CN107798093A - Image search method - Google Patents

Image search method Download PDF

Info

Publication number
CN107798093A
CN107798093A CN201711006969.XA CN201711006969A CN107798093A CN 107798093 A CN107798093 A CN 107798093A CN 201711006969 A CN201711006969 A CN 201711006969A CN 107798093 A CN107798093 A CN 107798093A
Authority
CN
China
Prior art keywords
node
content
view data
request
visitor
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201711006969.XA
Other languages
Chinese (zh)
Other versions
CN107798093B (en
Inventor
方引
杨洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chengdu Technology Co Ltd
Original Assignee
Chengdu Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chengdu Technology Co Ltd filed Critical Chengdu Technology Co Ltd
Priority to CN201711006969.XA priority Critical patent/CN107798093B/en
Publication of CN107798093A publication Critical patent/CN107798093A/en
Application granted granted Critical
Publication of CN107798093B publication Critical patent/CN107798093B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F16/51—Indexing; Data structures therefor; Storage structures
    • 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
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/48—Extraction of image or video features by mapping characteristic values of the pattern into a parameter space, e.g. Hough transformation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467—Encoded features or binary features, e.g. local binary patterns [LBP]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/513—Sparse representations

Abstract

The invention provides a kind of image search method, this method includes：The association established between facial image actual blocks of data and storage location, and the association established between retrieval information and actual storage data block；The solicited message sent to the data owner of facial image by client, is retrieved in the index, and the retrieval result of correlation then is fed back into client.The present invention proposes a kind of image search method, helps face blocks, sample size and quality are relatively low, improves the accuracy rate of recognition of face in the case of loss of learning, while to reduce the run time of identification.

