CN100357960C - Parallel and distributing type identifying human face based on net - Google Patents

Abstract

The present invention relates to a parallel distributed human face recognition method based on grids. An improved human method based on line sections is used for making the recognition calculation and the operation of human face images to be allocated on grids, the strong calculation ability, the memory ability and the mutual cooperation of each node in grids are used for carrying out the calculation processing of human face recognition in parallel in a distribution mode, the speed of human face recognition is improved, and human faces can be recognized in real time based on that human face bases with a large amount of data are searched. Meanwhile, more human face characteristic line sections are selected by increasing resource consumption, and recognition accuracy is improved. The method comprise the operation procedures that grids are collocated, and an adjusting calculation node pool is established; human faces are extracted from video streams and are stored in a buffer memory area; images of human face rectangular area are extracted from images with human faces, normalization processing is carried out to the images, and the images are transmitted to a managing node; each calculation node executes the calculation operation of human face recognition together; the managing node carries out comprehensive analysis and processing to human face recognition result data.

Description

A kind of method of parallel, the distributing type identifying human face based on grid

Technical field

The present invention relates to two Computer Applied Technologies of a kind of comprehensive utilization and carry out the technology of Intelligent Recognition, exactly, the method that relates to a kind of parallel, distributing type identifying human face based on grid belongs to the comprehensive application technology field of multimedia mode identification and grid computing.

Background technology

Grid is a kind of distributed computing technology, utilizes network that other network element device of computing machine and its place network is concentrated in together, and the resource in the shared network is to promote computing power.At present, grid does not also have uniform definition, and Sun Microsystems thinks that grid is a system that provides single sign-on to arrive powerful distributed resource.IBM Corporation thinks that grid computing is a technology of utilizing the intrasystem processor of Internet connection large-scale virtual, realizes utilizing craving for of shared resource.

The research of grid originates from the mode that present people deal with problems.In the time will solving certain problem, at first can form a VO of Virtual Organization (Virtual Organization), i.e. the dynamic combined of some individuals, tissue or resource.Then, the researchist utilizes the various means of communication among the VO to carry out mutual and shared resource, and discussion is dealt with problems.Grid is exactly in Virtual Organization dynamic change, that have a plurality of departments or group, flexibly, safety, synergistically shared resource with find the solution problem.Therefore, the principal feature of grid is shared resource and dynamic schema, has only the software and hardware resources of fully sharing in the grid, and it is just meaningful to build grid; And shared object not only comprises the computing power or the resource of supercomputer in the grid, also comprises other various resources (as various software resources, storage resources, database resource, expensive instrument or the like) of sharing in the grid.In the practical application, grid is set up grid environment exactly between high-performance computer, realizes that the abundant of related resource of these computing machines shared, for the resources consumption client in the grid (being thin-client) provides service.

At present, the planning of grid and research object all are the grids in the setting range, and grid is shared normally non-trivial resource, the for example computing power of the high-performance node of supercomputer class and other resource, these resources must be transmitted electricity as the power station and fully be used for the user.Therefore, real-time, interactivity, efficient are required than the higher Distributed Application the constructed grid of grid software (as globustoolkit3) to be installed on common pc, do not have actual application value.

The purpose of grid is to share the non-trivial resource, so that for solving Virtual Organization's service of particular problem.So grid research is from how formulating the communication protocol of sharing these resources to begin, and how specifically to move on resource as for calculation task, then is responsible for processing by job scheduling software (as LSF, openPbs etc.).That is to say that grid only provides agreement and standard, set for this platform, create conditions, allow the user can share use resource wherein for shared resource; For how using of resource, then be responsible for voluntarily by grid itself.So, in the grid operational process, there is not special new technology, just, allow original soft hardware equipment play a greater role: except finishing local user's task, also will finish long-distance user's task by " sharing " technology.

Now, grid is in the growth stage, gridding technique is also in continuous development, along with going deep into to grid research, understanding, two kinds of grids have successively appearred: physical grid (physical grid)-various isomerization hardware resources (CPU, storage etc.) are organically combined, under the support of grid middleware, these resources are realized sharing and interoperability by network.Logical nets (logical grid)-service and/or the set of using realize the shared and interoperability of using and/or serving.Wherein, physical grid is the basis of construction logic grid, and physical grid can be shared by a plurality of logical netses institute, and a logical nets also can comprise a plurality of physical grid.

The research of grid is at first from physical grid.The research object that physical grid begins most is concrete isomerization hardware resource, purpose is that exploitation one group of application programming interfaces API, SDK (Software Development Kit) SDK (as: globus toolkits 1.0 softwares) operate these resources, thereby make up a grid prototype system as early as possible, prove the feasibility of grid.Afterwards, the research emphasis of grid was transferred to research to agreement from the exploitation of API, SDK, caused very important first mesh architecture First Five-Year Plan layer hourglass configuration having occurred in the evolution of grid.

Grid based on protocol construction, the development of grid is entered a new phase, but, various core protocol (as GRAM, LDAP, GRIDFTP) function is overlapping, and implementation has nothing in common with each other, bring very big inconvenience for user's use, objectively require each agreement to integrate.Along with the expansion of grid scale, judge dependence between the different assemblies, provide the demand of end-to-end managing griddings such as QoS also more and more stronger.In addition, in physical grid, resource is hardware such as CPU, bandwidth; But,, do not wish these bottom physical resources of direct control for the developer of grid application.So, press for a kind of new resource abstract mechanism.

The appearance and the application of web service have brought dawn for addressing the above problem.As a kind of new abstract mechanism-" service ", be exactly realization to the agreement of one group of definition specific function, promptly define one group of abstraction interface+service semantics.Wherein, interface embodies agreement by the binding definition of standard, and service semantics has the realization independence, also promptly can realize this interface by various language, this service of operation on various hardware resources.

The appearance of " service ", make the possibility that is configured to of logical nets, research emphasis is also transferred to based on the logical nets system of serving from customization physical grid standard agreement, and has caused the appearance of another important architecture OGSA on the Grid development history (Open Grid Service Architecture).The core of OGSA framework is " service ", and what emphasize is to the sharing of service, and being different from what emphasize in the physical grid is to share physical resource.Like this, visited each physical resource in the past and all will call distinct interface, and had the how many kinds of resource just to need the how many kinds of interface, and be difficult to form unification, the interface of standard and the situation of behavior and obtained change.In the OGSA structure, by " service " each concrete resource, information, data etc. are all united, help realizing flexible, unified, dynamic shared mechanism, make Distributed System Management that the interface and the behavior of standard arranged.

