CN109325536B - A kind of biomimetic pattern recognition method and its device - Google Patents
A kind of biomimetic pattern recognition method and its device Download PDFInfo
- Publication number
- CN109325536B CN109325536B CN201811119602.3A CN201811119602A CN109325536B CN 109325536 B CN109325536 B CN 109325536B CN 201811119602 A CN201811119602 A CN 201811119602A CN 109325536 B CN109325536 B CN 109325536B
- Authority
- CN
- China
- Prior art keywords
- sample
- classifier
- area
- detected
- rejection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000012567 pattern recognition method Methods 0.000 title claims abstract description 13
- 230000003592 biomimetic effect Effects 0.000 title claims abstract description 11
- 238000000034 method Methods 0.000 claims abstract description 52
- 239000011664 nicotinic acid Substances 0.000 claims abstract description 29
- 238000003909 pattern recognition Methods 0.000 claims abstract description 25
- 230000007246 mechanism Effects 0.000 claims abstract description 15
- 230000006870 function Effects 0.000 claims description 40
- 230000008569 process Effects 0.000 claims description 16
- 238000012549 training Methods 0.000 claims description 14
- 238000005457 optimization Methods 0.000 claims description 9
- 230000009977 dual effect Effects 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 3
- 230000001131 transforming effect Effects 0.000 claims description 2
- 238000004806 packaging method and process Methods 0.000 claims 1
- 238000012163 sequencing technique Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 2
- 244000287680 Garcinia dulcis Species 0.000 abstract 1
- 238000012706 support-vector machine Methods 0.000 description 16
- 238000010586 diagram Methods 0.000 description 6
- 230000009286 beneficial effect Effects 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000004590 computer program Methods 0.000 description 3
- 230000003247 decreasing effect Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- 241001465754 Metazoa Species 0.000 description 2
- NAXKFVIRJICPAO-LHNWDKRHSA-N [(1R,3S,4R,6R,7R,9S,10S,12R,13S,15S,16R,18S,19S,21S,22S,24S,25S,27S,28R,30R,31R,33S,34S,36R,37R,39R,40S,42R,44R,46S,48S,50R,52S,54S,56S)-46,48,50,52,54,56-hexakis(hydroxymethyl)-2,8,14,20,26,32,38,43,45,47,49,51,53,55-tetradecaoxa-5,11,17,23,29,35,41-heptathiapentadecacyclo[37.3.2.23,7.29,13.215,19.221,25.227,31.233,37.04,6.010,12.016,18.022,24.028,30.034,36.040,42]hexapentacontan-44-yl]methanol Chemical compound OC[C@H]1O[C@H]2O[C@H]3[C@H](CO)O[C@H](O[C@H]4[C@H](CO)O[C@H](O[C@@H]5[C@@H](CO)O[C@H](O[C@H]6[C@H](CO)O[C@H](O[C@H]7[C@H](CO)O[C@@H](O[C@H]8[C@H](CO)O[C@@H](O[C@@H]1[C@@H]1S[C@@H]21)[C@@H]1S[C@H]81)[C@H]1S[C@@H]71)[C@H]1S[C@H]61)[C@H]1S[C@@H]51)[C@H]1S[C@@H]41)[C@H]1S[C@H]31 NAXKFVIRJICPAO-LHNWDKRHSA-N 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 241000287196 Asthenes Species 0.000 description 1
- 235000008429 bread Nutrition 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 241000894007 species Species 0.000 description 1
- 208000011580 syndromic disease Diseases 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A kind of biomimetic pattern recognition method and its device, due to introducing the recognition mechanism of classifier when identifying sample to be detected, so that judging that sample to be detected is to belong to when the identification region of certain specimen types still falls within the rejection region of classifier to be easier to realize, to improve the efficiency of pattern-recognition;Simultaneously, classifier makes full use of the characteristics of " homologous " similar sample continuity, the connectivity of region, formed it is a kind of solve " homologous " similar sample areas tight surface methodology and determine a little whether the discriminant function method in the tight curved surface, conducive to accuracy of the sample to be detected during category identification is improved, the pattern recognition model that a kind of correct recognition rata is high, effect is good is provided for bionic pattern identification.
Description
Technical Field
The invention relates to the technical field of classification, in particular to a bionic mode identification method and a device thereof.
Background
The bionic pattern recognition is also called as topological pattern recognition and is a brand new pattern recognition theoretical system. From the perspective of recognizing things by human, the method provides a recognition concept of 'recognizing' rather than 'dividing' according to the general rule that 'homologous' similar sample points in a feature space are continuously and continuously distributed and communicated, and provides a brand-new development direction for the development of pattern recognition.
In the case where there are two things that are "same source" but not exactly equal in nature (including things, images, sounds, languages, states, etc.), and the difference between these two things is either gradual or non-quantized (as is the case for most things in the macroscopic world, even though it is impossible to refine below an apple like "apple yield", but "apple yield" by weight is considered to be gradual), there must be at least one gradual change between these two things, and the things in the middle of this gradual change are of the same type. That is, the set of all feature points of a "homologous" class constitutes a continuous, connected "region". The bionic pattern recognition is used as 'priori knowledge' of sample point distribution, so that the recognition capability of objects can be improved. After introducing the continuity rules of the same type of samples in the feature space in the bionic pattern recognition, the 'recognition' of a type of things is essentially the analysis and 'recognition' of the 'shape' of an infinite point set formed by the whole of the type of things in the feature space. Thus, unlike traditional pattern recognition, which targets the optimal partitioning of different types of samples in a feature space, biomimetic pattern recognition targets the optimal coverage of the distribution of a "homogeneous" type of sample in a feature space (i.e., the coverage area contains all the "homogeneous" sample regions, and the difference between the coverage area volume and the "homogeneous" sample region volume is less than a small constant) based on high-dimensional space geometry and point set topology. Solving the optimal coverage is a key technical method for realizing the bionic pattern recognition.