Description

Image search method
Technical field
The present invention relates to cloud computing, more particularly to a kind of image search method.
Background technology
With the continuous development of society and the continuous progress of science and technology, oneself warp of the research of face information processing turns into current and ground Study carefully one of focus.The research contents of recognition of face is related to pattern-recognition, Computer Image Processing, machine learning and artificial intelligence Deng field, it is also obtained in commercial fields such as living things feature recognition, man-machine interaction, content retrieval, video monitoring, safety-protection systems It is widely applied.Although numerous kinds of face recognition algorithms are own at present obtains preferable recognition performance, face identification system is in reality Many challenges are still faced in the application of border, including：Recognition of face problem is blocked as caused by illumination variation, ornaments etc.；It is non-can The sample number that can be collected under the conditions of control is few, posture conversion etc., can all cause the missing of face information.How in existing method On the basis of overcome problem above further to improve the accuracy rate of recognition of face, while reduce identification run time improve actual effect Property, it is problem urgently to be resolved hurrily at present.
The content of the invention
To solve the problems of above-mentioned prior art, the present invention proposes a kind of image search method, including：
The association established between facial image actual blocks of data and storage location, and establish retrieval information and actual storage number According to the association between block；
The solicited message sent to the data owner of facial image by client, is retrieved, then in the index The retrieval result of correlation is fed back into client.
Preferably, data block storage address is divided using address cutting retrieval.
Preferably, the storage format of the facial image is according to the view data visitor/view data owner/content, The address information of data storage block is divided into 3 sections.
Preferably, the data of described image content and index are stored as a unit, the path of each content by The HASH values of picture material title and operating time composition.
Preferably, the flow of client node issue request content includes：
(1) the view data owner proposes content query requests to file system, and first hair is sent to local Agent for request, Whether the request after analysis, can cache according to the request, decide whether to be forwarded to high in the clouds；
(2) content for whether having request in service is locally stored in local Agent service-seekings, if then going to (12)；
(3) if the content for not finding request in service is locally stored, Content query service is forwarded the request to；
(4) Content query service check in view data visitor's node listing whether have it is similar to request content inquiry Path address；
(5) if there is the address similar to request content inquiry in view data visitor's node listing, pass directly to View data owner node corresponding to it；
(6) if the inquiry similar to request content inquiry, content are not looked into view data visitor's node listing Inquiry service sends a query to lower data storage and retrieval system module, inquires about the view data owner node of content search；
(7) search data memory system is by finding corresponding view data owner node；
(8) view data visitor node, view data owner node or picture material querying node its manipulative indexing Table, it is confirmed whether to exist the corresponding index of request content inquiry, if now view data visitor node and do not find please The index of content search is sought, then continues to inquire about next view data visitor node or image by search data memory system Content node；
(9) index that view data visitor node, view data owner node or picture material node will inquire Information returns to the Content query service for the node for initiating inquiry request；Now if the result returned includes requesting query View data visitor's node, then Content query service view data visitor's node listing is updated；
(10) return information is forwarded to be serviced to local Agent；
(11) if what is returned is sky, local Agent services on high in the clouds by obtaining content, otherwise according to the result of return Content is obtained from respective nodes.Then it is locally stored in service and stores the backup of this content, and issues corresponding index information and arrive Corresponding view data visitor node and view data owner node；
(12) file system is sent content to.
The present invention compared with prior art, has advantages below：
The present invention proposes a kind of image search method, contributes to face blocks, sample size and quality are relatively low, information The accuracy rate of recognition of face is improved in the case of missing, while reduces the run time of identification.
Brief description of the drawings
Fig. 1 is the flow chart of image search method according to embodiments of the present invention.
Embodiment
Retouching in detail to one or more embodiment of the invention is hereafter provided together with the accompanying drawing for illustrating the principle of the invention State.The present invention is described with reference to such embodiment, but the invention is not restricted to any embodiment.The scope of the present invention is only by right Claim limits, and the present invention covers many replacements, modification and equivalent.Illustrate in the following description many details with Thorough understanding of the present invention is just provided.These details are provided for exemplary purposes, and without in these details Some or all details can also realize the present invention according to claims.
An aspect of of the present present invention provides a kind of image search method.Fig. 1 is image retrieval according to embodiments of the present invention Method flow diagram.
The mass data storage searching system based on cloud computing of the present invention includes back end and index node.Index section Point safeguards face image data block index, the mapping relations between data block, data block's attribute, and back end is with different images number It is the actual face image data block of unit storage according to the owner.When the view data owner accesses storage system, obtain independent Space.Different ChunkID is distributed for each data block, each image block and copy are stored on each back end.Face The index of video data block is included with properties：When ChunkID, title, type, size, view data owner title, access Between and its positional information.
User's block recording image data owner storage ChunkID in systems, sharing mode, video data block name are right It should be related to, view data owner's title.System is obtained by the OwnerID of the access images data owner ChunkID mapped Take family block and give user's block space of view data owner's independence.View data owner user block only allows correspondingly The OwnerID view data owner just has permission to access its all user's block, and distributes a ChunkID for user's block, Each back end preservation data block of system is stored in after piecemeal is carried out to it.
Supernode is also set up in search data memory system, supernode has high speed bandwidth and high-performance node, each Supernode safeguards a routing table, can adjust its routing table according to the power of self-ability.The institute in search data memory system Some supernodes form a storage ring.M rows are included in supernode n routing table, often row includes x items, and x represents joint behavior Power, row k include the by stages such as x items, x [n+3i, n+3i+1), wherein 0<k<M, each supernode is according to itself current ability Dynamic adjusts its x value.Storage ring is used for routing inquiry request.Each supernode is responsible for safeguarding its forerunner, descendant node.Super section Point n descendant node is the descendant node immediately in storage node on the ring n, i.e., clockwise since n in storage ring First supernode, equally, n predecessor node are forerunner's supernodes immediately of n in storage ring.Each supernode also safeguards one Back end list, each single item points to a supernode in the routing table of supernode, in back end list record from itself to All back end of its descendant node, the data letter that signified node can be used on backup supernode in back end list Breath, or when supernode image data storage is overflowed, the view data of spilling is transferred to signified in its back end list On node.All solicited messages are all route by storage ring in search data memory system.Wherein back end is looked into Ask request and be first forwarded to its follow-up supernode, then route can be carried out in storage ring, eventually arrive at destination.Super section The inquiry request of point is directly route in storage ring.