At present, face recognition technology has become a research focus in the multimedia field.Because different factors such as age, attitude, expression, illumination, facial image has the characteristics of " people thousand faces ", and therefore, face recognition technology has great challenge.In recent years, face recognition technology has been obtained rapid progress.But under, the ill-matched situation of object uncontrollable at extensive face database, imaging environment, even the recognition performance of at present best recognition system also can descend rapidly, the recognition system of this moment could enter practical application at all! Therefore, the recognition performance decline problem that how face identification system solves under imperfect imaging conditions (illumination variation, change of background, picture pick-up device difference) and object targetedly when mismatching (visual angle change, expression shape change, wear jewelry and even cosmetic), and the recognition of face speed issue of extensive face database all becomes the new focus of this problem.

The method of identification people face can be divided into based on several big classes such as feature, template, three-dimensional model or random series models.Olivier de Vel and Stefan Aeherhard (publish in IEEE Trans.Pattem Analysis and Machine Intelligence in " (Line-based face recognition undervarying pose ", vol.21,1999) a kind of relatively more novel being fit to of proposition walks abreast, the method for Distribution calculation.This method is selective or at random draw N bar straight line (referring to Fig. 1) between the outline line of people's face, characterize people's face with this N bar straight line, again all straight-line pass interpolation are normalized to the line segment L of same length, and are the quadratic sum of picture element gray-scale value error the distance definition between two line segments.Like this, for a given line segment l _kIf can in each bar line segment of face images, search out a line segment l with its difference minimum _j, l then _kPromptly be included into l _jThe image at place.Because a width of cloth facial image has many line segments, define a discriminant function again, the classification that decides whole facial image to belong to according to the image category that many line segments belonged to.This method has obtained very high discrimination on ORL (Olivetti Research Ltd.) facial image.The author thinks that this method has the following advantages: the first, and facial image is a stochastic sampling, the rotation on vertical guide has robustness to this algorithm to people's face, promptly allows people's face that certain rotation error is arranged.The second, owing to the line segment two ends are positioned on the outline line of people's face, its length relative fixed, so this algorithm has the yardstick unchangeability.The 3rd because line segment obtains from the view picture facial image, this algorithm to expression shape change, whether have jewelry to have stability, promptly it does not have influence to identification or influence is very little.

But the realization of this algorithm must solve following problems: at first, calculated amount is very huge.Because arbitrarily angled setting-out, interpolation that will be on facial image, and every bit be carried out record, calculated amount is very big.The N value that the author selects in training is 400, and promptly every width of cloth facial image is chosen 400 line segments.In fact, handle the calculated amount of 400 segment datas much larger than the calculated amount of handling a frame facial image.Secondly, the author supposes that the outline line of people's face is known, still, accurately obtain the facial contour line at any visual angle, is not an easy thing; Yet the author does not point out specifically to adopt any method.Moreover, owing to directly adopt the gray-scale value of image to calculate, thereby said method is relatively more responsive to the variation of intensity of illumination.

Summary of the invention

In view of this, the purpose of this invention is to provide a kind of walking abreast based on grid, the method of distributing type identifying human face, this method is the improvement based on the face identification method of line segment that above-mentioned Olivier de Vel and Stefan Aeherhard are proposed, it is powerful calculating ability and a storage capacity of utilizing grid, by cooperating with each other of each node in the grid, the prior art amount of calculation that vastness is heavy in the enterprising pedestrian's face of the face database of big data quantity identifying and the requirement of mass memory have been solved preferably, adopt the technological means of higher facial image recognizer of accuracy and increase resource consumption again, solved obtaining of facial image outline line, intensity of illumination changes problems such as sensitivity, has realized the facial image identification of degree of precision.And method of the present invention has good resource scalability, as long as insert more other grid by the internet, just can improve its computing power and storage capacity.

In order to achieve the above object, the invention provides a kind of parallel, distribution face identification method based on grid, it is characterized in that: calculating and operating portion that recognition of face is handled are deployed on the grid platform, the calculation process parallel by each computing node in the grid, that distributed earth carries out recognition of face is to improve recognition of face speed; Choose more face characteristic line segment simultaneously, to improve the recognition of face precision; Comprise following concrete operations step:

(1) configuration grid, set up adjustable computing node pond: dispose grid earlier, that sets up face database in management node and the virtual shared memory of grid is connected that (wherein the scheduling of virtual shared memory and management are prior aries, not at the row of the present invention's discussion), again the computing node that is in idle condition in the grid is concentrated and be constructed as " adjustable computing node pond ", and the running status of each parts in time informed management node, so that start mesh services; " initially "/" work " state of each camera of user side is set then;

(2) in video flowing, extract facial image, deposit buffer zone in: the frame buffer zone that the continuous videos stream of user side camera collection is stored in internal memory, again the every two field picture in the video flowing being carried out people's face in proper order detects, extraction contains the image of people's face, be stored in people's face buffer zone of internal memory, obtain the image of a series of people's of containing faces;

(3) image that contains people's face is extracted people's face rectangular area image, and this people's face rectangular area image carried out standardization, be transferred to management node again: order is extracted each two field picture that contains people's face from people's face buffer zone, and after extracting people's face rectangular area image respectively, carry out standardization according to statistical law and setting form again, each frame front face image is sent to management node in proper order;

(4) arithmetic operation of the concurrent execution recognition of face of each computing node: for every frame front face image of request identification, all from adjustable computing node pond, take out the computing node of a free time respectively, be used for this front face image is discerned processing by management node; Simultaneously, this computing node of deletion from adjustable computing node pond;

(5) management node carries out comprehensive analysis processing to the face recognition result data: when not having new recognition of face request in the video flowing, management node is discerned each computing node to a series of front face images result carries out overall treatment, draws final face recognition result.