At present, the bionic pattern recognition system mostly adopts a simulation method to solve the optimal coverage method, for example, a finite number of hyper-spheres, hyper-ellipsoids or simplex models are used to simulate the optimal coverage. Obviously, in a high-dimensional space, the simulation gaps are huge, for example, a single-shaped area of a similar sample of the same source is covered by a hypersphere or a hypersphere of the same sample of the same source is covered by the hypersphere, the volume of the covering body is more than twice that of the hypersphere of the similar sample of the same source, and the correct recognition rate is lower than the expected rate. Therefore, solving the best coverage of the "homologous" homogeneous sample irregular area in the high dimensional space still does not solve the good international problem.
However, when designing classifiers for identifying serious diseases, identifying persons through biological features, identifying bills, identifying terrorists, and the like, an appropriate identification mechanism is often required to be introduced so that the classifiers either reject identification or have correct identification results when working. That is, it is required that (1) the rejection rate, which is expressed as the ratio of the number of samples rejected in the common test sample library to the total number of samples in the test sample library, is small; (2) the correct recognition rate is 100% or close to 100%, and the correct recognition rate is expressed as the proportion of the number of correctly classified samples in the test sample library after the rejected samples are removed to the total number of samples. If the rejection rate is large, the practical range and occasion of the classifier are limited. If the correct recognition rate cannot approach 100%, then one is discouraged from authenticating some particularly important thing or event directly with just the recognizer. Obviously, the improvement of the recognition rate of the bionic mode recognition enables the bionic mode recognition to have better application value.
At present, a Support Vector Machine (SVM) concept is adopted for pattern recognition, which mainly maps all feature vectors into a very high-dimensional space, a maximum interval hyperplane is established in the space, and an original space curved surface corresponding to the hyperplane is a classification decision surface. Two hyperplanes parallel to each other are built on both sides of the hyperplane separating the two types of feature vectors (data), and the hyperplane is separated to maximize the distance between the two parallel hyperplanes. Obviously, the same kind of feature region determined by the SVM is often an unbounded region, while the actual same kind of feature region is bounded. Therefore, the same-class feature region determined by the SVM encroaches on other actual feature regions of other classes or feature regions of unknown classes, and the encroachment is more serious, so that the risk of misclassifying samples is higher. Therefore, the method is not suitable for directly performing classifier design work such as serious disease authentication and identification, identity authentication and identification of people through biological characteristics, banknote authentication and identification, bill authentication and identification, terrorist authentication and identification and the like. And when training samples are added or new categories are added, the work of solving the SVM classification decision surface needs to be carried out again, so that the SVM does not have an incremental learning function. In the design of a multi-classification SVM classifier, a training sample is changed or a class is added, the learning and training process of the corresponding classifier needs to be started again, and the classifier cannot inherit any result of the previous training and learning, so that the SVM multi-classifier also has no incremental learning function. In consideration of the characteristics of unbalance of different types of feature areas and the like, in many improved SVM methods, for example, the concept of a hypersphere SVM classification algorithm is to map all feature vectors into a very high-dimensional space, establish a hypersphere with a minimum radius meeting a certain constraint in the space, and wrap almost all similar sample points with the hypersphere, so that an original space curved surface corresponding to the hypersphere or concentric hypersphere is a classification decision surface. In specific application, a plurality of embodiments show that the wrapping area of the classification decision surface of the hypersphere SVM encroaches on the feature area of an unknown class, namely the classification decision surface does not tightly wrap the actual feature area of a similar sample. In the SVM classifier described above, the following problems exist: (1) no suitable rejection mechanism was introduced: the correct recognition rate is not necessarily improved because it is inconvenient to determine the correct recognition rejection area or the recognition rejection area is determined reluctantly; (2) the classification decision-making bread wrapping area occupies a characteristic area of an unknown class, the unknown class is misjudged as a known class, and the correct recognition rate of a classifier cannot approach 100%; (3) the classifier has no incremental learning function: when the categories are increased or decreased, the learning and training work is thoroughly repeated, and when the training samples are increased or decreased, the learning and training work is basically thoroughly repeated.
Disclosure of Invention
The invention mainly solves the technical problem of how to overcome the defects of the prior art so as to improve the correct recognition rate of the bionic mode recognition. In order to solve the above technical problems, the present application provides a method and an apparatus for recognizing a bionic pattern.
According to a first aspect of the present application, there is provided a biomimetic pattern recognition method, comprising:
obtaining a sample to be detected;
performing type identification on the sample to be detected according to a classifier, wherein the classifier comprises at least one identification area of the sample type and a rejection area of the classifier, and the type of the sample to be detected is identified according to the identification area of the sample type and the rejection area of the classifier;
outputting the identification result of the sample to be detected
According to a second aspect of the present application, there is provided a biomimetic recognition apparatus comprising:
a memory for storing a program;
a processor for implementing the method as described in the first aspect by executing the program stored by the memory.
The beneficial effect of this application is:
according to the bionic pattern recognition method and the device thereof of the embodiment, as the recognition mechanism consisting of at least one sample type recognition area and the rejection area of the classifier is introduced when the sample to be detected is recognized, the judgment of whether the sample to be detected belongs to the recognition area of a certain type or the rejection area of the classifier is easy to realize, thereby improving the pattern recognition efficiency; meanwhile, the classifier fully utilizes the characteristics of continuity and area connectivity of the homologous samples to form a method for solving a compact package curved surface of the homologous sample area and a method for solving a discriminant function for judging whether a point is in the compact package curved surface, so that the accuracy of the sample to be detected in the process of identifying the type is improved, and a pattern identification model with high identification rate and good effect is provided for the bionic pattern identification.
Drawings
FIG. 1 is a flow chart of a biomimetic pattern recognition method;
FIG. 2 is a flow diagram of a classifier class identification process of an embodiment;
FIG. 3 is a flow diagram of a classifier class identification process of another embodiment;
FIG. 4 is a flow chart of a classifier design method;
FIG. 5 is a schematic diagram of an example of a classification decision surface of a hyper-sphere support vector machine not tightly wrapping homogeneous features;
FIG. 6 is a schematic diagram demonstrating the existence of tightly packed sets of homogeneous feature sets;
FIG. 7 is a schematic diagram of solving a similar feature area tight wrapping surface;
fig. 8 is a schematic structural diagram of a biomimetic pattern recognition apparatus.