The Indexing Mechanism of face image data block includes the foundation and block retrieval of index.Index, which is established, to be used to establish actual number According to the association between block and storage location, the association established between retrieval information and actual storage data block, and store necessary The information of data block.Block retrieval includes examining the solicited message that the view data owner is sent by client in the index Rope, the retrieval result of correlation is then fed back into client.Block retrieval will be related to the storage format of video data block in itself, section The composition structure and block retrieval mode of point.The present invention is divided using address cutting retrieval to data block storage address.
Facial image storage format is according to the view data visitor/view data owner/content, by data storage block Address information is divided into 3 sections.All picture numbers that the data-base recording of view data visitor view data visitor possesses According to the inventory of the owner；All the elements in the database maintenance of view data owner view data owner's server；Figure As storage, deletion and the picture material in search node equipment are responsible in content service.The data and index of picture material are stored in Together, preserved as a unit.The path of each content is made up of the HASH values of picture material title and operating time.
The method of the present invention is that keyword identifiers are divided into 3 parts.By keyword identifiers position from high to low Sequentially, this three parts is view data visitor k respectively1, view data owner k2, picture material k3, query by image content value Keyword ID according to k1k2k3Order connection.Its absolute value：|k1|, | k2|, | k3| calculate in the following order, first by figure As the order of content search feature calculates the HASH values of first content search value and is taken by the order from a high position to low level | k1| Position is used as k1Value, next calculate second content search value HASH values take | k2| position is as k2Value, then content is looked into The remainder of inquiry value calculates HASH and taken | k3| position is as k3Value.Similar inquiry is arranged in close proximity scope.
Divided when video data block stores according to its keyword, its ID is divided into s1/s2/s33 component arrangement from high to low In route storage ring, s1It is view data visitor's address field, s2It is view data owner's address field, s3It is picture material Address field.By data block address cutting, view data visitor or figure where the view data owner determines data block Behind data owner position, the quick particular location for determining video data block.
The node for storing visitorID is denoted as view data visitor's node of content search, storage by the present invention OwnerID node is designated as the view data owner node of content search.The node that storage content is finally inquired about to ID is denoted as The picture material node of this query by image content.The index information of content search can be stored in its all view data and visit The person's of asking node, view data owner node and picture material node.
Each back end safeguards local storage content, safeguards view data visitor's content indexing table, view data Owner's content indexing table and content indexing table, while safeguard view data visitor's node listing, view data owner section Point list, picture material node listing.Include content search address, image in view data visitor's content indexing table per a line Node corresponding to content last access time, multiple content searches where picture material etc.；Equally, in the view data owner Hold concordance list and content indexing table also includes these, but the picture material node of content search therein is in this node；Image Content node list includes picture data content.View data visitor's node listing safeguards the nearest access frequency of this node most High view data owner node；This node of view data owner node list maintenance is accessed frequency highest recently Picture material node；Picture material node storage content particular content.
Each node is operated local storage server by the proxy server of a local；Server is locally stored to lead to Cross search data memory system platform and other nodes sharing storage contents in network.In the search data memory system of bottom Platform carries out Content query service, there is provided retrieves information request and returned data storage and retrieval system retrieval result is to locally Agent.All search data memory systems for being routed through bottom provide.Wherein, content search includes view data visitor Service, view data owner service and content service, and include inquiry, insertion and deletion action.Content query service is appointed Business is to judge to be forwarded to view data owner node or transmission according to the information in view data visitor's node listing To lower floor's routing mechanism.The view data owner node and picture material node that insertion operation returns according to Content query service Information issues the index information of corresponding query image data owner node to corresponding node.
Subsidiary the inquired about address of search data memory system queries operation, if the path image content section of requesting node When there is no the similar address of respective request content search in point list, the Query Information of search data memory system is sent to first Query image data access person's node, after view data visitor's node is reached, if not finding corresponding index entry, The ID of view data owner node is calculated, and is continued since view data visitor's node in search data memory system Inquired about, until finding result or reaching picture material node.If view data visitor's node listing of requesting node Inside there is the similar path address of respective request content search, be then forwarded directly to corresponding view data owner node, and from This view data owner node starts a query at.
According to above-mentioned block retrieval mode, its block retrieval flow can be divided into two steps, first, node issues content information；Two It is that data block retrieves data block and feedback request according to content information.
The workflow of node issue request content includes：
(1) the view data owner proposes content query requests to file system, and first hair is sent to local Agent for request, Whether the request after analysis, can cache according to the request, decide whether to be forwarded to high in the clouds；
(2) content for whether having request in service is locally stored in local Agent service-seekings, if then going to (12)；
(3) if the content for not finding request in service is locally stored, Content query service is forwarded the request to；
(4) Content query service check in view data visitor's node listing whether have it is similar to request content inquiry Path address；
(5) if there is the address similar to request content inquiry in view data visitor's node listing, pass directly to View data owner node corresponding to it；
(6) if the inquiry similar to request content inquiry, content are not looked into view data visitor's node listing Inquiry service sends a query to lower data storage and retrieval system module, inquires about the view data owner node of content search；
(7) search data memory system is by finding corresponding view data owner node；
(8) view data visitor node, view data owner node or picture material querying node its manipulative indexing Table, it is confirmed whether to exist the corresponding index of request content inquiry, if now view data visitor node and do not find please The index of content search is sought, then continues to inquire about next view data visitor node or image by search data memory system Content node；
(9) index that view data visitor node, view data owner node or picture material node will inquire Information returns to the Content query service for the node for initiating inquiry request；Now if the result returned includes requesting query View data visitor's node, then Content query service view data visitor's node listing is updated.
(10) return information is forwarded to be serviced to local Agent；
(11) if what is returned is sky, local Agent services on high in the clouds by obtaining content, otherwise according to the result of return Content is obtained from respective nodes.Then it is locally stored in service and stores the backup of this content, and issues corresponding index information and arrive Corresponding view data visitor node and view data owner node；
(12) file system is sent content to.
And wherein, block retrieval request content and feedback information, including：