Described step (1) further comprises following operation:

(11) use grid software with router and various computer combined, to build be a grid, in the virtual shared memory of grid face database is set, this face database is used for the front face image after the storage standards processing, and the front face image after every frame standardization all picked at random N bar straight line characterize it; Between the entrance and exit-management node of this grid and virtual shared memory, connect simultaneously, make management node can visit face database at any time; This management node also can connect other grids on the Internet, makes this grid have extendability;

(12) each computing node in the management node traversal grid according to each computing node state, adds adjustable computing node pond with the computing node free time, that can carry out recognition of face;

(13) when the user proposes beginning or stops the recognition of face task, management node asks to open or close each camera that is positioned at user side according to the user, and initialize routine and comprise that at least people's face characterizes each predefine parameter of the whole grid system of the reference coefficient in line segment parameter N, determinant width s and the comprehensive recognition of face;

(14) associated components that includes but not limited to user side, camera, adjustable computing node pond and face database in the grid is incited somebody to action running status announcement management node separately respectively, management node according to the status information after converging is judged whether can start or close mesh services, deletion or restore data.

Described step (2) further comprises following operation:

(21) the user side camera enters the camera work state according to the enabled instruction of management node;

(22) be that unit sequence is stored in frame buffer zone with the image in the continuous videos stream of camera collection with " frame ", wherein, present frame does not detect a two field picture of handling for this camera collection carries out people's face constantly the earliest, as yet;

(23) by people's face sorter every two field picture in the frame buffer zone is carried out people's face and detect, simultaneously, camera works on, and constantly the video flowing of gathering is write frame buffer zone; Described people's face sorter is to train the face characteristic file that obtains to constitute by loading to people's face sample storehouse, and people's face sample storehouse comprises that a large amount of people's faces and easy misunderstanding are the non-face two class images of people's face;

(24) if in present frame, detect and contain people's face, then present frame is stored in people's face buffer zone, and shows that at user side this present frame and mark are extracted the position and the size of human face region; Otherwise, only show this present frame at user side;

(25) circulation execution in step (23), (24) operation detect until people's face of finishing whole two field pictures.

Described step (3) further comprises following operation:

(31) according to human eye feature file and human eye sample storehouse, every two field picture in people's face buffer zone is detected human face region and determine the operation of these two positions of people in facial image, if have only a glance in the facial image or do not have eyes or eye feature not obvious, then calculate its Position Approximate of two in the face ratio; Wherein the human eye feature file then forms by human eye sample storehouse is trained, and human eye sample storehouse comprises a large amount of human eyes and easily is misinterpreted as the non-human eye two class images of human eye;

(32) obtain the connecting line of eyes according to two positions in the facial image, judge again whether this eyes connecting line is parallel to horizontal line, if, redirect execution in step (34); Otherwise order is carried out subsequent operation;

(33) rotate this facial image, make its eyes line be parallel to horizontal line; Image rotating can cause losing the edge pixel point of this facial image, and the value of the pixel of losing is changed to 0 without exception; Because people's face rectangular area generally can be near the edge, the pixel of losing does not have influence for identification or influence is little;

(34) with the human eye be benchmark, with eyes be yardstick apart from d, from this facial image, intercept the square area behaviour face of a 2d * 2d according to " five in three front yards " principle, unify convergent-divergent according to predetermined size again, obtain the positive criteria facial image;

(35) will be sent to management node through the positive criteria facial image that above-mentioned steps is handled.

Described step (4) further comprises following operation:

(41) every frame front face image of discerning for request, management node takes out first idle computing node respectively from adjustable computing node pond, be used for this front face image is discerned processing, simultaneously, this first computing node of deletion from adjustable computing node pond;

(42) first computing node adopts improved face identification method based on line segment that facial image is discerned processing, be about to a front face image and be divided into N bar line segment, characterize facial image with this N bar line segment, thus with a front face image be converted to this N bar line segment one by one respectively with face database in the relatively identification of all line segments of each facial image;

(43) first computing node takes out another idle computing node from adjustable computing node pond, be used for comparison process to all line segments of each facial image in certain the bar line segment of described request identification facial image and the face database, simultaneously, this invoked another computing node of deletion from adjustable computing node pond; This another computing node continues to take out the comparison between each line segment that other computing node carries out each facial image in other line segment and the face database respectively from adjustable computing node pond, simultaneously, this invoked other computing node of deletion from adjustable computing node pond, after returning to another computing node that calls it when comparative result, this other computing node then is added to adjustable computing node pond; After another computing node is finished the comparison of each line segment in certain bar line segment and the face database, find in the face database with this certain bar line segment recently like a line segment, write down the facial image that the nearest liny section of this certain bar is belonged to, and the result returns to first computing node that calls it as a comparison; Thereupon, this another computing node is added to adjustable computing node pond;

(44) operation of repetition above-mentioned steps (43) after the comparison process of finishing whole N bar line segments, obtains N comparative result: N frame facial image; At this moment, each computing node is sent to management node with the relevant information of the facial image that occurrence number is maximum in the N frame facial image separately respectively.

Improved face identification method is that every frame facial image is characterized with N bar line segment in the described step (42), again with the grayvalue transition of every line chart picture for intensity of illumination is changed insensitive singular value, be used to carry out the computing of recognition of face; This recognition methods also comprises following initialization operation step:

(A1) after choosing 2 lines on the facial image outline line arbitrarily, can form a line segment, utilize the resulting N bar of this method line segment to characterize facial image, N is made as the natural number greater than 200 usually;

(A2) utilize the linear interpolation algorithm of Flame Image Process, this N bar line segment is carried out the interpolation normalizing handle, it all is converted to same length is s ²Collection of pixels, promptly every line segment all is made up of s little line segment, each little line segment all has s pixel, s is a positive integer; Pixel numerical value with every line segment changes into a s * s matrix like this: the pixel value of preceding s pixel in first trip and the line segment is corresponding, and second row is corresponding to the pixel value of s+s pixel with the s+1 in the line segment, and by that analogy, it is capable to have s; Find the solution the singular value of this s * s matrix again, the grayvalue transition of every line chart picture for intensity of illumination is changed insensitive singular value, is handled so that replace the relatively identification that gray-scale value in the facial image carries out each line segment of facial image with the singular value of this line segment.

