high-dimensional color transform | |
CN109815801A (en) | Face identification method and device based on deep learning | |
WO2019207557A1 (en) | Method for distinguishing a real three-dimensional object from a two-dimensional spoof of the real object | |
Liu et al. | Multimorphological superpixel model for hyperspectral image classification | |
CN108140107B (en) | Quickly, high-precision large-scale fingerprint verification system | |
CN108629373A (en) | A kind of image classification method, system, equipment and computer readable storage medium | |
Chawathe | Rice disease detection by image analysis | |
Raghukumar et al. | Comparison of machine learning algorithms for detection of medicinal plants | |
CN109726725A (en) | The oil painting writer identification method of heterogeneite Multiple Kernel Learning between a kind of class based on large-spacing | |
Teymournezhad et al. | Detection of counterfeit banknotes by security components based on image processing and googlenet deep learning network | |
Bressane et al. | Statistical analysis of texture in trunk images for biometric identification of tree species | |
CN112651364A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
Gong et al. | Class-agnostic object counting robust to intraclass diversity | |
Shashank et al. | Identifying epiphytes in drones photos with a conditional generative adversarial network (C-GAN) | |
CN110458200A (en) | A kind of flower category identification method based on machine learning | |
CN109271544A (en) | Automatically the method and device of artist's masterpiece is selected | |
CN112070079B (en) | X-ray contraband package detection method and device based on feature map weighting | |
CN112381730A (en) | Remote sensing image data amplification method | |
Lino et al. | Performance of haar and LBP features in cascade classifiers to whiteflies detection and counting | |
Hannan et al. | Analysis of Detection and Recognition of Human Face Using Support Vector Machine | |
Iwata et al. | Tree recognition based on leaf images | |
Weng et al. | Online image classification using IHDR | |
Devesh et al. | Retrieval of monuments images through ACO optimization approach |
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 |