The path indexing information of this content is issued after content obtaining in corresponding view data visitor node and image Hold node.When a node n sends a content query requests content, this node first checks for whether content needs to pass through Lower floor's routing mechanism carries out inquiry acquisition, and if necessary to inquire about, node, which will be checked in its view data visitor's node listing, is View data visitor's node of the no content search for having a request.If without in request in view data visitor's node listing View data visitor's node of appearance, now, according to the visitorID of content search address computation content and it is sent to lower floor's number Go to inquire about the visitor picture material nodes of content according to storage and retrieval system, by inquiry, search data memory system will be found View data visitor's node of this content, then view data visitor node check its view data visitor's content rope Draw the index for whether thering is requested content to inquire about in table, view data owner's content indexing table and content indexing table, if rope Draw presence, then the address of node where one requested content of this path image content node return to node n, node n By asking this agent node to obtain content；If not having requested content on the agent node returned, request will hair It is sent to cloud server.Otherwise, if index is not present, calculate ownerID and be sent to visitor picture material nodes, number Continue to inquire about since view data visitor's node according to storage and retrieval system, again by inquiry, view data institute will be found The person's of having node, view data owner node equally check index table information thereon, if corresponding index information, then returned An agent node information is returned to local node, otherwise continues to inquire about next view data owner node, follows storage successively Ring until have on some view data owner node requested content inquire about index information, or find requested content inquiry Picture material node, poll-final.If there is no requested content in the concordance list of the picture material node of content search yet Index, then request be sent to cloud server.In this query process, each view data owner node will return Itself address information is returned to local node n.
If there is the view data owner node of a request content inquiry in view data visitor's node listing, this When inquiry is forwarded directly to this view data owner node, this view data owner node checks its concordance list, If there is index, then an agent node is returned, otherwise continues to inquire about next view data owner node until finding Picture material node.
After a node obtains a content, it is by into the view data owner node and image of issue request Hold node to release news, informing these nodes, it has the request content of this content search, view data owner node and figure As content node accordingly updates their concordance list.In this query process, if node n view data owner node The view data owner node of request content inquiry is stored with list, then is forwarded directly to this view data owner node Start a query at.
In terms of facial image feature extraction, the present invention is poor using regional center pixel and the annular neighborhood territory pixel point of storage Value size characterizes the texture eigenvalue of pixel, by image using in units of neighborhood of pixel points as texture unit, then pass through Two-value numerical value quantifies to the texture unit, obtains Local textural feature value, is gone forward side by side by the texture unit in statistical picture Row normalization operation, obtains describing the texture feature vector of image, and the detailed step that feature extraction is carried out using this method is as follows It is described：
First, binary-coding is carried out to image.A region is randomly selected in the facial image of collection, is appointed in the region Meaning pixel can be described with G (y, z), and its geometric center point can use hcIt is described, to the neighborhood in 3 × 3 windows Pixel h0To h7Binary transform processing is carried out, it is as follows：
hd=t (h0-hc) ... t (h7-hc)；
Wherein
Processing is weighted to above-mentioned binary transform result, obtains the local binary patterns value of the window center：
Q is set to be used to describe K kind characteristic types, Q ∈ (0,1,2 ..., K-1).The facial image of collection is divided into n × p Block, the occurrence number of each pattern in each piecemeal is counted, that is, the characteristic type in every piece of facial image subregion is carried out Statistics, obtain by n × p set of histograms into facial image characteristic component U=(U1,U2,…Un×p).Wherein, Molecule Pj(Q) it is used to describe the quantity that local binary patterns value in j-th of subregion is Q feature,For describing the The binary pattern histogram of j sub-regions.
Method according to being set forth above establishes facial image feature histogram, so as to provide data base for facial image retrieval Plinth.
In order to reduce influence of noise in image processing process, the present invention is carried out based on medium filtering visual characteristics of human eyes Denoising.Noise spot is determined first, if image R sizes are m × n, is slided using the window of 3 × 3 sizes on image.
The window center grey scale pixel value is defined as, then all pixels point value set is in the window：
wI, j=g (i+k, j+r) | k, r=(1,0, -1) }
Pixel average in calculation window
Image R maximum gradation value and minimum gradation value is found out, is designated as I respectivelymax(m×n)、Imin(m×n).In mark The threshold value of imago vegetarian refreshments is Hi,j。
Then when central pixel point gray value meets following condition, the pixel can be determined as noise spot：
If | g (i, j)-wm| ＞ Hi,j, then the pixel is noise spot.
If | g (i, j) |=Imax(m × n) or Imin(m × n), then the pixel is noise spot.
For above-mentioned condition, the present invention determines threshold value H according to noise-sensitive coefficient lambdai,jSize.Define in window Imago vegetarian refreshments g (i, j) noise-sensitive coefficient lambda is
Now, judge whether pixel is noise spot, as long as the noise-sensitive coefficient lambda calculatedi,jIf | g (i, j)-wm| ＞ λi,jMeet condition.
After image pixel point is divided into noise spot and the non-noise class of point two, NURBS functions are used for image g (i, j) It is carried out smoothly, image is considered as the uniform sampling of curved surface, be original image and k batten and l spline function from The result of convolution is dissipated, is described as：
Wherein, BkAnd B (x-i)l(y-i) be respectively NURBS k batten and l Spline convolution template, if g (i, j) is Noise spot, then 3 × 3 filter window is taken, obtain filtered value and carry out medium filtering again, obtain end value.
If i takes [0,255], it is as follows that force function is differentiated in definition：
Fr(i)=N (i)/max [N (i)]
Define target area membership function be：
Wherein, f (i) is monotonically increasing function, and meets condition f (a)=0, f (b)=1.Gray value is at [0, a] section The pixel belongs to background area, and at [b, 255] section, the pixel belongs to target area, pixel when between section [a, b] Need further to represent by ambiguity function.
According to the feature of above method extraction facial image, and human eye visual perception model is established according to features described above, from And realize facial image and retrieve.
The present invention is using feature operator using the peak value of the pixel gradient direction histogram in feature vertex neighborhood as this feature The principal direction of point, and reference axis is rotated to be to the principal direction of characteristic point.Calculate two histogram vector HiAnd H (x)j(x) similar Degree：
Wherein, | | Hi| | and | | Hj| | represent the length of histogram feature vector.
Off-note point pair is detected then in conjunction with dimension, it is abnormal right finally to be abandoned with stochastical sampling uniformity. Whole process is exactly as transformation matrix by sample data set fitted figure.Initial sample data n=min { n0,max{ns,nslog2 μn0}}。n0It is according to the quantity of the matching characteristic point of k nearest neighbor algorithm judgement, nsTo abandon off-note point to matching characteristic before The quantity of point, μ is adjustment parameter.(the x of original image1, y1) and target image (x2, y2) transformation relation is as follows：
It is the transformation matrix of 8 parameters, obtaining the matrix parameter at least needs four character points pair, The preferred embodiment of the invention uses weighted least-squares method solution matrix parameter, if
K=[k1k2k3k4k5k6k7k8]
L=- [x2y2]T/μ
Then it is transformed to:
K=- [GTG]-1GTL
Schilling μ initial values are 1 and obtain K initial value, then proceed to iterate to calculate μ, finally give stable K.It is specific to calculate Method is as follows：
(1) the matching characteristic point pair of Different Plane is randomly selected, calculates the transformation matrix K of these points pair；
(2) for matching double points (x, y) to be detected, if meeting condition | Kx-y |<ε, ε are tolerance, then the point is interior Point.If interior points are more than given threshold t, matrix K is recalculated by iteration weighted least-squares method, and update interior count Amount, if interior quantity is less than t, return to step (1)；
(3) if after W iteration, most imperial palace point set quantity determines and is more than t, is combined according to interior point and calculates change Change matrix K.
In further aspect, The present invention gives the condition that threshold value meets, handled by criterion, prevent noise Point erroneous judgement.False coordinate (x, y) and (x*, y*) represent the feature operator of source images and target image, each characteristic point pair respectively Feature operator can obtain with the following methods.
Δ x=x-sm(x*·cos(Δθm)-y*·sin(Δθm))
Δ y=y-sm(x*·cos(Δθm)-y*·sin(Δθm))