Detailed Description
The present invention will be described in further detail with reference to the following detailed description and accompanying drawings. In the following description, numerous details are set forth in order to provide a better understanding of the present application. However, those skilled in the art will readily recognize that some of the features may be omitted in different instances or may be replaced by other methods. In some instances, certain operations related to the present application have not been shown or described in detail in order to avoid obscuring the core of the present application from excessive description, and it is not necessary for those skilled in the art to describe these operations in detail, so that they may be fully understood from the description in the specification and the general knowledge in the art.
Furthermore, the features, operations, or characteristics described in the specification may be combined in any suitable manner to form various embodiments. Also, the various steps or actions in the method descriptions may be transposed or transposed in order, as will be apparent to one of ordinary skill in the art. Thus, the various sequences in the specification and drawings are for the purpose of describing certain embodiments only and are not intended to imply a required sequence unless otherwise indicated where such sequence must be followed.
The invention of the application is characterized in that in order to overcome the problem of low correct recognition rate in the existing bionic pattern recognition process, a bionic pattern recognition model is reconstructed, a rejection algorithm of a classifier is adopted in the new bionic pattern recognition model, the region outside the rejection region of the classifier is considered to be the recognizable region of the sample to be detected, the sample to be detected is considered to be located in the recognizable region when the sample to be detected is judged not to be in the rejection region of the classifier, and then the sample type to which the sample to be detected belongs is found according to the distinguishing mode (note: sorting learning algorithm) of various sample types.
Referring to fig. 4, the present application discloses a bionic pattern recognition method, which includes steps S100-S300, which are described below.
And step S100, obtaining a sample to be detected. The sample to be tested here can be something with certain characteristics, such as a car, a building, a person, a bill, a banknote, etc.
Step S200, identifying the type of the sample to be detected according to a classifier, wherein the classifier comprises at least one identification area of the sample type (any sample type comprises an identification area and a rejection area, and the two areas are in a dual relationship) and the rejection area of the classifier (namely the intersection area of the rejection areas of all the sample types), and identifying the type of the sample to be detected according to the identification area of each sample type and the rejection area of the classifier. In one embodiment, see FIG. 5, the step S200 may include steps S210-S260.
Step S210, obtaining a sample to be detected for species identification.
In step S220, it is determined whether the sample to be detected is within the identification region of the sample type 1 (i.e., the region outside the rejection region of the sample type 1), and if the sample is within the identification region of the sample type 1 (i.e., the sample is not within the rejection region), the process proceeds to step S260, otherwise, the process proceeds to step S230.
In one embodiment, the feature of the sample to be tested is compared with the identification region of the sample type 1, and if the feature is within the identification region of the sample type 1, the sample to be tested is considered to be included in the identification region of the sample type 1.
If the number of sample types in the classifier is one, the sample to be detected is identified according to the identification area of the sample type, that is, the steps S230 to S240 are omitted.
It should be noted that, if the number of the sample types in the classifier is more than one, the sample types are sorted according to a preset priority, and the types of the samples to be detected are sequentially identified according to the identification areas of the sorted sample types; to facilitate the comparison process of the samples to be detected in this embodiment, the set classifier includes N (N > -2) sample types and corresponding identification regions, the sample types are sorted according to a preset priority, and preferably, the sample types are arranged according to priority numbers from 1 to N (it can be considered that the smaller the number is, the higher the priority is). Further, the priority here may be a user-defined level such as a size level of an automobile, an age level of a person, a denomination level of bills, and the like.
Step S230, comparing the samples to be detected with the identification areas or rejection areas of the sample types in sequence, and determining whether the samples are in the identification areas of a certain sample type.
Step S240, according to step S230, until it is determined whether the sample to be detected is in the identification area or the rejection area of the sample type N (i.e. the last sample type), if the sample to be detected is in the identification area (i.e. not in the rejection area of the sample type N), step S260 is performed, otherwise step S250 is performed.
Step S250, after the steps S220-S240, determining that the sample to be detected does not belong to the identification area of any sample type, and then determining that the sample to be detected belongs to the rejection area of the classifier, and at this time, regarding the sample to be detected as the rejection sample. The sample to be detected can also be considered as a sample to be recognized, so as to temporarily classify the sample which is not accurately recognized, and facilitate the later incremental learning of the system (the learning method will be specifically described below).
Step S260, determining that the sample to be detected belongs to the sample type corresponding to the identification area.
And step S300, outputting the identification result of the sample to be detected. In one embodiment, which sample class the sample to be detected belongs to is output, or a rejected sample or a sample to be recognized of the classifier is output.
In another embodiment, referring to fig. 3, in order to enhance the sample identification process of the classifier, a determination step S212 may be added between step S210 and step S220, where the determination step S212 includes: identifying the type of the sample to be detected according to the rejection area of the classifier, and entering step S220 when the sample to be detected is detected as being in the identification area of the classifier (i.e., not in the rejection area of the classifier), otherwise entering step S260 (i.e., determining that the sample to be detected belongs to the rejection sample or the sample to be detected of the classifier). In an embodiment, when the determination result of step S212 is no, it is determined in step S250 that the sample to be detected belongs to the rejected sample of the classifier, and when the determination result of step S240 is no, it is determined in step S250 that the sample to be detected belongs to the sample to be recognized of the classifier, so that the sample to be detected which is not accurately recognized can be clearly classified.
The bionic pattern recognition method disclosed by the application mainly introduces a recognition area of sample types and a rejection area of a classifier when a sample to be detected is recognized, the sample to be detected can be rapidly recognized according to recognition mechanisms of the recognition area and the rejection area, correct recognition rate is improved according to a discriminant function and a tight wrapping curved surface of the recognition area of various sample types constructed in the classifier, and the structure and the working principle of the classifier are specifically explained below.