The computing formula that in the described step (A2) the pixel numerical value of every in facial image line segment is changed into a s * s matrix is:

That is:

$M_{L_{i,k}}=\left|\begin{array}{cccc}{L}_{i,k}^{\left(1\right)}& L_{i,k}^{\left(2\right)}& ...& L_{i,k}^{\left(s\right)}\\ L_{i,k}^{(s+1)}& L_{i,k}^{(s+2)}& ...& L_{i,k}^{(2\×s)}\\ .................& ...................& ............& ................\\ L_{i,k}^{(s^2-s+1)}& L_{i,k}^{(s^2-s+2)}& ...& L_{i,k}^{(s\×s)}\end{array}\right|$

In the formula, L _{I, k}Be any line segment in the front face image, it contains s ²Individual pixel, natural number i, k and j are respectively the frame number of the every frame facial image in the face database, the sequence number of each line segment in every frame facial image and the pixel sequence number on each line segment, and the maximal value of k is N; L _{I, k} ^(j)K pixel in the j bar line segment in the expression front face image storehouse in the i frame facial image; M _{Li, k}Expression is by line segment L _{I, k}S * s matrix of changing and getting.

The concrete grammar that the relatively identification of using the s * s singular values of a matrix of finding the solution each line segment of facial image to carry out this each line segment of facial image in the described step (A2) is handled is:

(A21) suppose joint owner's face image M frame in the face database, its sequence number is respectively F ₁, F ₂... F _M, on every frame facial image, choose N bar straight line arbitrarily, the total line segment of relatively identification that needs is in the face database like this: M * N=MN bar is respectively f ₁, f ₂... f _{M * N}In addition, N bar line segment is arranged also on the facial image X to be identified, be respectively x ₁, x ₂... x _N

(A22) with any line segment x among the facial image X to be identified _kWith the M * N bar line segment f in the face database ₁, f ₂,,,,, f _{M * N}Compare processing respectively, wherein, natural number k is the line segment sequence number, and maximal value is N, and a such matching process will carry out the relatively calculating of M * N to line segment:＜x _k, f ₁,＜x _k, f ₂...,＜x _k, f _{M * N}, therefrom find out and this line segment x _kA most similar line segment f _kIf this line segment f _kThe j frame that belongs to facial image in the face database, then in this facial image X to be identified and the face database j frame facial image the match is successful once;

(A23) according to the comparative approach of step (A22), successively to the N bar line segment on the facial image X to be identified respectively with face database in MN bar line segment compare processing, promptly finish matching process N time;

(A24) if having certain frame facial image and the facial image X number of times that the match is successful to be identified maximum in the face database, then this frame facial image is exactly the people face the most similar to facial image X to be identified.

Described step (5) further comprises following operation:

(51) user side is communicated by letter with management node, and informing does not have new recognition of face request in the management node video flowing, and stops to transmit video image to management node;

(52) management node result images and correlation parameter that each computing node is discerned a series of facial images gathers comprehensively, obtain with face database in the maximum facial image of matching times, and record matching times, find the solution this matching times again divided by the resulting ratio of a series of facial image sums, if this ratio, thinks then that the maximum facial image of this matching times is final face recognition result greater than the reference frame numerical value in the comprehensive recognition of face of described setting; Otherwise, think and can not discern;

(53) management node is communicated by letter with user side, informs the net result and the related data information of user side recognition of face;

(54) user side is shown to the user with the net result and the related data statistical information of recognition of face.

The present invention is a kind of improved face identification method based on line segment, promptly based on the method for parallel, the distributing type identifying human face of grid, characteristics are that recognition of face is incorporated into grid platform, make full use of the grid characteristic: the collaboration method of multicompartment in powerful calculating ability, mass memory ability and the grid, solved the requirement of recognition of face preferably for computing power and storage capacity, thereby can ignore calculating strength, improve recognition of face speed greatly, fully satisfy the real-time identification requirement; And this method improves the speed and the precision of recognition of face by complexity that improves face recognition algorithms and resource consumptions such as increasing storer, can satisfy the real-time identification requirement of carrying out recognition of face in the magnanimity face database.In addition, the management node in the grid makes this method have certain " examining oneself " ability, i.e. the fault-tolerance and the stability of system have been strengthened in sustainable operation under no artificial intervention situation.In addition, this method also has good resource scalability, can improve the computing power and the storage capacity of native system by the mode that inserts more other grids.

The inventive method all has good application prospects in a lot of fields.For example:

Safe examination system: take the facial photo of being examined the people with camera, and the image library that is distributed in a plurality of places is carried out real-time retrieval, can be used for the counterterrorist activity in important places such as airport, customs, station, Olympic venue.

The news program management: at present, news video data accumulative total was above ten million hour.The news program that the TV station that has broadcasts every day surpasses 20 hours, and send into the video new data of TV station above 300 hour from various channels every day.TV station will provide detailed report to spectators, necessarily requires the news production personnel can utilize new live video data and old historical summary fast, efficiently.And adopt facial image retrieval technique based on grid that the news video program is managed and file, can realize news program producer's above-mentioned requirements.

Network video image information retrieval: on the internet a large amount of video image information (photo or dynamic video) is arranged at present, present search engine can only be retrieved the key word of text, search for according to facial image, will improve the retrieval capability of the network information greatly.

Therefore, the present invention adopts the method for gridding technique identification facial image can produce good social benefit and economic benefit.

Description of drawings

Fig. 1 is selective or choose the synoptic diagram of N bar line segment arbitrarily between the outline line of facial image.

Fig. 2 is the method overall procedure block scheme that the present invention is based on parallel, the distributing type identifying human face of grid.

Fig. 3 is the grid system structural representation of utilization the inventive method.

How Fig. 4 (A), (B) intercept a square area that is of a size of 2d * 2d as the signal of front face image and the standardization front face image of an embodiment according to " five in three front yards " principle from facial image.

Embodiment

For making the purpose, technical solutions and advantages of the present invention clearer, the present invention is described in further detail below in conjunction with accompanying drawing.

As everyone knows, the calculated amount that facial image identification relates to is very huge, the present invention is as a kind of parallel, distributed face identification method based on grid, it is that calculating and operating portion that recognition of face is handled are deployed on the grid platform, the calculation process parallel by each computing node in the grid, that distributed earth carries out recognition of face is to reduce the recognition of face time; Simultaneously, can choose more straight line, to improve the precision of recognition of face.

Referring to Fig. 2, introduce the concrete operations step of the inventive method:

Step (1): configuration grid, set up adjustable computing node pond-elder generation's configuration grid, set up being connected of face database in management node and the virtual shared memory of grid, again the computing node that is in idle condition in the grid is concentrated and be constructed as " adjustable computing node pond ", and the running status of each parts in time informed management node, so that start mesh services; " initially "/" work " state of each camera of user side is set then.

This step can be subdivided into following content of operation again;

(11) use grid software with router and various computer combined, to build be that (referring to Fig. 3, wherein management node is the entrance and exit of grid to a grid, and its support programs are grid software globustoolkits; The virtual shared memory that also includes a plurality of computing nodes and form) by data server and memory node, in the virtual shared memory of grid face database is set, this face database is used for front face image after the storage standardsization, and every frame facial image all picked at random N bar straight line characterize it; Between the entrance and exit-management node of this grid and virtual shared memory, connect simultaneously, make management node can visit face database at any time; This management node also can connect other grids on the Internet, makes this grid have extendability;

(12) each computing node in the management node traversal grid according to each computing node state, adds adjustable computing node pond with the computing node free time, that can carry out recognition of face;

(13) when the user proposes beginning or stops the recognition of face task, management node asks to open or close each camera that is positioned at user side according to the user, and initialize routine and each predefine parameter that comprises the whole grid systems such as reference coefficient in people's face sign line segment parameter N, determinant width s and the comprehensive recognition of face;

(14) associated components such as the user side in the grid, camera, adjustable computing node pond and face database are incited somebody to action running status announcement management node separately respectively, management node according to the status information after converging is judged whether can start or close mesh services, deletion or restore data.

Step (2): in video flowing, extract facial image, deposit the buffer area-continuous videos stream of user side camera collection the is stored in frame buffer zone of internal memory in, again the every two field picture in the video flowing being carried out people's face in proper order detects, extraction contains the image of people's face, be temporarily stored in people's face buffer zone of internal memory, obtain the image of a series of people's of containing faces.

This step can be subdivided into following content of operation again;

(21) the user side camera enters the camera work state according to the enabled instruction of management node;

(22) be that unit sequence is stored in frame buffer zone with the image in the continuous videos stream of camera collection with " frame ", wherein, present frame does not detect a two field picture of handling for this camera collection carries out people's face constantly the earliest, as yet;

(23) by people's face sorter every two field picture in the frame buffer zone is carried out people's face and detect, simultaneously, camera works on, and constantly the video flowing of gathering is write frame buffer zone; Wherein people's face sorter is to train the face characteristic file that obtains to constitute by loading to people's face sample storehouse, and people's face sample storehouse comprises that a large amount of people's faces and easy misunderstanding are the non-face two class images of people's face;

(24) if in present frame, detect and contain people's face, then present frame is stored in people's face buffer zone, and shows that at user side this present frame and mark are extracted the position and the size of human face region; Otherwise, only show this present frame at user side;

(25) circulation execution in step (23), (24) operation detect until people's face of finishing whole two field pictures.

Step (3): the image that contains people's face is extracted people's face rectangular area image, and it is carried out standardization, be transferred to the order of management node-from people's face buffer zone again and extract each two field picture that contains people's face, and after extracting the human face region image respectively, carry out standardization according to statistical law and setting form again, each frame front face image is sent to management node in proper order.

This step can be subdivided into following content of operation again;

(31) according to human eye feature file and human eye sample storehouse, every two field picture in people's face buffer zone is detected human face region and determine the operation of these two positions of people in facial image, if have only a glance in the facial image or do not have eyes or eye feature not obvious, then calculate its Position Approximate of two in the face ratio; Wherein the human eye feature file then forms by human eye sample storehouse is trained, and human eye sample storehouse comprises a large amount of human eyes and easily is misinterpreted as the non-human eye two class images of human eye;

(32) obtain the connecting line of eyes according to two positions in the facial image, judge again whether this eyes connecting line is parallel to horizontal line, if, redirect execution in step (34); Otherwise order is carried out subsequent operation;

(33) rotate this facial image, make its eyes line be parallel to horizontal line; Image rotating can cause losing the edge pixel point of this facial image, and the value of the pixel of losing is changed to 0 without exception;

(34) with the human eye be benchmark, is yardstick with eyes apart from d, from this facial image, intercept the square area behaviour face of a 2d * 2d (referring to Fig. 4 according to " five in three front yards " principle, this figure has showed how to intercept a square area that is of a size of 2d * 2d as a front face image and an embodiment facial image from facial image according to " five in three front yards " principle), unify convergent-divergent according to predetermined size again, obtain the positive criteria facial image;

(35) will be sent to management node through the positive criteria facial image that above-mentioned steps is handled.

Step (4): the arithmetic operation-management node of the concurrent execution recognition of face of each computing node is for every frame front face image of request identification, from adjustable computing node pond, take out the computing node of a free time respectively, this front face image is discerned processing; Simultaneously, this computing node of deletion from adjustable computing node pond.

This step can be subdivided into following content of operation again;

(41) every frame front face image of discerning for request, management node takes out first idle computing node respectively from adjustable computing node pond, be used for this front face image is discerned processing, simultaneously, this first computing node of deletion from adjustable computing node pond;

(42) first computing node adopts improved face identification method based on line segment that facial image is discerned processing, be about to a front face image and be divided into N bar line segment, characterize facial image with this N bar line segment, thus with a front face image be converted to this N bar line segment one by one respectively with face database in the relatively identification of all line segments of each facial image; Wherein improvements be grayvalue transition with every line chart picture for intensity of illumination is changed insensitive singular value, be used to carry out the computing of recognition of face;

Following mask body is introduced each operation steps of this recognition methods:

(A1) after choosing 2 lines on the facial image outline line arbitrarily, can form a line segment, utilize the resulting N bar of this method line segment to characterize facial image, N is made as the natural number greater than 200 usually;

(A2) utilize the linear interpolation algorithm of Flame Image Process, this N bar line segment is carried out the interpolation normalizing handle, it all is converted to same length is s ²Collection of pixels, promptly every line segment all is made up of s little line segment, each little line segment all has s pixel, s is a positive integer; Pixel numerical value with every line segment changes into a s * s matrix like this: the pixel value of preceding s pixel in first trip and the line segment is corresponding, and second row is corresponding to the pixel value of s+s pixel with the s+1 in the line segment, and by that analogy, it is capable to have s;

The computing formula of line segment pixels point is:

That is:

$M_{L_{i,k}}=\left|\begin{array}{cccc}{L}_{i,k}^{\left(1\right)}& L_{i,k}^{\left(2\right)}& ...& L_{i,k}^{\left(s\right)}\\ L_{i,k}^{(s+1)}& L_{i,k}^{(s+2)}& ...& L_{i,k}^{(2\×s)}\\ ...............& ...............& ..........& ................\\ L_{i,k}^{(s^2-s+1)}& L_{i,k}^{(s^2-s+2)}& ...& L_{i,k}^{(s\×s)}\end{array}\right|$

Formula

In,

L _{I, k}Be any line segment in the front face image, it contains s ²Individual pixel, natural number i, k and j are respectively the frame number of the every frame facial image in the face database, the sequence number of each line segment in every frame facial image and the pixel sequence number on each line segment, and the maximal value of k is N; L _{I, k} ^(j)K pixel in the j bar line segment in the expression front face image storehouse in the i frame facial image; M _{Li, k}Expression is by line segment L _{I, k}S * s matrix of changing and getting;

Then, find the solution the singular value of above-mentioned s * s matrix, the grayvalue transition of every line chart picture for intensity of illumination is changed insensitive singular value, is handled so that replace the relatively identification that gray-scale value in the facial image carries out each line segment of facial image with the singular value of this line segment.

(A3) compare identification respectively with each line segment in each facial image in the s * s singular values of a matrix of each line segment of facial image and the face database, the concrete operations step of this process is:

(A31) suppose joint owner's face image M frame in the face database, its sequence number is respectively F ₁, F ₂... F _M, on every frame facial image, choose N bar straight line arbitrarily, the total line segment of relatively identification that needs is in the face database like this: M * N=MN bar is respectively f ₁, f ₂... f _{M * N}In addition, N bar line segment is arranged also on the facial image X to be identified, be respectively x ₁, x ₂... x _N

(A32) with any line segment x among the facial image X to be identified _kWith the M * N bar line segment f in the face database ₁, f ₂,,,,, f _{M * N}Compare processing respectively, wherein, natural number k is the line segment sequence number, and maximal value is N, and a such matching process will carry out the relatively calculating of M * N to line segment:＜x _k, f ₁,＜x _k, f ₂...,＜x _k, f _{M * N}, therefrom find out and this line segment x _kA most similar line segment f _kIf this line segment f _kBelong to j frame facial image in the face database, then in this facial image X to be identified and the face database j frame facial image the match is successful once;

(A33) according to the comparative approach of step (A32), with the N bar line segment on the facial image X to be identified successively respectively with face database in MN bar line segment compare processing, promptly finish matching process N time;

(A34) if having certain frame facial image and the facial image X number of times that the match is successful to be identified maximum in the face database, then this frame facial image is exactly the people face the most similar to facial image X to be identified.

(43) first computing node takes out another idle computing node from adjustable computing node pond, be used for comparison process to all line segments of each facial image in certain the bar line segment of described request identification facial image and the face database, simultaneously, this invoked another computing node of deletion from adjustable computing node pond; Other computing node of taking-up carries out the comparison between any line segment in other line segment and the face database from adjustable computing node pond and another computing node can continue, simultaneously, this invoked other computing node of deletion from adjustable computing node pond, after returning to another computing node that calls it when comparative result, this other computing node is added to adjustable computing node pond; After another computing node is finished the comparison of each line segment in this line segment and the face database, find in the face database with this line segment recently like a line segment, write down the facial image that this nearest liny section is belonged to, and the result returns to first computing node that calls it as a comparison; Thereupon, this another computing node is added to adjustable computing node pond;

(44) operation of repetition above-mentioned steps (43) after the comparison process of finishing whole N bar line segments, obtains N comparative result: N frame facial image; At this moment, each computing node is sent to management node with the relevant information of the facial image that occurrence number is maximum in the N frame facial image separately respectively.

Step (5): management node carries out comprehensive analysis processing-when not having new recognition of face request in the video flowing to the face recognition result data, management node is discerned each computing node to a series of front face images result carries out overall treatment, draws final face recognition result.

This step can be subdivided into following content of operation again;

(51) user side is communicated by letter with management node, and informing does not have new recognition of face request in the management node video flowing, and stops to transmit video image to management node;

(52) management node result images and correlation parameter that each computing node is discerned a series of facial images gathers comprehensively, obtain with face database in the maximum facial image of matching times, and record matching times, find the solution this matching times again divided by the resulting ratio of a series of facial image sums, if this ratio thinks then that greater than the reference frame numerical value in the comprehensive recognition of face of setting the maximum facial image of this matching times is final face recognition result; Otherwise, think and can not discern;

(53) management node is communicated by letter with user side, informs the net result and the related data information of user side recognition of face;

(54) user side is shown to the user with the net result and the related data statistical information of recognition of face.

The applicant has carried out implementing test to method of the present invention.Test is to have built a grid system with China invites the person's gigabit LAN, the management node of this grid system be to use the CPU model be Xeon 3.2GHz, in to save as 2G byte, hard disk be that two active and standby structural devices of 73G byte are formed, computing node is to use 40 computing machines of same specification to form, and the capacity of storage array is 2T.Wherein server side operation system and program compiler are respectively Linux and gcc service routine and javac service framework, client operating system and program compiler be respectively Windows and VC, java SDK (Software Development Kit) JDK (Java Developer ' sKit), based on the build instrument Ant of Java.Database software is Postgresql, and Web publishing software is Apache tomcat, and support platform software is Globus and CGSP.

The picture that this pilot system is tested has 210, is divided into 35 groups, every group 6 pictures.Test result data is: the average test response time of every pictures is about 1.5 seconds, wherein discerns successfully 176, and mistake identification maybe can not be discerned 34, and success ratio is 83.81%.Therefore, under the situation of using high complexity face recognition algorithms and big data quantity face database to compare, guaranteed the real-time identification requirement of people's face; And, not having under the artificial situation of intervening, this grid system is continuous service for a long time, can increase the grid cluster of grid internal calculation node or grid outside easily, improves computing power and storage capacity.Test findings shows: face identification method of the present invention is successful, has realized goal of the invention.

Claims (9)

1, a kind of parallel, distributed face identification method based on grid, it is characterized in that: calculating and operating portion that recognition of face is handled are deployed on the grid platform, the calculation process parallel by each computing node in the grid, that distributed earth carries out recognition of face is to improve recognition of face speed; Choose more face characteristic line segment simultaneously, to improve the recognition of face precision; Comprise following concrete operations step:

(1) configuration grid, set up adjustable computing node pond: dispose grid earlier, set up being connected of face database in management node and the virtual shared memory of grid, again the computing node that is in idle condition in the grid is concentrated and be constructed as " adjustable computing node pond ", and the running status of each parts in time informed management node, so that start mesh services; " initially "/" work " state of each camera of user side is set then;

(2) in video flowing, extract facial image, deposit buffer zone in: the frame buffer zone that the continuous videos stream of user side camera collection is stored in internal memory, again the every two field picture in the video flowing being carried out people's face in proper order detects, extraction contains the image of people's face, be stored in people's face buffer zone of internal memory, obtain the image of a series of people's of containing faces;

(3) image that contains people's face is extracted people's face rectangular area image, and this people's face rectangular area image carried out standardization, be transferred to management node again: order is extracted each two field picture that contains people's face from people's face buffer zone, and after extracting people's face rectangular area image respectively, carry out standardization according to statistical law and setting form again, each frame front face image is sent to management node in proper order;

(4) arithmetic operation of the concurrent execution recognition of face of each computing node: for every frame front face image of request identification, all from adjustable computing node pond, take out the computing node of a free time respectively, be used for this front face image is discerned processing by management node; Simultaneously, this computing node of deletion from adjustable computing node pond;

(5) management node carries out comprehensive analysis processing to the face recognition result data: when not having new recognition of face request in the video flowing, management node is discerned each computing node to a series of front face images result carries out overall treatment, draws final face recognition result.

2, face identification method according to claim 1 is characterized in that: described step (1) further comprises following operation:

(11) use grid software with router and various computer combined, to build be a grid, in the virtual shared memory of grid face database is set, this face database is used for the front face image after the storage standards processing, and the front face image after every frame standardization all picked at random N bar straight line characterize it; Between the entrance and exit-management node of this grid and virtual shared memory, connect simultaneously, make management node can visit face database at any time; This management node also can connect other grids on the Internet, makes this grid have extendability;

(12) each computing node in the management node traversal grid according to each computing node state, adds adjustable computing node pond with the computing node free time, that can carry out recognition of face;

(13) when the user proposes beginning or stops the recognition of face task, management node asks to open or close each camera that is positioned at user side according to the user, and initialize routine and comprise that at least people's face characterizes each predefine parameter of the whole grid system of the reference coefficient in line segment parameter N, determinant width s and the comprehensive recognition of face;

(14) associated components that includes but not limited to user side, camera, adjustable computing node pond and face database in the grid is incited somebody to action running status announcement management node separately respectively, management node according to the status information after converging is judged whether can start or close mesh services, deletion or restore data.

3, face identification method according to claim 1 is characterized in that: described step (2) further comprises following operation:

(21) the user side camera enters the camera work state according to the enabled instruction of management node;

(22) be that unit sequence is stored in frame buffer zone with the image in the continuous videos stream of camera collection with " frame ", wherein, present frame does not detect a two field picture of handling for this camera collection carries out people's face constantly the earliest, as yet;

(23) by people's face sorter every two field picture in the frame buffer zone is carried out people's face and detect, simultaneously, camera works on, and constantly the video flowing of gathering is write frame buffer zone; Described people's face sorter is to train the face characteristic file that obtains to constitute by loading to people's face sample storehouse, and people's face sample storehouse comprises that a large amount of people's faces and easy misunderstanding are the non-face two class images of people's face;

(24) if in present frame, detect and contain people's face, then present frame is stored in people's face buffer zone, and shows that at user side this present frame and mark are extracted the position and the size of human face region; Otherwise, only show this present frame at user side;

(25) circulation execution in step (23), (24) operation detect until people's face of finishing whole two field pictures.

4, face identification method according to claim 1 is characterized in that: described step (3) further comprises following operation:

(31) according to human eye feature file and human eye sample storehouse, every two field picture in people's face buffer zone is detected human face region and determine the operation of these two positions of people in facial image, if have only a glance in the facial image or do not have eyes or eye feature not obvious, then calculate its Position Approximate of two in the face ratio; Wherein the human eye feature file then forms by human eye sample storehouse is trained, and human eye sample storehouse comprises a large amount of human eyes and easily is misinterpreted as the non-human eye two class images of human eye;

(32) obtain the connecting line of eyes according to two positions in the facial image, judge again whether this eyes connecting line is parallel to horizontal line, if, redirect execution in step (34); Otherwise order is carried out subsequent operation;

(33) rotate this facial image, make its eyes line be parallel to horizontal line; Image rotating can cause losing the edge pixel point of this facial image, and the value of the pixel of losing is changed to 0 without exception;

(34) with the human eye be benchmark, with eyes be yardstick apart from d, from this facial image, intercept the square area behaviour face of a 2d * 2d according to " five in three front yards " principle, unify convergent-divergent according to predetermined size again, obtain the positive criteria facial image;

(35) will be sent to management node through the positive criteria facial image that above-mentioned steps is handled.

5, face identification method according to claim 1 is characterized in that: described step (4) further comprises following operation:

(41) every frame front face image of discerning for request, management node takes out first idle computing node respectively from adjustable computing node pond, be used for this front face image is discerned processing, simultaneously, this first computing node of deletion from adjustable computing node pond;

(42) first computing node adopts improved face identification method based on line segment that facial image is discerned processing, be about to a front face image and be divided into N bar line segment, characterize facial image with this N bar line segment, thus with a front face image be converted to this N bar line segment one by one respectively with face database in the relatively identification of all line segments of each facial image;

(43) first computing node takes out another idle computing node from adjustable computing node pond, be used for comparison process to all line segments of each facial image in certain the bar line segment of described request identification facial image and the face database, simultaneously, this invoked another computing node of deletion from adjustable computing node pond; This another computing node continues to take out the comparison between each line segment that other computing node carries out each facial image in other line segment and the face database respectively from adjustable computing node pond, simultaneously, this invoked other computing node of deletion from adjustable computing node pond; After returning to another computing node that calls it when comparative result, this other computing node then is added to adjustable computing node pond; After another computing node is finished the comparison of each line segment in certain bar line segment and the face database, find in the face database with this certain bar line segment recently like a line segment, write down the facial image that the nearest liny section of this certain bar is belonged to, and the result returns to first computing node that calls it as a comparison; Thereupon, this another computing node is added to adjustable computing node pond;

(44) operation of repetition above-mentioned steps (43) after the comparison process of finishing whole N bar line segments, obtains N comparative result: N frame facial image; At this moment, each computing node is sent to management node with the relevant information of the facial image that occurrence number is maximum in the N frame facial image separately respectively.

6, face identification method according to claim 5, it is characterized in that: improved face identification method is that every frame facial image is characterized with N bar line segment in the described step (42), again with the grayvalue transition of every line chart picture for intensity of illumination is changed insensitive singular value, be used to carry out the computing of recognition of face; This method also comprises following initialization operation step:

(A1) after choosing 2 lines on the facial image outline line arbitrarily, can form a line segment, utilize the resulting N bar of this method line segment to characterize facial image, N is made as the natural number greater than 200 usually;

(A2) utilize the linear interpolation algorithm of Flame Image Process, this N bar line segment is carried out the interpolation normalizing handle, it all is converted to same length is s ²Collection of pixels, promptly every line segment all is made up of s little line segment, each little line segment all has s pixel, s is a positive integer; Pixel numerical value with every line segment changes into a s * s matrix like this: the pixel value of preceding s pixel in first trip and the line segment is corresponding, and second row is corresponding to the pixel value of s+s pixel with the s+1 in the line segment, and by that analogy, it is capable to have s; Find the solution the singular value of this s * s matrix again, the grayvalue transition of every line chart picture for intensity of illumination is changed insensitive singular value, is handled so that replace the relatively identification that gray-scale value in the facial image carries out each line segment of facial image with the singular value of this line segment.

7, face identification method according to claim 6 is characterized in that: the computing formula that in the described step (A2) the pixel numerical value of every in facial image line segment is changed into a s * s matrix is:

That is:

$M_{L_{i,k}}=\left|\begin{array}{cccc}{L}_{i,k}^{\left(1\right)}& L_{i,k}^{\left(2\right)}& \·\·\·& L_{i,k}^{\left(s\right)}\\ L_{i,k}^{(s+1)}& L_{i,k}^{(s+2)}& \·\·\·& L_{i,k}^{(2\×s)}\\ \·\·\·\·\·\·& \·\·\·\·\·\·& \·\·\·\·\·\·& \·\·\·\·\·\·\\ L_{i,k}^{(s^2-s+1)}& L_{i,k}^{(s^2-s+2)}& \·\·\·& L_{i,k}^{(s\×s)}\end{array}\right|$

In the formula, L _{I, k}Be any line segment in the front face image, it contains s ²Individual pixel, natural number i, k and j are respectively the frame number of the every frame facial image in the face database, the sequence number of each line segment in every frame facial image and the pixel sequence number on each line segment, and the maximal value of k is N; L _{I, k} ^(j)J pixel in the k bar line segment in the expression front face image storehouse in the i frame facial image; M _{Li, k}Expression is by line segment L _{I, k}S * s matrix of changing and getting.

8, face identification method according to claim 6 is characterized in that: the concrete grammar that the relatively identification of using the s * s singular values of a matrix of finding the solution each line segment of facial image to carry out this each line segment of facial image in the described step (A2) is handled is:

(A21) suppose joint owner's face image M frame in the face database, its sequence number is respectively F ₁, F ₂... F _M, on every frame facial image, choose N bar straight line arbitrarily, the total line segment of relatively identification that needs is in the face database like this: M * N=MN bar is respectively f ₁, f ₂... f _{M * N}In addition, N bar line segment is arranged also on the facial image X to be identified, be respectively x ₁, x ₂... x _N

(A22) with any line segment x among the facial image X to be identified _kWith the M * N bar line segment f in the face database ₁, f ₂,,,,, f _{M * N}Compare processing respectively, wherein, natural number k is the line segment sequence number, and maximal value is N, and a such matching process will carry out the relatively calculating of M * N to line segment:＜x _k, f ₁,＜x _k, f ₂...,＜x _k, f _{M * N}, therefrom find out and this line segment x _kA most similar line segment f _kIf this line segment f _kThe j frame that belongs to facial image in the face database, then in this facial image X to be identified and the face database j frame facial image the match is successful once;

(A23) according to the comparative approach of step (A22), successively to the N bar line segment on the facial image X to be identified respectively with face database in MN bar line segment compare processing, promptly finish matching process N time;

(A24) if having certain frame facial image and the facial image X number of times that the match is successful to be identified maximum in the face database, then this frame facial image is exactly the people face the most similar to facial image X to be identified.

9, face identification method according to claim 1 is characterized in that: described step (5) further comprises following operation:

(51) user side is communicated by letter with management node, and informing does not have new recognition of face request in the management node video flowing, and stops to transmit video image to management node;

(52) management node result images and correlation parameter that each computing node is discerned a series of facial images gathers comprehensively, obtain with face database in the maximum facial image of matching times, and record matching times, find the solution this matching times again divided by the resulting ratio of a series of facial image sums, if this ratio, thinks then that the maximum facial image of this matching times is final face recognition result greater than the reference frame numerical value in the comprehensive recognition of face of described setting; Otherwise, think and can not discern;

(53) management node is communicated by letter with user side, informs the net result and the related data information of user side recognition of face;

(54) user side is shown to the user with the net result and the related data statistical information of recognition of face.

Images