Wherein Δ x and Δ y represents that the histogram of extraction feature represents.Quadrinomial (sm, Δ θm, Δ xm, Δ ym) represent to delete The conversion approximation of off-note pair, specific implementation meet following condition：
|Δx-Δxm|>Δxt；|Δy-Δym|>Δyt
According to the width of histogram, Δ xtWith Δ ytThe horizontal and vertical poor threshold value of histogram is represented respectively.
After carrying out specific Point matching by the conversion of above-mentioned image, decision tree is applied to facial modeling by the present invention In.First, facial modeling is trained using shape indexing pixel grey scale feature.The part established with two reference points Two coordinate points of stochastical sampling in coordinate system, make the Pixel gray difference between the two points, then choose in two reference points A midpoint additional random offset generate characteristic point, feature is used as using the Pixel gray difference of the two characteristic points.
When training decision tree, the input of tree is facial image I, corresponds to the shape S and reference point of reference point coordinate composition True shape S ', and export target for prediction reference point offset Δ S.Training for decision tree determines first The division of non-leaf nodes in tree.With I, (p, Δ x, Δ y) is represented after current reference point shape is carried out into similarity transformation, is obtained Picture in using (Δ x, the gray value of Δ y) pixels in the local coordinate system that p-th of reference point is established as origin.Set and work as The segmentation threshold of front nodal point, its span is [- 255,255].In the local coordinate system established using reference point p as origin, take The difference of the shape indexing pixel grey scale value tag of two points and a threshold valueIt is compared, if being less than the threshold value, then will training Sample is divided into left child node, is otherwise assigned in right child node.
Optimum feature functions f0And optimal thresholdSelection can be described with equation below：
Wherein Δ SLForWhen Δ S portion；ΔSRForWhen Δ S portion；F (I)=I (p, Δ x1, Δy1)-I (p, Δ x2, Δ y2)；
Var(ΔSL) represent in left child node corresponding p-th of reference point offset variance yields, Var (Δ SR) represent It is the variance yields of the offset of corresponding p-th of reference point in right child node.
For each non-leaf nodes, a characteristic function f is selected to come to all sample extraction shapes corresponding to the node Index feature, then need to select a threshold valueThese shape indexing features are split, by the training sample of present node (I, S, S ') it is divided into left and right child node two parts (IL,SL,S’L) and (IR,SR,S’R)。
Each internal node of decision tree is to be trained in the above described manner, is so trained for each key point Decision tree has just been combined into a decision forest.The information output for the sample that leaf node in decision tree is included is expressed as one Individual binary feature vector, one-dimensional characteristic vector is combined into by being connected before and after the binary feature of all decision trees in decision forest. One facial image corresponding Feature Mapping in t layer decision forests is represented with following equation：
δt={ δt iI=1 ... L
Wherein t represent decision tree residing for the number of plies, L represent face shape in reference point quantity.δt iRepresent be by The characteristic vector that the binary feature that all decision trees corresponding to i-th of reference point extract is in series, referred to as local binary Feature.By extracting each δ corresponding to reference point in facet iAfter feature, by all δt iIt is connected into final binary feature Vector represents the Feature Mapping relation δ of facet。
Image is randomly selected from face images as training sample set, remaining image is as test sample collection； Carry out SIFT feature extraction and DCT feature extractions respectively to all training images, it is special to include SIFT phases for wherein SIFT feature Seek peace SIFT amplitude Characteristics.
By nonlinear function Φ by facial image DUAL PROBLEMS OF VECTOR MAPPING into high-dimensional feature space F, it is then empty in high dimensional feature Between principal component analysis conversion is carried out in F.When carrying out principal component analysis conversion, introducing meets that the nonlinear function E of core condition carrys out generation For the inner product operation of vector, i.e. E (xi,xj)=Φ (xi)·Φ(xj).The process of principal component analysis is：
The training sample face vector x that m is tieed up1,x2,...,xtHigh-dimensional feature space is mapped to using nonlinear function Φ F, obtain Φ (x1),Φ(x2),...,Φ(xt)；
To Φ (x in Fi) enter line translation.Solve characteristic equation l λΦα=K α, wherein K=(E (xi,xj))l×i, so as to It is to characteristic vector：
Corresponding characteristic value is λΦ 1, λΦ 2... λΦ l；Preceding m characteristic value and the corresponding characteristic vector in characteristic value are taken, Obtain eigenmatrix MΦ=(DΦ)1/2(VΦ)T, wherein：
DΦ=diag (λΦ 1, λΦ 2... λΦ m)
VΦ=(v1,v2,...,vm)
So training sample is after conversion in the F of space:
Obtain corresponding separation matrix WΦ；
To any one test sample y, it is Φ (y) to map that to space F, extracts its characteristic vector and is
After completing principal component analysis process, core independent characteristic vector sum proper subspace is obtained；Core independent characteristic is carried out Fusion Features, one-dimensional characteristic vector is obtained, finally give all characteristic vectors of training sample set；Using obtained characteristic vector It is trained SVM models；
After the characteristic vector that all test sample collections are obtained using same procedure, the characteristic vector of test sample collection is distinguished Projected to its subspace, obtain the core independent characteristic vector of test sample collection；
Class test will be carried out in the vectorial SVM models for being used to train of core independent characteristic, obtain the preliminary of facial image Recognition result.
Training sample is preferably further divided into overlapping block by the present invention, calculates the differentiation rate of each block respectively, Then the block construction template that differentiation rate is higher is selected, and training sample is filtered, is constructed by the training sample after filtering new Dictionary, finally classified with rarefaction representation.
Set A=[the A of the given n sample comprising C class1*,A2*,…An*]。Ai* i-th of image array is represented.Often Individual training image is divided into k overlapping block, after the block matrix of each image is converted into vector, i.e. Ai*=[ai,1,ai,2,… ai,k].Whole training dictionary set A is expressed as A=[A1,A2,…An], wherein AiRepresent l-th of module of all images to Measure the matrix formed.
To each module collection Ai, useMean vector corresponding to expression,Represent i-th of mould of all images in c classes Block vector ac,iAverage, c ∈ [1, C].Then modules AiDifferentiation rate it is as follows：
The size of module differentiation rate is ranked up from high to low, and h has module structure into template T before only retaining. Test and training sample image are filtered with the template.Training set fA=[fa after filtering1,fa2,...,fah], wherein faiIt is Filtered image Ai* vector representation, h are the template numbers included in template.
Further to reduce amount of calculation, using principal component analysis extraction principal component on fA, and projection matrix P is constructed, then Training image and the dimension of test sample y can yojan be further：
fpA=P'fA
fpY=P'fy
fpY can be expressed as fpA linear combination
fpY=fpA·X
X is sparse matrix, test sample is classified as by minimal reconstruction residual error according to class residual error corresponding to class：
WhereinTo select function.||()||2For l2Norm constraint.
In summary, the present invention proposes a kind of image search method, helps to block in face, sample size and quality The accuracy rate of recognition of face is improved in the case of relatively low, loss of learning, while reduces the run time of identification.
Obviously, can be with general it should be appreciated by those skilled in the art, above-mentioned each module of the invention or each step Computing system realize that they can be concentrated in single computing system, or be distributed in multiple computing systems and formed Network on, alternatively, they can be realized with the program code that computing system can perform, it is thus possible to they are stored Performed within the storage system by computing system.So, the present invention is not restricted to any specific hardware and software combination.
It should be appreciated that the above-mentioned embodiment of the present invention is used only for exemplary illustration or explains the present invention's Principle, without being construed as limiting the invention.Therefore, that is done without departing from the spirit and scope of the present invention is any Modification, equivalent substitution, improvement etc., should be included in the scope of the protection.In addition, appended claims purport of the present invention Covering the whole changes fallen into scope and border or this scope and the equivalents on border and repairing Change example.