To facilitate the description of the classifier design method related to the present embodiment, an example of not tightly wrapping homogeneous features is first given. Two areas are arranged on a two-dimensional plane, one area is a disc with the radius of 1, a coordinate system XOY is set up by taking the circle center as a coordinate origin, and the circle center is (0, 0); the other region is the inscribed regular triangle region of the disk, wherein one vertex of the triangle is positioned on the Y-axis (0,1) point. For the two regions, uniform sampling is performed, and the triangular region sampling points are denoted by "+". In a hypersphere Support Vector Machine (SVM) model, a kernel function is assumed to be a gaussian function, a penalty function coefficient is assumed to be 1000, a "+" point set hypersphere is trained, and a similar circle as in fig. 5 is obtained by mapping to an original space. Obviously, there is a large gap between the similar circle and the regular triangle area, i.e. the similar circle wraps the regular triangle area loosely, for example, the intervals [2/3,5/6] in the X-axis direction in fig. 4 and the like are all located in the gap formed by the similar circle and the regular triangle area, i.e. the area wrapped by the classification decision surface encroaches on the unknown characteristic-like area.
The above description only gives an example that the SVM classification decision surface does not tightly wrap the homogeneous features, but does not deny the existence of the tightly wrapped set of homogeneous feature sets, which is demonstrated below.
Let C be an N-dimensional feature space RNEpsilon is a small constant greater than zero and r is a constant greater than 1. If there are always different points X in C for any two points X, Y in C1,…,XhMake the distance delta (X, X)1),δ(X1,X2),…,δ(XhY), delta (X, Y) are not more thanThen it is called CAnd (4) the components are communicated compactly. Is provided withIs C plus all RNDistance from middle to C is not more thanThe area of the point of (2) whenIs a single connected region in the mathematical sense, then it is called CCompact single communication. In a pictorial sense, a single connection is an area without "holes". If the connection line using any two points in C as the end points is all inAnd any two points on these different connecting lines (segments) are also onIn, then, it is called CAnd (5) compacting the convex set. At this time, C is said to be inIs equal toIs called CAnd (5) compacting the boundary points. Let CεIs C plus all RNAnd (C) a point having a distance to C of no more than epsilon.
Let C be an N-dimensional feature space RNOne ofA dense set of single connected points, and C isAnd (5) compacting the convex set. Let C beThe number of dense boundary points is gamma, and X is (X) for any point1,…,xi,…,xN) E.g., C, defines 2N points (x)1±ε,x2,…,xi,…,xN),......, (x1,…,xi-1,xi±ε,xi+1,…,xN),......,(x1,…,xi,…,xN-1,xNε). Is provided withIs tightly wrapped asThen whenAt least Γ points, corresponding to the boundary points of each C, at least one of the 2N points is in i (C), and X ═ X1,…,xi,…,xN) E C toA distance between the inner and outer surfaces is greater than
The specific demonstration process is as follows:
let X be (X)1,…,xi,…,xN) Is one of CDense boundary point ofThe distance of the boundary isAs shown in fig. 6, point X ═ X1,…,xi,…,xN) Establishing a new coordinate system by using the epsilon C as an origin, and changing the X into (X)1,…,xi,…,xN) And (x)1,…,xi+ε,…,xN) Coordinate axes of (1) are denoted as XXiI is 1, …, N. Then with (x)1,…,xi-1,xi±ε,xi+1,…,xN) As a center, as a radiusBall B (x)1,…,xi-1, xi±ε,xi+1,…,xN) Then the ball B has a coordinate axisAnd (4) an intersection point is obtained. ByFor definition of (1), there must be a C midpoint (x)1',…,xi-1',(xi'±ε)',xi+1',…,xN') is contained in a ball B (x)1,…,xi-1,xi±ε,xi+1,…,xN) In (1). As being perpendicular to the coordinate plane XiXXj…, i, j ═ 1, …, N sphere B (x)1,…,xi-1,xi+ε,xi+1,…,xN) And B (x)1,…,xj-1,xj+ε,xj+1,…,xN) Tangent plane IIij…, when i, j is 1, …, N, piij… have coordinate axesAndtwo points of intersection.
From the hyperplane definition of the high dimensional space, when N points do not share the N-2 dimensional hyperplane,they must determine an N-1 dimensional hyperplane and share the hyperplane. Such as Has decidedAn iso-dotted N-1 dimensional hyperplane. Thus 2N dots Therein is provided withN point combination modes are adopted, wherein some combination modes comprise the following steps: the N points share an N-1 dimensional hyperplane and define the hyperplane (i.e., do not share an N-2 dimensional hyperplane). From the permutation and combination formula, however, the origin X ═ X (X)1,…,xi,…,xN) C and do not exceedThe number of the above N-1 dimensional hyperplanes isAnd (4) respectively. And it is easy to know that the solid enclosed by the hyperplane is a convex polyhedron which is 2N points The minimum volume convex polyhedron omega being the vertex. X ═ X1,…,xi,…,xN) E C is the center of the convex polyhedron and is an inner point.
Contrarily, assume 2N points (x)1±ε,x2,…,xi,…,xN),..., (x1,…,xi-1,xi±ε,xi+1,…,xN),...,(x1,…,xi,…,xN-1,xNε) is allIn (1).