Claims (5)

  1. A kind of 1. image search method, it is characterised in that including：
    The association established between facial image actual blocks of data and storage location, and establish retrieval information and actual storage data block Between association；
    The solicited message sent to the data owner of facial image by client, is retrieved in the index, then by phase The retrieval result of pass feeds back to client.
  2. 2. according to the method for claim 1, it is characterised in that data block storage address is carried out using address cutting retrieval Division.
  3. 3. according to the method for claim 1, it is characterised in that the storage format of the facial image is visited according to view data The person of the asking/view data owner/content, 3 sections are divided into by the address information of data storage block.
  4. 4. according to the method for claim 3, it is characterised in that the data and index of described image content are used as a unit Stored, the path of each content is made up of the HASH values of picture material title and operating time.
  5. 5. according to the method for claim 3, it is characterised in that the flow of client node issue request content includes：
    (1) the view data owner proposes content query requests to file system, and first hair is sent to local Agent for request, should ask Ask after analysis, whether can be cached according to the request, decide whether to be forwarded to high in the clouds；
    (2) content for whether having request in service is locally stored in local Agent service-seekings, if then going to (12)；
    (3) if the content for not finding request in service is locally stored, Content query service is forwarded the request to；
    (4) whether Content query service is checked in view data visitor's node listing the path similar to request content inquiry Address；
    (5) if there is the address similar to request content inquiry in view data visitor's node listing, its institute is passed directly to Corresponding view data owner node；
    (6) if the not inquiry similar to request content inquiry in view data visitor's node listing, content search clothes Business sends a query to lower data storage and retrieval system module, inquires about the view data owner node of content search；
    (7) search data memory system is by finding corresponding view data owner node；
    (8) its manipulative indexing table of view data visitor node, view data owner node or picture material querying node, really Recognize with the presence or absence of the corresponding index of request content inquiry, if now view data visitor node and not finding request content The index of inquiry, then continue to inquire about next view data visitor node or picture material section by search data memory system Point；
    (9) index information that view data visitor node, view data owner node or picture material node will inquire Return to the Content query service for the node for initiating inquiry request；Now if the result returned includes the image of requesting query Data access person's node, then Content query service view data visitor's node listing is updated；
    (10) return information is forwarded to be serviced to local Agent；
    (11) if what is returned is sky, local Agent services on high in the clouds by obtaining content, otherwise according to the result slave phase of return Node is answered to obtain content.Then it is locally stored in service and stores the backup of this content, and issues corresponding index information to accordingly View data visitor node and view data owner node；
    (12) file system is sent content to.
CN201711006969.XA 2017-10-25 2017-10-25 Image retrieval method Active CN107798093B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711006969.XA CN107798093B (en) 2017-10-25 2017-10-25 Image retrieval method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711006969.XA CN107798093B (en) 2017-10-25 2017-10-25 Image retrieval method

Publications (2)

Publication Number Publication Date
CN107798093A true CN107798093A (en) 2018-03-13
CN107798093B CN107798093B (en) 2022-05-03

Family

ID=61533611

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711006969.XA Active CN107798093B (en) 2017-10-25 2017-10-25 Image retrieval method

Country Status (1)

Country Link
CN (1) CN107798093B (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110688510A (en) * 2018-06-20 2020-01-14 浙江宇视科技有限公司 Face background image acquisition method and system
CN111695419A (en) * 2020-04-30 2020-09-22 华为技术有限公司 Image data processing method and related device
CN113159211A (en) * 2021-04-30 2021-07-23 杭州好安供应链管理有限公司 Method, computing device and computer storage medium for similar image retrieval
CN113868440A (en) * 2020-06-30 2021-12-31 华为技术有限公司 Method, device, equipment and medium for managing feature library
CN115953584A (en) * 2023-01-30 2023-04-11 盐城工学院 End-to-end target detection method and system with learnable sparsity

Citations (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060072830A1 (en) * 2004-02-26 2006-04-06 Xerox Corporation Method for automated image indexing and retrieval
CN101217565A (en) * 2008-01-04 2008-07-09 华中科技大学 A network organization method of classification retrieval in peer-to-peer network video sharing system
US20090063445A1 (en) * 2007-08-27 2009-03-05 Arimilli Lakshminarayana B System and Method for Handling Indirect Routing of Information Between Supernodes of a Multi-Tiered Full-Graph Interconnect Architecture
CN101447937A (en) * 2009-02-27 2009-06-03 北京理工大学 Rapid data positioning method based on path division and multi-distributed-directory
US20090198956A1 (en) * 2008-02-01 2009-08-06 Arimilli Lakshminarayana B System and Method for Data Processing Using a Low-Cost Two-Tier Full-Graph Interconnect Architecture
CN101572726A (en) * 2008-05-01 2009-11-04 英特尔公司 Method and apparatus for hierarchical routing in multiprocessor mesh-based systems
CN102609434A (en) * 2010-12-20 2012-07-25 微软公司 Face recognition using social data
CN102982165A (en) * 2012-12-10 2013-03-20 南京大学 Large-scale human face image searching method
CN103036964A (en) * 2012-12-04 2013-04-10 杭州顺网科技股份有限公司 Internet bar server data updating method based on peer-to-peer (P2P)
CN103164452A (en) * 2011-12-15 2013-06-19 深圳市腾讯计算机系统有限公司 Method, device and system of picture storage and reading
CN103544261A (en) * 2013-10-16 2014-01-29 国家计算机网络与信息安全管理中心 Method and device for managing global indexes of mass structured log data
CN104301215A (en) * 2014-10-10 2015-01-21 北京邮电大学 Construction method of overlay network
CN104933154A (en) * 2015-06-25 2015-09-23 北京邮电大学 Method for quickly searching image according to content based on distributed network
CN105159895A (en) * 2014-05-28 2015-12-16 国际商业机器公司 Method and system for storing and inquiring data
CN105243084A (en) * 2015-09-07 2016-01-13 广东欧珀移动通信有限公司 Photographed image file storage method and system and photographed image file search method and system
CN105976399A (en) * 2016-04-29 2016-09-28 北京航空航天大学 Moving object detection method based on SIFT (Scale Invariant Feature Transform) feature matching
CN106599311A (en) * 2016-12-29 2017-04-26 广州市奥威亚电子科技有限公司 Cloud computation-based internet education platform resource library image retrieval method
CN106875325A (en) * 2017-02-14 2017-06-20 四川理工学院 One kind can search for resume image

Patent Citations (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060072830A1 (en) * 2004-02-26 2006-04-06 Xerox Corporation Method for automated image indexing and retrieval
US20090063445A1 (en) * 2007-08-27 2009-03-05 Arimilli Lakshminarayana B System and Method for Handling Indirect Routing of Information Between Supernodes of a Multi-Tiered Full-Graph Interconnect Architecture
CN101217565A (en) * 2008-01-04 2008-07-09 华中科技大学 A network organization method of classification retrieval in peer-to-peer network video sharing system
US20090198956A1 (en) * 2008-02-01 2009-08-06 Arimilli Lakshminarayana B System and Method for Data Processing Using a Low-Cost Two-Tier Full-Graph Interconnect Architecture
CN101572726A (en) * 2008-05-01 2009-11-04 英特尔公司 Method and apparatus for hierarchical routing in multiprocessor mesh-based systems
CN101447937A (en) * 2009-02-27 2009-06-03 北京理工大学 Rapid data positioning method based on path division and multi-distributed-directory
CN102609434A (en) * 2010-12-20 2012-07-25 微软公司 Face recognition using social data
CN103164452A (en) * 2011-12-15 2013-06-19 深圳市腾讯计算机系统有限公司 Method, device and system of picture storage and reading
CN103036964A (en) * 2012-12-04 2013-04-10 杭州顺网科技股份有限公司 Internet bar server data updating method based on peer-to-peer (P2P)
CN102982165A (en) * 2012-12-10 2013-03-20 南京大学 Large-scale human face image searching method
CN103544261A (en) * 2013-10-16 2014-01-29 国家计算机网络与信息安全管理中心 Method and device for managing global indexes of mass structured log data
CN105159895A (en) * 2014-05-28 2015-12-16 国际商业机器公司 Method and system for storing and inquiring data
CN104301215A (en) * 2014-10-10 2015-01-21 北京邮电大学 Construction method of overlay network
CN104933154A (en) * 2015-06-25 2015-09-23 北京邮电大学 Method for quickly searching image according to content based on distributed network
CN105243084A (en) * 2015-09-07 2016-01-13 广东欧珀移动通信有限公司 Photographed image file storage method and system and photographed image file search method and system
CN105976399A (en) * 2016-04-29 2016-09-28 北京航空航天大学 Moving object detection method based on SIFT (Scale Invariant Feature Transform) feature matching
CN106599311A (en) * 2016-12-29 2017-04-26 广州市奥威亚电子科技有限公司 Cloud computation-based internet education platform resource library image retrieval method
CN106875325A (en) * 2017-02-14 2017-06-20 四川理工学院 One kind can search for resume image