Remember 2N points (x)1',…,xi-1',(xi±ε)',xi+1',…,xN'), i is 1, …, and N is the smallest convex polyhedron with a vertex is Ω'. Since C isCompact convex set, thusObviously, X ═ X1,…,xi,…,xN) E.g.. OMEGA', with (x)1',…,xi-1',(xi±ε)',xi+1',…,xN'), i 1, …, and N2NA hyperplane. Common vertex (x)1',…,xi-1',(xi+ε)',xi+1',…,xN') each hyperplane with XXiThe intersection of the axes is not X ═ X1,…,xi,…,xN) Andon the connecting line segment and on the outer extension of the line segment. Thus the sum of Ω' is (x)1',…,xi-1',(xi+ε)',xi+1',…,xN') is the vertex angle of the vertex and covers Completely similar can proveThus, there is a minimum convex set definition
As demonstrated belowWhen X is (X)1,…,xi,…,xN) Distance to omega surface is greater than
Can be verified The hyperplane (note: one of the boundary surfaces of Ω) equation of
From the point-to-plane distance equation, X ═ X1,…,xi,…,xN) The distance to the Ω surface is:
this is in conjunction withAnd X ═ X1,…,xi,…,xN) Is CThe tight boundary points contradict. After the syndrome is confirmed. Wherein,is the decreasing number of N, not greater than 5.8289.
Based on the above description, it is known that a compact package set of a similar feature set exists, and a design method of a classifier is described below, as shown in fig. 4, step a1 is a compact package set construction method of a similar feature set, where step a11 is an optimization process of a compactness parameter ∈:
provided with regions of similar characteristicsThe collected same-class feature point set is C, and the set isAnd (5) compacting the convex set. In the case where the homogeneous feature region T containing C is fixed, the smaller epsilon, the more C midpoints, and the more closely distributed C midpoints (i.e., the more samples taken from T). Given T is the convex set and given ε, we construct satisfyC of a dense convex set nature is not difficult. However, the set of points C is known and C is assumed to have somethingIt is difficult to obtain a minimum epsilon for dense convex set. Herein according to RNThe simplex determined by N +1 points on the same hyperplane is a theory of the smallest volume convex polyhedron containing the N +1 points, and an algorithm for estimating epsilon is given. If C has M points, use X1,X2,…,XMAnd (4) showing. The algorithm for estimating epsilon is:
first, calculate XjFirst neighbor of
Second, calculate XjSecond neighbor of (2)
......
Step N +1, calculating XjIs adjacent to the N +1 th
Step (N + 2), calculating XjMaximum distance to neighbor:
and (N + 3) calculating a suboptimal estimation of epsilon:
ε, i.e. satisfying that C is calculated according to the above algorithmAnd (5) compacting the convex set.
Step A12, constructionTightly packed collection
First, M points are constructedHypersphere neighborhood discriminant function:
here when fjWhen (X) is not less than 0, it can be judged that X is equal to XjIs used as the center of the device,radius hyper-sphere neighborhood Π (X)j) And (4) the following steps.
Second, from each Xj=(xj1,…,xji,…,xjN) And j is more than or equal to 1 and less than or equal to M, and 2N points (x) are derivedj1,…,xji±ε,…,xjN),1≤i≤N;
Third, all derived points (x) are detectedj1,…,xji±ε,…,xjN) I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to M is greater than or equal to pi (X)j) J is more than or equal to 1 and less than or equal to M, and the handle is not in any hypersphere neighborhood Π (X)j) And j is more than or equal to 1 and less than or equal to M are combined to obtainTightly packed set i (c).
Therefore, according to the above theorem, the number of midpoints in I (C) is greater than the number of boundary points in C.
The following further introduces a solving process of the similar feature region tightly wrapped curved surface based on the similar feature point set and the wrapped point set, namely step a2, which includes a step a21 of solving a classification decision function and a step a22 of solving a classification decision curved surface, as follows:
by transforming phi: RN→ H to map the eigenspace to a higher dimensional space, k: RN×RN→ R is the corresponding kernel function. FIG. 7 is a schematic view of an interface between two concentric superspheres, wherein the radius of the small supersphere is r, and the radius of the large concentric supersphere is rThe points inside the small hypersphere are the corresponding points of the C midpoint transformed to the high dimensional space, and the points outside the large hypersphere are the corresponding points of the I (C) midpoint transformed to the high dimensional space. Our goal is to find a suitable transformation phi RN→ H makes the small ultra-sphere contain almost all the points "+" and r is the smallest,maximum, i.e. p2And max. Thus, the original space curved surface corresponding to the small hypersphere in the high-dimensional space is the close wrapping curved surface of C.
For convenience of description, let C have m1Dots of which I (C) has m2=n-m1Point, c, is the sphere center of the high dimensional space. The following optimization model is established, and the similar characteristic region tightly-wrapped curved surface is constructed by solving an optimization solution.
ξ thereini,ξjIn order to be a function of the relaxation variable,is a penalty factor. To solve the model, a Lagrange function (Lagrange) was introduced:
the extremum points of the Lagrange function should satisfy:
thus, it is possible to provide
This gives a dual problem model:
with K (X)i,Xj) Substitution of phi (X)i)·φ(Xj) Known from the nature of the coreable function
The dual problem is a quadratic optimization problem and can be solved by a quadratic optimization problem solving method. After solving the above problem, to find r, ρ2And r2+ρ2Consider two sets:
let n be1=|S1|,n2=|S2Under the KKT condition
Wherein
Then, in step a21, the classification decision function is:
step A22, classify the decision surface as
The curved surface (the point on the curved surface can be solved by a numerical calculation method to solve an approximate solution) tightly wraps the same kind of characteristic areas. The distance from the feature region boundary to the classification decision boundary is less than epsilon.
Step A3, a rejection mechanism is set. Aiming at the problem that the correct rejection area is inconvenient to determine in the existing classification method, or even if the rejection area is reluctantly determined, the correct recognition rate is not necessarily improved, a suitable rejection mechanism suitable for a multi-class classifier is set based on the solution method for the similar characteristic area tightly wrapped curved surface.
For η type omega1,…,ωηThe classification problem, the parameter epsilon as small as possible is found by the method. Determining the discriminant function corresponding to each type of sample by the method described above
And sort decision tightly wrapped curved surface
When in usejWhen f (X) is not less than 0, it is judged that X is omegajAnd (4) class. In order to prevent the phenomenon that the same point belongs to different classes (i.e. different tightly wrapped areas have overlapping) from occurring, provision is made forjf (X) has priority, and the smaller the label, the higher the priority. When in useWhen X is ωjAnd (4) class. In addition, when there are a plurality of j pairs for the same point XjWhen f (X) is not less than 0, the class classification can also be carried out by a neighbor method. Otherwise, setting all common areas outside the tightly wrapped curved surface as rejection areas of the classifier, namely whenWhen so, the classifier rejects X.
Those skilled in the art will appreciate that the η -type ω is illustrated for equations (19), (20)1,…,ωηA classification problem is understood to be a classification problem of 1-N samples, then j in the formula is represented as [1, N ]]Number of sample types within the range. Substituting the characteristic set X of the sample to be detected into a formula (19) to obtain a function value corresponding to the sample to be detected, and judging that the characteristic set is shown in a formula (20) when the function value is more than or equal to zeroThe method comprises the following steps that an intentional classification decision is tightly wrapped in a curved surface, namely a sample to be detected is in an identification area of a sample type j; when the function value is less than zero, judging that the feature set is positioned outside the classification decision tight wrapping curved surface indicated by the formula (20), namely that the sample to be detected is in the rejection area of the sample type j; and substituting the feature set X into a formula (19) under any j condition to obtain function values corresponding to the sample to be detected when j is each numerical value, and if each obtained function value is smaller than zero, judging that the sample to be detected is positioned in a public area outside each classification decision compact wrapping curved surface indicated by a formula (20) under any j condition, namely is positioned outside the compact wrapping curved surface indicated by a formula (18), namely judging that the sample to be detected is included in the rejection area of the classifier.
In this embodiment, the identification area of any sample type can be represented by a classification decision-based tightly-wrapped curved surface as shown in formula (20), the area in the classification decision-based tightly-wrapped curved surface is used as the identification area of the current sample type, and the area outside the classification decision-based tightly-wrapped curved surface is used as the rejection area of the current sample type, where the identification area and the rejection area of the sample type are dual areas. Since the common region outside the compact package curved surface of the classification decision corresponding to each sample type does not include the identification region of any sample type, the common region can be represented by the region outside the compact package curved surface indicated by the formula (18), and taking the common region as the rejection region of the classifier can be beneficial to uniformly classifying the samples to be detected which cannot be identified, and is beneficial to forming a complete identification mechanism for classifying all the samples.
The existing classification method has the problems that when a new class sample is added or the class of the sample is reduced, the learning training work is completely repeated, namely when a training sample is added or reduced, even if a large number of samples are previously learned and trained, all samples are subjected to learning training again. In another embodiment of the invention, based on the above process method, an incremental learning method is introduced, and the specific steps are as follows:
(1) solving a tight package set of a sample set, wherein the sample set comprises an added new sample type set, a set of types corresponding to misclassification and rejection samples when training samples are added, and/or a subtracted set of types corresponding to misclassification samples;
(2) based on the corresponding classes of the sample set and the obtained compact wrapping set, solving a classification decision function and a classification decision compact wrapping curved surface of the sample set;
(3) and adjusting the rejection area according to the rejection mechanism of the classifier.
The incremental learning step is described below in three cases:
when adding new classes, the work results of all previous training studies can be retained. Only the added new category sample set is required to be solved for the parcel set, and the classification decision function and the classification decision compact wrapping curved surface of the new category are solved based on the new category sample set and the parcel set, and meanwhile, the rejection area is correspondingly adjusted according to the rejection mechanism of the classifier.
When the training samples are added and the original classifier can classify correctly, the classifier does not need to make any adjustment. Otherwise, only the package set (note: most of the previous calculation results can be used) is solved again for the class set corresponding to the misclassification and rejection samples (note: the new boundary point can be judged), the classification decision function and the classification decision compact package curved surface of the new class are solved based on the new class sample set and the package set, and meanwhile, the rejection area is correspondingly adjusted according to the rejection mechanism of the classifier.
When the misclassification sample is subtracted, only the class set corresponding to the misclassification sample (note: the boundary point can be judged) is solved again for the parcel set (note: most of the calculation results can be used in the past) and the classification decision function and the classification decision compact parcel curved surface of the new class are solved based on the new class sample set and the parcel set, and meanwhile, the rejection area is correspondingly adjusted according to the rejection mechanism of the classifier.
In an embodiment, the incremental learning function is added to the above-described disclosed bionic pattern recognition method, so that when sample types are added or deleted, particularly when the sample types corresponding to the samples in the sample to be recognized in the embodiment of the present application are added, the related operations (such as recognition processing, tight wrapping curve solving, classification function solving) of other sample types do not change, which is beneficial to dynamically adjusting the sample types of the classifier for bionic pattern recognition, and is beneficial to enhancing the adaptability of the bionic pattern recognition in different application occasions.
Referring to fig. 8, correspondingly, the present application also discloses a bionic pattern recognition apparatus, which includes a sample acquiring unit 41, a recognition unit 43 and an output unit 44, which will be described separately below.
The sample acquiring unit 41 is configured to acquire a sample to be detected, and in an embodiment, an image having specific information, such as a banknote, a bill, a human face, a certificate, an animal, a bionic model, and the like, may be used as the sample to be detected, and the sample to be detected may be acquired in a sequential or continuous manner.
The identification unit 43 is configured to identify the type of the sample to be detected according to a classifier, where the classifier includes identification areas of multiple sample types and rejection areas of the classifier, and the type of the sample to be detected is identified according to the identification area of each sample type and the rejection area of the classifier. In one embodiment, the recognition unit 43 comprises 1-N sample classes, each sample class having its own recognition area and rejection area, which constitute the primary database of the classifier.
The output unit 44 is configured to output the identification result of the sample to be detected, including outputting which sample category the sample to be detected belongs to, or outputting a rejected sample or a sample to be detected belongs to. In a specific embodiment, if a sample to be detected is listed as a rejected sample or a sample to be detected, the system records the characteristics of the sample to be detected, so that the system can improve the own basic database through an incremental learning method conveniently, and classify the recorded characteristics into the attributes of a new sample type. It should be noted that the specific form of the recognition result may be a text message or a prompt symbol that is convenient for the user to know, and is not limited herein.
In an embodiment, the above-mentioned bionic pattern recognition device is applied to a specific recognition occasion, for example, bionic pattern recognition experiments are performed on animal and vehicle models with various shapes, and experiments show that when 8800 × 2 times of recognition is performed in an all-round way, the correct recognition rate reaches 99.85%, and the rejection rate is only about 0.2%, so that the feasibility and reliability of the bionic pattern recognition method and the bionic pattern recognition device thereof are proved, and the application value is also shown to be high.
Those skilled in the art will appreciate that all or part of the functions of the various methods in the above embodiments may be implemented by hardware, or may be implemented by computer programs. When all or part of the functions of the above embodiments are implemented by a computer program, the program may be stored in a computer-readable storage medium, and the storage medium may include: a read only memory, a random access memory, a magnetic disk, an optical disk, a hard disk, etc., and the program is executed by a computer to realize the above functions. For example, the program may be stored in a memory of the device, and when the program in the memory is executed by the processor, all or part of the functions described above may be implemented. In addition, when all or part of the functions in the above embodiments are implemented by a computer program, the program may be stored in a storage medium such as a server, another computer, a magnetic disk, an optical disk, a flash disk, or a removable hard disk, and may be downloaded or copied to a memory of a local device, or may be version-updated in a system of the local device, and when the program in the memory is executed by a processor, all or part of the functions in the above embodiments may be implemented.
The present invention has been described in terms of specific examples, which are provided to aid understanding of the invention and are not intended to be limiting. For a person skilled in the art to which the invention pertains, several simple deductions, modifications or substitutions may be made according to the idea of the invention.
Claims (6)
1. A bionic pattern recognition method is characterized by comprising the following steps:
acquiring a sample to be detected, wherein the sample to be detected is image information of a real object to be detected;
performing type identification on the sample to be detected according to a classifier, wherein the classifier comprises at least one identification area of the sample type and a rejection area of the classifier, and the type of the sample to be detected is identified according to the identification area of the sample type and the rejection area of the classifier;
outputting an identification result of the sample to be detected, wherein the identification result comprises a sample type to which the sample to be detected belongs, or a rejected sample or a sample to be identified;
the acquisition process of the identification area of each sample type of the classifier and the rejection area of the classifier comprises the following steps: constructing a compact package set of the same kind of feature set, which comprises an optimized compactness parameter, and constructing the compact package set; solving a similar characteristic region tightly wrapped curved surface, which comprises solving a classification decision function and a classification decision curved surface; setting a rejection mechanism of the classifier, wherein the rejection mechanism comprises a discriminant function and a classification decision tight wrapping curved surface which are corresponding to each sample type, setting an area in the classification decision tight wrapping curved surface which is corresponding to each sample type as an identification area of the sample type, setting an area outside the classification decision tight wrapping curved surface which is corresponding to each sample type as a rejection area of the sample type, setting a public area outside the classification decision tight wrapping curved surface as a rejection area of the classifier, and setting the rejection area of the classifier as an intersection area of the rejection areas of the sample types;
collecting a similar characteristic point set C from a similar characteristic area, wherein the C has M points and uses X1,X2,…,XMAnd (3) expressing that the optimization method of the compactness parameter comprises the following steps:
calculating XjFirst neighbor X ofj1:
Calculating XjSecond neighbor X ofj2:
......
Calculating XjN +1 neighbor X ofjN+1:
Calculating XjMaximum distance to neighbor:
computing a suboptimal estimate of ε:wherein epsilon is a compactness parameter, and N is a characteristic space dimension;
the construction steps of the compact packaging set are as follows:
constructing M pointsHypersphere neighborhood discriminant function: 1≤j≤M;
from each Xj=(xj1,…,xji,…,xjN) And j is more than or equal to 1 and less than or equal to M, and 2N points (x) are derivedj1,…,xji±ε,…,xjN),1≤i≤N;
Detecting all derived points (x)j1,…,xji±ε,…,xjN) I is more than or equal to 1 and less than or equal to N, j is more than or equal to 1 and less than or equal to M is greater than or equal to pi (X)j) J is more than or equal to 1 and less than or equal to M, and the handle is not in any hypersphere neighborhood Π (X)j) And j is more than or equal to 1 and less than or equal to M are combined to obtain the derivative point setTightly packed set i (c);
the solving process of the similar characteristic region tightly wrapped curved surface comprises the following steps: transforming the feature space to a higher dimension and establishing an optimization model; introducing a Lagrange function to solve the optimization model; solving the dual problem by a quadratic optimization problem solving method; and solving the classification decision function and the classification decision curved surface.
2. The biomimetic pattern recognition method according to claim 1, wherein the recognizing the type of the sample to be detected according to the recognition area of the type of the sample and the rejection area of the classifier comprises:
if the number of the sample types in the classifier is one, identifying the sample to be detected according to the identification area of the sample type;
if the number of the sample types in the classifier is more than one, sequencing all the sample types according to a preset priority, and sequentially identifying the types of the samples to be detected according to the identification areas of the sequenced sample types;
and for each sample type, detecting whether the sample to be detected belongs to the identification area of the sample type, and determining that the sample to be detected belongs to the sample type when the sample to be detected is detected to be in the identification area of the sample type.
3. The biomimetic pattern recognition method according to claim 1 or 2, wherein the recognizing the type of the sample to be detected according to the recognition area of the type of the sample and the rejection area of the classifier further comprises:
and identifying the type of the sample to be detected according to the rejection area of the classifier, and determining that the sample to be detected belongs to the rejection sample or the sample to be detected of the classifier when the sample to be detected is detected to be in the rejection area of the classifier.
4. The biomimetic pattern recognition method of claim 3, wherein the acquisition process of the recognition area for each sample class of the classifier further comprises an incremental learning step, the incremental learning step comprising:
solving a tight package set of a sample set, wherein the sample set comprises a set of added new sample types, a set of types corresponding to misclassification and rejection samples when training samples are added, and/or a set of types corresponding to subtracted misclassification samples;
solving a classification decision function and a classification decision compact wrapping surface of the sample set;
and adjusting the rejection area according to the rejection mechanism of the classifier.
5. A biomimetic pattern recognition apparatus, comprising:
a memory for storing a program;
a processor for implementing the method of any one of claims 1-4 by executing a program stored by the memory.
6. A computer-readable storage medium, comprising a program executable by a processor to implement the method of any one of claims 1-4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811119602.3A CN109325536B (en) | 2018-09-25 | 2018-09-25 | A kind of biomimetic pattern recognition method and its device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811119602.3A CN109325536B (en) | 2018-09-25 | 2018-09-25 | A kind of biomimetic pattern recognition method and its device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109325536A CN109325536A (en) | 2019-02-12 |
CN109325536B true CN109325536B (en) | 2019-09-17 |
Family
ID=65265836
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811119602.3A Active CN109325536B (en) | 2018-09-25 | 2018-09-25 | A kind of biomimetic pattern recognition method and its device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109325536B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110084790B (en) * | 2019-04-17 | 2020-11-17 | 电子科技大学成都学院 | Bionic mode identification improvement method for judging pneumonitis in imaging |
CN110598578B (en) * | 2019-08-23 | 2024-06-28 | 腾讯云计算(北京)有限责任公司 | Identity recognition method, training method, device and equipment of identity recognition system |
CN112446486A (en) * | 2019-08-30 | 2021-03-05 | 中国科学院半导体研究所 | Construction method of high-order neuron model |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
CN102608589A (en) * | 2012-03-16 | 2012-07-25 | 南京理工大学 | Radar target identification method on basis of biomimetic pattern identification theory |
CN103400149A (en) * | 2013-08-02 | 2013-11-20 | 武汉大学 | Biomimetic pattern recognition method based on kernel function extension |
CN103559542A (en) * | 2013-10-31 | 2014-02-05 | 华北水利水电大学 | Extension neural network pattern recognition method based on priori knowledge |
CN105069485A (en) * | 2015-08-26 | 2015-11-18 | 中国科学院深圳先进技术研究院 | Extreme-learning-machine-based mode identification method in tensor mode |
-
2018
- 2018-09-25 CN CN201811119602.3A patent/CN109325536B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
CN102608589A (en) * | 2012-03-16 | 2012-07-25 | 南京理工大学 | Radar target identification method on basis of biomimetic pattern identification theory |
CN103400149A (en) * | 2013-08-02 | 2013-11-20 | 武汉大学 | Biomimetic pattern recognition method based on kernel function extension |
CN103559542A (en) * | 2013-10-31 | 2014-02-05 | 华北水利水电大学 | Extension neural network pattern recognition method based on priori knowledge |
CN105069485A (en) * | 2015-08-26 | 2015-11-18 | 中国科学院深圳先进技术研究院 | Extreme-learning-machine-based mode identification method in tensor mode |
Non-Patent Citations (5)
Title |
---|
Geometrical learning, descriptive geometry, and biomimetic pattern recognition;Wang Shoujue;《ELSEVIER》;20050613 |
Two-dimensional discriminant locality preserving projections (2DDLPP) and its application to feature extraction via fuzzy set;Minghua Wan;《Springer》;20151110 |
仿生模式识别(拓扑模式识别)一种模式识别;王守觉;《电子学报》;20021031;第30卷(第10期) |
仿生模式识别技术研究与应用进展;陈阳;《智能***学报》;20160228;第11卷(第1期) |
基于仿生模式识别与传统模式识别的人脸识别效果比较研究;王守觉;《电子学报》;20040731;第32卷(第7期) |
Also Published As
Publication number | Publication date |
---|---|
CN109325536A (en) | 2019-02-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2019015246A1 (en) | Image feature acquisition | |
Bouguila et al. | A countably infinite mixture model for clustering and feature selection | |
CN109325536B (en) | A kind of biomimetic pattern recognition method and its device | |
Zhao et al. | Fingerprint image synthesis based on statistical feature models | |
Bouadjenek et al. | Robust soft-biometrics prediction from off-line handwriting analysis | |
CN106951825A (en) | A kind of quality of human face image assessment system and implementation method | |
Wenjing et al. | Research on areal feature matching algorithm based on spatial similarity | |
CN111339935B (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
Bouadjenek et al. | Histogram of Oriented Gradients for writer's gender, handedness and age prediction | |
Khoury et al. | Adversarial training with voronoi constraints | |
Tian et al. | Learning the distribution preserving semantic subspace for clustering | |
CN104751463A (en) | Three-dimensional model optimal visual angle selection method based on sketch outline features | |
CN108776809A (en) | A kind of dual sampling Ensemble classifier model based on Fisher cores | |
CN111126617B (en) | Method, device and equipment for selecting fusion model weight parameters | |
Mallet et al. | Deepfake detection analyzing hybrid dataset utilizing cnn and svm | |
CN108596245A (en) | It is a kind of that the complete face identification method for differentiating sub-space learning is cooperateed with based on multiple view | |
Liu et al. | Remote sensing image vehicle detection based on pre-training and random-initialized fusion network | |
CN105160358B (en) | A kind of image classification method and system | |
CN111401440A (en) | Target classification recognition method and device, computer equipment and storage medium | |
Lemhadri | A neural network with feature sparsity | |
Shnitzer et al. | Log-euclidean signatures for intrinsic distances between unaligned datasets | |
Zhang et al. | Wasserstein generative recurrent adversarial networks for image generating | |
Yadav et al. | Cross-covariance based affinity for graphs | |
Lethikim et al. | Fuzzy cluster analysis for interval data based on the overlap distance | |
Ramathilaga et al. | Two novel fuzzy clustering methods for solving data clustering problems |
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 |