Non-Patent Citations (9)

* Cited by examiner, † Cited by third party
Title
VIJAYAKALYARI NATARAJAN EL: "Indexing and Retrieval of Images Using Wavelets-Segmentation Using Bisection K-Means Clustering", 《2011 FIRST INTERNATIONAL CONFERENCE ON INFORMATICS AND COMPUTATION INTELLIGENCE》 *
Z.TAN 等: "A New Routing Protocol of Structured Peer-to-Peer Overlay Networks", 《2008 FIFTH IEEE INTERNATIONAL SYMPOSIUM ON EMBEDDED COMPUTING》 *
刘欣 等: "BGP路由表中环形路由现象分析", 《计算机工程》 *
张一鸣 等: "《虚拟计算环境中的覆盖网构建技术》", 31 October 2010, 国防科技大学出版社 *
曹慧 等: "基于IPv6、簇和超节点的P2P路由模型研究", 《计算机工程》 *
曾维江 等: "一种基于路由表的工作流动态性能优化方法", 《计算机工程与应用》 *
林坤辉 等: "图像数据库基于内容检索的索引方法研究", 《厦门大学学报(自然科学版)》 *
蔡晓东 等: "基于图像特征索引的并行检索技术研究", 《电视技术》 *
邵美晶: "P2P网络关键技术研究与应用", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110688510A (en) * 2018-06-20 2020-01-14 浙江宇视科技有限公司 Face background image acquisition method and system
CN110688510B (en) * 2018-06-20 2022-06-14 浙江宇视科技有限公司 Face background image acquisition method and system
CN111695419A (en) * 2020-04-30 2020-09-22 华为技术有限公司 Image data processing method and related device
CN113868440A (en) * 2020-06-30 2021-12-31 华为技术有限公司 Method, device, equipment and medium for managing feature library
CN113868440B (en) * 2020-06-30 2023-06-27 华为技术有限公司 Feature library management method, device, equipment and medium
CN113159211A (en) * 2021-04-30 2021-07-23 杭州好安供应链管理有限公司 Method, computing device and computer storage medium for similar image retrieval
CN115953584A (en) * 2023-01-30 2023-04-11 盐城工学院 End-to-end target detection method and system with learnable sparsity

Also Published As

Publication number Publication date
CN107798093B (en) 2022-05-03

Similar Documents

Publication Publication Date Title
CN107798093A (en) Image search method
CN107330451B (en) Clothing attribute retrieval method based on deep convolutional neural network
US9348877B2 (en) Methods, systems, and data structures for performing searches on three dimensional objects
Li et al. Generating holistic 3d scene abstractions for text-based image retrieval
Rashno et al. Content-based image retrieval system with most relevant features among wavelet and color features
CN107679235A (en) searching system based on cloud platform
CN107729855B (en) Mass data processing method
KR102244086B1 (en) System for visual commonsense reasoning using knowledge graph
JP6173754B2 (en) Image search system, image search apparatus, and image search method
CN108595649A (en) The textile image search method of local invariant textural characteristics based on geometry
Srinivasa Rao et al. Content and context based image retrieval classification based on firefly-neural network
Gao et al. Occluded person re-identification based on feature fusion and sparse reconstruction
Wei et al. A novel color image retrieval method based on texture and deep features
Yu et al. The Related Techniques of content-based image retrieval
Manoharan et al. A comparison and analysis of soft computing techniques for content based image retrieval system
WO2012077818A1 (en) Method for determining conversion matrix for hash function, hash-type approximation nearest neighbour search method using said hash function, and device and computer program therefor
Lu et al. L 1-medial skeleton-based 3D point cloud model retrieval
Yoganand et al. Face detection approach from video with the aid of KPCM and improved neural network classifier
Dhobale et al. Semantic understanding of image content
Nagthane Image mining techniques and applications
Guo et al. Holistic indoor scene understanding by context-supported instance segmentation
Manso et al. Integrating planning perception and action for informed object search
Sarker et al. Transformer-Based Person Re-Identification: A Comprehensive Review
Arhid et al. A novel approach for partial shape matching and similarity based on data envelopment analysis
Li et al. The depth estimation of 3D face from single 2D picture based on manifold learning constraints

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant