CN106778801A - Method and apparatus for analyzing the classification of object of observation - Google Patents
Method and apparatus for analyzing the classification of object of observation Download PDFInfo
- Publication number
- CN106778801A CN106778801A CN201611006639.6A CN201611006639A CN106778801A CN 106778801 A CN106778801 A CN 106778801A CN 201611006639 A CN201611006639 A CN 201611006639A CN 106778801 A CN106778801 A CN 106778801A
- Authority
- CN
- China
- Prior art keywords
- observations
- classification
- characteristic information
- training
- information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Exemplary embodiment of the invention is related to the method and apparatus of the classification for analyzing object of observation.It is a kind of to include for analyzing the class method for distinguishing of object of observation：Obtain the characteristic information of one or more object of observations；According to the characteristic information of one or more of object of observations, the classification prediction to one or more of object of observations is implemented using the classification policy based on neutral net；And the result that the classification is predicted is based at least partially on, determine the respective classes of one or more of object of observations.
Description
Technical field
Exemplary embodiment of the invention is usually related to data analysis and process technology, and is more particularly to used to analyze
The method and apparatus of the classification of object of observation.
Background technology
Grading analysis numerous industries and field (for example including car industry, chemical industry, mining industry, forestry, agricultural, nutrition industry,
Pharmacy industry, energy industry, comprehensive industry etc.) in all have suitable importance.By taking car industry as an example, the group from automobile can be directed to
The particle fallen on part or parts is counted, it is also possible to measure the size of these particles.For example, it is possible to use aobvious
Micro mirror (such as scanning electron microscope) obtains the information of quantity and size about these particles.Such as the size of fruit granule
It is excessive, then it is considered that the poor quality of the component or parts.Furthermore it is also possible to the size according to particle is come to these
Grain is classified.But, the size according only to particle cannot track the source of particle, also not know that particle is fallen from what component
Get off.Therefore it provides a kind of can be Worth Expecting to the effective scheme that object of observation as such as particle is classified
's.
The content of the invention
It is according to an exemplary embodiment of the present invention in a first aspect, a kind of class method for distinguishing for analyzing object of observation can be with
Including：Obtain the characteristic information of one or more object of observations；According to the characteristic information of one or more of object of observations, profit
Implement the classification prediction to one or more of object of observations with the classification policy based on neutral net；And at least partly
Result of the ground based on the classification prediction, determines the respective classes of one or more of object of observations.
Second aspect according to an exemplary embodiment of the present invention, a kind of device for analyzing the classification of object of observation can be with
Including：Data obtaining module, the characteristic information for obtaining one or more object of observations；Classification prediction module, for basis
The characteristic information of one or more of object of observations, implemented using the classification policy based on neutral net to one or
The classification prediction of multiple object of observations；And category determination module, the result for being based at least partially on the classification prediction,
Determine the respective classes of one or more of object of observations.
The third aspect according to an exemplary embodiment of the present invention, a kind of device for analyzing the classification of object of observation can be with
Including：At least one processor and at least one memory for storing computer program code.At least one memory
With the computer program code can be configured as promoting described device at least to implement together with least one processor with
Lower operation：Obtain the characteristic information of one or more object of observations；According to the characteristic information of one or more of object of observations,
Implement the classification prediction to one or more of object of observations using the classification policy based on neutral net；And at least portion
Divide result of the ground based on the classification prediction, determine the respective classes of one or more of object of observations.
Foregoing any aspect according to an exemplary embodiment of the present invention, the characteristic information of one or more of object of observations
The spectral information of one or more of object of observations can be included.For example, the spectral information can be normalized
Spectral information.In the exemplary embodiment, the spectral information can be energy dispersive spectrometry (Energy Dispersive
Spectrometer, EDS) spectral information.
Foregoing any aspect according to an exemplary embodiment of the present invention, the feature of one or more object of observations of acquisition
Information can include：The positional information of one or more of object of observations is collected, wherein, the positional information indicates described
The position of one or more object of observations；And the spy of one or more of object of observations is extracted according to the positional information
Reference ceases.
Foregoing any aspect according to an exemplary embodiment of the present invention, the respective classes of one or more of object of observations
Can indicate it is following at least one：The source of one or more of object of observations, the material of one or more of object of observations
The size of matter, one or more of object of observations, and one or more of object of observations part.
Foregoing any aspect according to an exemplary embodiment of the present invention, the classification policy based on neutral net can be wrapped
Include the algorithm based on SVMs (Support Vector Machine, SVM).
Foregoing any aspect according to an exemplary embodiment of the present invention, the classification policy based on neutral net can lead to
Cross the mapping relations that network training process comes between defined feature information and object type.
Foregoing any aspect according to an exemplary embodiment of the present invention, the network training process can include：Obtain one
It is individual or it is multiple training object characteristic information, wherein, it is one or more of training objects respective classes be known；At least
The characteristic information and respective classes of one or more of training objects are based in part on, calculating is associated with the mapping relations
Training network one or more parameters；And the training network with one or more parameters for being calculated is true
It is set to the neutral net, the classification results for predicting one or more of object of observations.
Foregoing any aspect according to an exemplary embodiment of the present invention, can be based at least partially on the minimum optimization of sequence
(Sequential Minimum Optimization, SMO) principle calculates one or more of parameters.
For example, can be implemented by microscopical at least a portion (such as microscopical one or more parts)
The method of first aspect according to an exemplary embodiment of the present invention.Similarly, according to an exemplary embodiment of the present invention second and/
Or the device of the third aspect can be microscopical at least a portion or can be communicatively coupled to the microscope.It is described
Microscope can be charged particle microscope or electron microscope (such as scanning electron microscope).
By the method and apparatus using being provided according to exemplary embodiment of the present, can be with convenient, efficiently and accurate
True mode realizes the automatic classification to the object of observation of particle etc..
Brief description of the drawings
In order to illustrate more clearly of the technical scheme of exemplary embodiment of the present, below by the attached of exemplary embodiment
Figure is briefly described.It should be evident that drawings in the following description be merely exemplary with it is illustrative, and do not mean that to this
Invention carries out any limitation.For those of ordinary skill in the art, other accompanying drawings can also be obtained according to these accompanying drawings.When
When being read in conjunction with the figure, by referring to the following detailed description to illustrative embodiment, the present invention is better understood with exemplary
The various aspects of embodiment and its further objects and advantages, in the accompanying drawings：
Fig. 1 is according to the flow present embodiment illustrates the class method for distinguishing for analyzing object of observation
Figure；
Fig. 2 is according to present embodiment illustrates neural network training process；
Fig. 3 is according to present embodiment illustrates neural network prediction process；
Fig. 4 shows the signal of the device of the classification for analyzing object of observation according to one example embodiment of the present invention
Figure；And
Fig. 5 shows showing for the device of the classification for analyzing object of observation according to another exemplary embodiment of the invention
It is intended to.
Specific embodiment
To make the purpose, technical scheme and advantage of the embodiment of the present invention clearer, come to retouch in detail below with reference to accompanying drawings
State embodiments of the invention.Obviously, described embodiment is only a part of embodiment of the invention, rather than whole implementation
Example.
Through this specification, refer to feature, advantage or similar wording be not meant as can using the present invention and
The all features realized should be in any single embodiment of the invention with advantage.Conversely, it is understood that be related to
The wording of feature and advantage mean in conjunction with the embodiments described specific features, advantage or characteristic be included in it is of the invention at least
In one embodiment.Thus, through this specification, discussion and similar wording to feature and advantage can refer to same
Embodiment, but not necessarily refer to same embodiment.Additionally, described feature of the invention, advantage and characteristic can be used appointing
What suitable mode merges in one or more embodiments.Those skilled in the relevant art will recognize that not have
The present invention is put into practice in the case of one or more specific features or advantage of specific embodiment.In other examples, Ke Yi
Additional feature and advantage are realized in some embodiments, it is not necessarily come across among all embodiments of the invention.
Exemplary embodiment of the invention, term " sample ", " sample ", " determinand " and similar terms are interchangeably
The object applicable for representing the technical scheme proposed according to exemplary embodiment of the present.Further, when entering in the text
During row description, various information, data, image, message or other communications can be sent or be sent to from a part or device
Another part or device.It should be appreciated that transmission information, data, image, message or other communications not only can include to the letter
Breath, data, image, message or other communication transmission, but also can include to described information, data, image, message or its
The preparation that it communicates.Therefore, the use to any such term is all not considered as limiting the spirit and model of the embodiment of the present invention
Enclose.
The composition of particle can be known by grading analysis, particle is classified, review the source of particle, and thus
Judge the quality and performance of associated components or product.For example, it is possible to use light microscope technique and electron microscope skill
Art carries out grading analysis.
Using light microscope situation in, by comparing light field, polarisation under particle image difference, and combine
The shape of grain, can be divided into metallic particles, non-metallic particle and fiber by particle.Additionally, according to the size of particle, can be by not
The gradation of same type.However, broadly can only be divided three classes particle by the program, and grade according to size, but cannot
Know the composition of particle, and then the source of particle cannot be reviewed.
In the situation using electron microscope, by the image or photo of electron microscope, the position of particle can be obtained
Put, the information such as size, recycle the spectral information of particle, particle can be classified and graded.If the principle of classification is
Source based on particle, then the source of particle can also be followed the trail of.However, the program needs to set up the data for granules
Storehouse.Setting up the process of the database needs to combine the experience of particle material composition in itself and researcher, and due to material sheet
The complexity of body, it is difficult to set up a good database to classify particle.
For example, it is a kind of to use energy dispersive spectrometry (Energy Dispersive Spectrometer, EDS) information
Scheme to realize granules particle is classified and is tracked the source of particle according to chemical composition.It is every obtaining
In the case of the EDS results of individual particle, it is possible to use they are divided into different types by EDS data storehouse.Specifically, can make
The element composition of each particle is obtained with EDS.For example, as it is known that particle is made up of 90% iron (Fe) and 10% sulphur (S).
Because there are different materials different elements to constitute, these particles can be divided into by different materials based on EDS information, and then
The source of particle can be tracked.
As can be seen here, the sorting technique based on EDS information needs to set the element model of each particle in various grain types
Enclose, and specific classifying rules just depends on the setting to the scope.The conventional method for setting EDS data storehouse is based on specific unit
The scope of plain percentage composition.If for example, to classify FeMnCr particles, needing manually to carry out database as follows
Set：{ iron (Fe)：90-98%；Manganese (Mn)：1-2%；Chromium (Cr)：0.1-2% }.
However, the feasibility of this particle sorting method and the degree of accuracy are limited.On the one hand, the number of particle and its type
Amount is always excessive, thus is difficult to effective manual sort.On the other hand, the EDS results changes of every kind of grain type are more.
Due to this limitation, it is difficult to the chemical composition for finding a kind of clear and definite rule or standard to utilize particle is classified to particle.
In fact, as fruit granule comprises only specific element, such as FeMnCr, then for each element rational ratio of setting is feasible
's.However, when the quantity of grain type increases, this just becomes infeasible, because all types of particles can not be taken into account.
In addition, it is contemplated that the component difference very little between some materials, its particle may contain a large amount of similar elements, such as CuZn16,
CuZn33, CuZn16Si, this can make granules process situation become even worse.The particle of even identical material, due to material
Natural inhomogeneities, the percentage composition between the main component of these particles there is also difference, in addition some particles composition
Content error is more than 20%.As can be seen here, the complexity due to particle in itself, it is artificial to be difficult to the particle compared with multiple types or classification
Set up an accurate criteria classification database.
A kind of exemplary embodiment of the invention, there is provided scheme for analyzing the classification of object of observation, it leads to
Cross using the classification policy based on neutral net realize to the determinands such as particle, particulate, sample or object of observation oneself
Dynamic classification.Without manually participating in, user is also without knowing object of observation specific material in itself for the implementation of this classification schemes
Or element is constituted, thus the complexity of operation is also reduced while providing compared with high-class accuracy.
Fig. 1 is according to the flow present embodiment illustrates the class method for distinguishing for analyzing object of observation
Figure.The method shown in Fig. 1 can be realized by various devices.The device can be any kind of electricity
Sub- device, including but not limited to data processing equipment, data analyzer, computing device, visualizer, microscope, power spectrum
Instrument, measuring instrument, personal digital assistant, digital recorder, monitor, sensor, multimedia equipment and/or console etc..
Method according to Fig. 1, in a step 102, can obtain the characteristic information of one or more object of observations.Lift
For example, the characteristic information of one or more of object of observations can include that the power spectrum of one or more of object of observations is believed
Breath.Alternatively, or in addition, the characteristic information can also be including the source with object of observation, classification, composition and/or material etc.
Associated any other information.Using various appropriate devices (such as scanning electron microscope and energy disperse spectroscopy),
By scanning one or more object of observations, the gamma-spectrometric data or information of one or more of object of observations can be obtained.Root
According to exemplary embodiment of the invention, one or more of object of observations can include one or more particles, for example, come from
The particle of identical or different material/objects.Needs analyze its corresponding classification by observing the particle.It is appreciated that removing
Outside particle, method as shown in Figure 1 applies also for any other determinand for needing to analyze or determine classification.
Exemplary embodiment of the invention, 102 can the step of obtain the characteristic information of one or more object of observations
To specifically include：The positional information of one or more of object of observations is collected, wherein, the positional information indicates described one
The position of individual or multiple object of observations；And the feature of one or more of object of observations is extracted according to the positional information
Information.
For example, it is possible to use electron microscope scans object of observation (such as testing sample particle) and obtains corresponding
Image.By appropriate image processing method, positional information (the such as absolute coordinate and/or phase of object of observation can be obtained
To coordinate), and thereby determine that the position of object of observation.By using energy disperse spectroscopy, can obtain being located at the object of observation of diverse location
Initial characteristic data (such as gamma-spectrometric data).Alternatively, or in addition, resulting initial characteristic data can be fitted
When treatment, such as normalized or any other treatment carried out according to default rule.Optionally, can be by original
The category analysis process that beginning characteristic or the characteristic through processing are used to subsequently carry out as characteristic information.
Exemplary embodiment of the invention, can need not collect the positional information of one or more object of observations
In the case of, from other devices or equipment (such as energy disperse spectroscopy, sensor of characteristic information that is acquired or storing object of observation
And/or database) directly or indirectly obtain one or more of sights (such as by reception, retrieval and/or reading etc.)
Survey the characteristic information of object.It is alternatively possible to be used in combination the characteristic information for obtaining in this way and according to positional information
The characteristic information of extraction.
At step 104, according to the characteristic information of one or more of object of observations, it is possible to use based on neutral net
Classification policy implement to the classification of one or more of object of observations prediction.Exemplary embodiment of the invention,
The classification policy based on neutral net can be by network training process come between defined feature information and object type
Mapping relations.In other words, defined mapping relations can be indicated or shown between the characteristic information of object of observation and classification
Relevance, and can to a certain extent embody the classifying rules for object of observation.
For example, the network training process can include：The characteristic information of one or more training objects is obtained, its
In, the respective classes of one or more of training objects are known；It is based at least partially on one or more of training
The characteristic information and respective classes of object, one or more parameters for the training network that calculating is associated with the mapping relations；
And the training network with one or more parameters for being calculated is defined as the neutral net, for predicting
State the classification results of one or more object of observations.It is one or more of training objects can include with it is one or more of
Particle of the object of observation from identical or different article/sample.
In practice, can simultaneously or gathered according to sequencing training object and object of observation characteristic information.
In other words, when the classification of object of observation is analyzed, it is possible to use the training network for building in advance and storing；Or do not having
In the case of available training network, but the standard sample with some Known Species, it is also possible to utilize these standard samples
Build training network temporarily.For example, it is possible to use electron microscope trains object (such as standard sample to scan
Grain) and obtain corresponding image.By appropriate image processing method, can obtain training the positional information of object (such as exhausted
To coordinate and/or relative coordinate), and thereby determine that the position of training object.By using energy disperse spectroscopy, can obtain being located at different positions
The initial characteristic data (such as gamma-spectrometric data) of the training object put, and optionally by initial characteristic data or through treatment
Characteristic is used for network training process as the characteristic information of training object.
As can be seen here, the training process of neutral net make use of the characteristic information of the training object of known class or species with
Relevance between object type, and by building the modes such as training pattern or training network, it is right that such relevance is used for
The classification prediction of the object of observation of unknown classification or species.Especially, in order to improve the accuracy that classification is predicted, training process can
With including the parameter configuration and the optimization process of value to training network.For example, SMO principles can be based at least partially on
To calculate one or more parameters of training network.Further, since different sorting algorithms can have different parameter settings, because
This, can select suitable sorting algorithm, to reflect different classification gauges according to different application scenarios and test object
Then.
Exemplary embodiment of the invention, the classification policy based on neutral net can be included based on SVM
Algorithm.According to the algorithm based on SVM, one group of training object is given, each training object is marked as belonging to a classification, passes through
SVM training process, can construct the SVM models or neutral net that can be classified to object of observation.For example, SVM models
Or each object can be mapped to neutral net the multiple points in space, so that different classes of object is by as wide as possible
Intervallum demarcate come.Then, the same space is mapped to by by object of observation, can be fallen at interval based on object of observation
The which side of band predicts the classification of the object of observation.
For example, SVM neutral nets can have two parameters, and the coordinate and classification of be input into object are represented respectively,
And the neutral net can automatically for two different classes of objects find suitable boundary to separate each other.In this way, working as
When being input into the coordinate of object of observation, SVM neutral nets can export the classification of object of observation.The coordinate of object of observation can be with sight
The feature for surveying object is associated, and if the feature of object of observation is related with multiple parameters, then the coordinate of object of observation can be by
Multi-C vector represents that correspondingly, SVM neutral nets can realize the division of object type using many dimensional planes or hyperplane.
When SVM neutral nets are trained, it is only necessary to provide a small amount of training object for every kind of object type.Through instruction
Experienced SVM neutral nets can automatically be classified (for example, it is only necessary to be input into the characteristic information of object of observation) to object of observation,
And without artificially finding classifying rules.Additionally, SVM neutral nets can also provide classification accuracy higher, for example, classify
Precision can reach more than 90%.It is appreciated that classification policy according to an exemplary embodiment of the present invention can also include other bases
In the classifying rules or algorithm of neutral net, or SVM algorithm and other classifying rules or algorithm can be combined to come in fact
Existing classification policy according to an exemplary embodiment of the present invention.
In step 106, it is based at least partially on the result of the classification prediction, it may be determined that one or more of sights
Survey the respective classes of object.For example, the classification obtained by implementation steps 104 can be directly utilized to predict the outcome to refer to
Show the classification of object of observation, such as classification predicts the outcome can include the unique identifier for indicating object type.Alternatively or
Additionally, can consider that the classification obtained by implementation steps 104 is predicted the outcome and various contextual informations (are for example grasped
Make environment, design conditions and/or data precision etc.), to determine the respective classes of object of observation.One or more of observations
The respective classes of object can indicate it is following at least one：It is the source of one or more of object of observations, one or many
The size of the material of individual object of observation, one or more of object of observations, and one or more of object of observations group
Into part.
It is according to an exemplary embodiment of the present invention to be conducive to user lighter for analyzing the scheme of the classification of object of observation
Set up characteristic information (such as EDS information) database of various samples or particle, and constituent content model need not be manually set
Enclose, and can be by automatic classification and determine the classification of particle, and then follow the trail of its source.According to an exemplary embodiment of the present
Scheme involved by network training process be also easy to realize, only lead to too small amount of standard sample data (composition of such as particle
Element and content), suitable neutral net just can be trained to classify particle automatically.The program can avoid setting
Complicated but less accurate classifying rules, so as to reduce implementation complexity and save the operating time.
Can be implemented according to this hair by microscopical at least a portion (such as microscopical one or more parts)
The method of bright exemplary embodiment.For example, method according to an exemplary embodiment of the present invention as shown in Figure 1 can pass through
Electron microscope analyzes the characteristic information (such as gamma-spectrometric data) of particle, and nerve is trained by the gamma-spectrometric data of standard particle
Network, and utilize housebroken neutral net, gamma-spectrometric data according to candidate particles is automatically classified to particle, from
And obtain accurate granules result.Described in detail below in conjunction with Fig. 2 and Fig. 3 according to an exemplary embodiment of the present invention
Neural network training process and neural network prediction process.
Fig. 2 is according to present embodiment illustrates neural network training process.As shown in Fig. 2 in step 202
Middle acquisition training set data.For example, using scanning electron microscope and energy disperse spectroscopy, by the standard for scanning m types
Sample, each type of t (such as 60) particle of standard sample scanning, can obtain the m*t gamma-spectrometric data of particle.Table 1
Exemplarily give i-th the one of particle group of gamma-spectrometric data.
| Cu | Zn | Si | Pb | Fe | S | Al | P | Mg | Br | Cr | Mn | Sn |
| 0 | 0 | 3.944 | 0 | 95.247 | 0 | 0 | 0 | 0 | 0 | 0 | 0.809 | 0 |
Table 1
In table 1, the corresponding element term of symbology of the first row, the digitized representation of the second row element is in particle i
In degree.Due to scanning is standard sample particle, therefore the corresponding grain type of every group of gamma-spectrometric data is known
's.Can be using the gamma-spectrometric data of these standard sample particles and its corresponding grain type as training set data.
In step 204, the pretreatment to training set data is implemented.For example, can return to all of gamma-spectrometric data
One change is processed.Assuming that including P kinds essential element (in this example, P=13 as shown in table 1) in training set data altogether.Here
Essential element can be determined according to default rule, it may for instance be considered that degree>More than 0.1% be only is main
Element.It is original by the element for finding all samples particle in training set data for pth kind element (p=1,2 ..., P)
The maximum x of contentp,maxWith minimum value xp,min, it is possible to use formula (1) realizes the normalized of constituent content.
Wherein, xi,p,orignalRepresent the original amount of the pth kind element in particle i, xi,p,normRepresent the pth in particle i
Plant the normalization content of element, and i=1,2 ..., m*t.For particle i, can be by vector xiTo represent that it owns
The content of essential element, and vector xiDimension P depend on all particles essential element number (such as 13).
In step 206, training set is divided.For example, being q by normalized m*t groups gamma-spectrometric data random division
Individual (such as 3) training set, and each training set containsGroup gamma-spectrometric data and its corresponding grain type.Especially,
Still contain m kind grain types in each divided training set, therefore, the gamma-spectrometric data of each training set still being capable of body
Now all related grain types.
In a step 208, it is that training network selects appropriate kernel function.Linearly can not in view of current gamma-spectrometric data
Point, therefore kernel function can be utilized, the data of linearly inseparable are mapped to higher-dimension so as to linear separability.For example, can be with
Using the Gaussian function as shown in formula (2) as kernel function.
Wherein, kernel function k (z1,z2) it is as two vector z of |input paramete1And z2Space after implicit mapping
In inner product function representation, σ is the control parameter of kernel function.Can be by normalized gamma-spectrometric data (such as vector xi) as core
The |input paramete of function, the inner product operation in mapping space is simplified by kernel function, direct in higher dimensional space so as to avoid
Computing.In other words, it is believed that kernel function is the expression-form of the inner product of supporting vector.
In step 210, the relevant parameter of kernel function is selected.It can be seen from the description above in conjunction with Fig. 1, svm classifier
Process is related to find Optimal Separating Hyperplane f=wTxi+ b, the Optimal Separating Hyperplane is bigger from the interval of data point, then certainty factor is higher, because
And need to find geometry intervalMaximum hyperplane.By mathematic(al) manipulation, svm classifier algorithm can be changed into by public affairs
Formula (3) is solved to realize the classification to input data.
s.t.yi(wTxi+b)≥1-εi, i=1...n (3)
Wherein, εiIt is the slack variable introduced for the noise in processing data, its control data or data point deviate
Influence of the degree of hyperplane to classifying；I is specimen number；N is total sample number；C is weight factor；Vectorial w and constant b be with
The related parameter (for example, in two-dimensional space, w is slope, b is intercept) of Optimal Separating Hyperplane, wTThe transposition of vector w is represented, | | w
| | represent norm；Vector xiIt is the given data (such as i-th gamma-spectrometric data of particle) of object to be sorted；And yiIt is classification knot
Really.So, for SVM neutral nets, it is thus necessary to determine that two variables：Kernel function variable σ and slack variable εi。
In the step 212, higher-dimension Lagrangian formulation is introduced, to determine the parameter of training network.For example, it is right
In each (σ, εi) combination, formula (3) can be converted to by formula (4) by Lagrangian decomposition.
Wherein, L (w, b, αi) it is Lagrangian formulation；αiIt is the coefficient of expression formula；γiIt is geometry interval, it represents number
According to a distance from hyperplane.So, the training process of SVM neutral nets just can be understood as bringing training data into Lagrange
Expression formula solves factor alphaiProcess.For example, by calculatingCan be by formula
(4) object function as shown in formula (5) is converted to.
Wherein,<xi,xj>Represent vector xiAnd xjInner product；I and j represent that the different specimen numbers in same training set are (all
Such as i-th particle and j-th particle).
In step 214, SMO solution procedurees are carried out.For example, by way of iteration, a pair of α are choseniAnd αj(example
Such as according to didactic selection mode), then it is fixed except αiAnd αjOutside other parameters, determine the α under w extremum conditionsi, so
α afterwardsjBy αiRepresent.If meeting institute's Prescribed Properties, terminate SMO solution procedurees, otherwise, update αi, proceed iteration meter
Calculate.
In the step 216, in order to reach the more preferable accuracy of training network, can obtain optimal by cross validation
(σ, εi) combination.For example, into Lagrangian formulation can be brought the training data of 3 training sets.For each (σ, εi)
Combination, after training data is brought into, can obtain one group of αi, and thereby determine that corresponding SVM training networks.Therefore, for 3
1st, 2 groups of data are merged into training set by the cross validation of training set first, are calculated αi, then using the 3rd group of data as testing
Card collection, brings into the αiThe training network of value, is predicted the outcome, (i.e. known with actual result by that will predict the outcome
Classification results) it is compared, first accuracy rate can be obtained.Keep (σ, εi) constant, then respectively by the 2nd, 3 groups of data, the
1st, used as training set, remaining one group collects 3 groups of data as checking, obtains two other accuracy rate.These three can be chosen accurate
The average value of rate is used as the group (σ, εi) accuracy rate.
Different (σ, ε can be attemptedi) combination (for example randomly select (σ, εi) combination), by repeat step 210-216,
(σ, ε until obtaining optimal (i.e. predictablity rate highest)i) combination, or until obtaining predictablity rate more than appointed threshold
(such as accuracy rate>90%) one group of (σ, εi), as final optimal (σ, the ε for determiningi) combination.
In step 218, all training set datas are brought into Lagrangian formulation, so that it is determined that optimized training net
The parameter of network.For example, according to optimal (σ, the ε for determining in the step 216i) combination, and in step 218 by all instructions
Practice collection data (such as 60 particle specimens of every kind of grain type) and all bring Lagrangian formulation into, then in step 220
In, by SMO methods, α can be calculatediCoefficient, and be saved as the parameter of training network in step 222.It is real
On border, in SVM algorithm, the corresponding α of only so-called " supporting vector "iIt is just meaningful, most αiEqual to 0.
Using training process and corresponding step as shown in Figure 2, it may be determined that for classifying to object of observation
Neutral net.It is appreciated that housebroken neutral net and its relevant parameter can be saved to predetermined storage location,
And it is alternatively possible to it is updated accordingly as needed (for example, updating one or more related to neutral net
Parameter, sorting criterion and/or algorithm) so that classification prediction is more convenient and accurate.
Fig. 3 is according to present embodiment illustrates neural network prediction process.Using process as shown in Figure 2
The neutral net (such as SVM neutral nets) that training is obtained, can analyze and determine the respective classes of object of observation.In order to obtain
The information relevant with object of observation (such as particle), in step 302, scans whole filter membrane image.For example, by sweep type electricity
Sub- microscope, can obtain the whole image of filter membrane, and the image contains the relevant information of the particle on filter membrane and filter membrane.
In step 304, particle position information is extracted.By appropriate image processing method, such as according to gray value and
Grad, can automatically extract out the profile of particle on filter membrane, so as to obtain position of the particle on image.For example, exist
During scanning filter membrane image, the positional information (such as with image-related coordinate and scaling) of objective table can be recorded, so
Can just image be combined and obtain coordinate of each particle on objective table.
Within step 306, particle spectral information is obtained.According to the particle position information for obtaining in step 304, by particle
The focal point of electron gun of electron microscope is moved to, using energy disperse spectroscopy, the spectral information of particle can be obtained (for example, in particle
Ratio shared by every kind of element).
In step 308, spectral information is input into neutral net (such as SVM neutral nets).For example, it is input into SVM
The spectral information of neutral net can be the element ratio through the particle after normalized.Here the method for normalizing that uses with
The method for normalizing used during training neutral net is identical.
In the step 310, neural network prediction is implemented.For the neutral net for training, the classification prediction that it is used
The relevant parameter of algorithm has determined, thus can enter data to export corresponding classification prediction knot according to suitable
Really.For example, in SVM neutral nets, it is possible to use such asIt is such to be related to what classification was predicted
Expression formula.Due to factor alpha being determined in network training processiWith kernel function k (xi, x) in variable σ, and b is known
Parameter, xiAnd yiIt is training set data, therefore, bring the gamma-spectrometric data x of candidate particles into the expression formulas and calculated, and tie
Predetermined classification decision criteria (for example, using the hyperplane in svm classifier algorithm) is closed, can obtain dividing the particle
Class predicts the outcome.
By repeat step 304-310, can be pre- to all particles corresponding neural network classification of implementation for scanning one by one
Survey.In step 312, output predicts the outcome.According to predicting the outcome for all particles, the classification of all particles can be obtained
Data.
With reference to Fig. 1-Fig. 3 description as can be seen that the scheme that is proposed according to an exemplary embodiment of the present by using
Classification policy (such as SVM neural network algorithms) based on neutral net, can simplify the taxonomy database for object of observation
The process of foundation, and the classification accuracy of (being greater than 90%) higher can be reached.
Exemplary embodiment of the invention, during training set data and/or test set data are prepared, can be with
Sample is put into electron microscope to scan successively.Alternatively, by appropriate image processing method, sample can be divided into position
In the particle of different zones.The spectral information of variable grain can be obtained by energy disperse spectroscopy.For example, for every kind of sample,
About 500-1000 particle can be scanned.In the data obtained from each sample/region, such as 50 can be randomly choosed
, used as training set, the gamma-spectrometric data of remaining particle is used as test set for the gamma-spectrometric data of grain.So, for training set and test set,
Corresponding gamma-spectrometric data and its corresponding correct granules can be obtained.The corresponding correct particle of training set can be utilized
Classify to train neutral net, and housebroken nerve net can be verified using the corresponding correct granules of test set
The classification accuracy of network.
It is normalized by by the gamma-spectrometric data of training set and test set, it is possible to use training set data is trained
Neutral net (such as process to train SVM networks) according to Fig. 2.Table 2 exemplarily gives two types (sample 1
With sample 2) sample particle gamma-spectrometric data (without numerical value represent content be 0), wherein each type include 5 sample particles.
| Cu | Zn | Si | Pb | Fe | S | Al | P | Mg | Br | Cr | Mn | Sn | Sample |
| 3.944 | 95.247 | 0.809 | 1 | ||||||||||
| 3.878 | 95.249 | 0.873 | 1 | ||||||||||
| 3.724 | 94.537 | 0.413 | 0.446 | 0.88 | 1 | ||||||||
| 3.287 | 95.612 | 1.101 | 1 | ||||||||||
| 3.64 | 95.419 | 0.94 | 1 | ||||||||||
| 78.476 | 17.321 | 2.808 | 1.125 | 0.271 | 2 | ||||||||
| 78.783 | 17.646 | 2.424 | 0.86 | 0.286 | 2 | ||||||||
| 79.33 | 17.44 | 2.83 | 0.286 | 0.1 | 2 | ||||||||
| 79.362 | 17.458 | 2.573 | 0.36 | 0.248 | 2 | ||||||||
| 79.366 | 17.547 | 2.813 | 0.274 | 2 |
Table 2
According to the neural network training method (such as SVM network training methods) described by exemplary embodiment of the present,
Gaussian function can be selected as kernel function, neutral net is trained using training set data, obtained by way of cross validation
To optimal network parameter (such as kernel function variable σ and slack variable εi), and preserve the neutral net for completing training.Will test
The neutral net that collection data input is trained, can automatically obtain corresponding classification and predict the outcome.
In this example, for the particle of test set, because its corresponding correct classification is also known, therefore, will just
True classification is compared with the classification via neural network prediction, can obtain the accurate of the classification prediction based on neutral net
Property.Table 3 exemplarily gives the assessment result predicted for neural network classification, wherein being listed accordingly for different materials
Essential element, the number of samples (such as total number of particles) of every group of testing of materials collection, number of errors, main error classification (for example
Erroneous judgement classification) and classification accuracy.
Table 3
The assessment result shown from table 3 can be seen that proposed according to an exemplary embodiment of the present based on nerve
The accuracy rate major part of the classification policy of network has reached at a relatively high accuracy all 95% or so.Additionally, realizing basis
During the classification schemes of exemplary embodiment of the present, participate in or interfere without artificial, without knowing object of observation or treat test sample
Product material in itself is constituted, and just automatically the object of observation can be entered according to the spectral information of object of observation (such as particle)
Row classification.This not only avoids the foundation to cumbersome and complicated classifying rules and database, but also with the side of high efficiency and time conservation
Formula provides good classification performance.
Fig. 4 shows the device 400 of the classification for analyzing object of observation according to one example embodiment of the present invention
Schematic diagram.For example, independent analytical equipment can be disposed or be designed to device 400, or device 400 is disposed or collected
Into to equipment such as microscope and/or energy disperse spectroscopies, to realize according to the scheme of exemplary embodiment of the present.In exemplary implementation
In example, device 400 can be microscopical at least a portion or be communicatively coupled to microscope.As shown in figure 4, device
400 can include data obtaining module 401, classification prediction module 402 and category determination module 403.Example of the invention
Property embodiment, data obtaining module 401 can be used to obtain the characteristic information of one or more object of observations.Classification prediction module
402 can be used for the characteristic information according to one or more of object of observations, using the classification policy based on neutral net come real
Apply the classification prediction to one or more of object of observations.Category determination module 403 can be used to being based at least partially on described
The result of classification prediction, determines the respective classes of one or more of object of observations.
It is appreciated that the module of the device 400 shown in Fig. 4 can be added, deletes, replaces, merges and/or be torn open
Point, to realize as combined method and step and/or function shown in Fig. 1-3.For example, data obtaining module 401 can be only
The positional information of one or more of object of observations is collected on the spot or by means of one or more submodules, wherein, institute
State the position that positional information indicates one or more of object of observations；And extract described one according to the positional information
The characteristic information of individual or multiple object of observations.Classification prediction module 402 can be independently or by means of one or more submodules
Block, by network training process come the mapping relations between defined feature information and object type.Alternatively, or in addition, classification
Determining module 403 can be right using identified one or more observations independently or by means of one or more submodules
The respective classes of elephant come indicate it is following at least one：The source of one or more of object of observations, one or more of sights
Survey object material, and one or more of object of observations part.
Fig. 5 shows the device 500 of the classification for analyzing object of observation according to another exemplary embodiment of the invention
Schematic diagram.For example, independent analytical equipment can be disposed or be designed to device 500, or device 500 is disposed or
The equipment such as microscope and/or energy disperse spectroscopy are integrated into, to realize according to the scheme of exemplary embodiment of the present.In exemplary reality
Apply in example, device 500 can be microscopical at least a portion or be communicatively coupled to microscope.As shown in figure 5, device
500 can include at least one processor 501, and at least one memory 503 including computer program code 502.Institute
State at least one memory 503 and the computer program code 502 can be configured to together with least one processor 501
So that device 500 performs the method and step and/or function combined described by Fig. 1-3.For example, processor 501 can be via
Communicated with memory 503 for the bus of the transmission information between the component of device 500.Memory 503 for example can be with
Including volatibility and/or nonvolatile memory.Memory 503 can be configured as storage information, data, content, using, refer to
Order etc., for enabling the exemplary embodiment of the invention of device 500 to perform various functions.
Processor 501 can be embodied by various ways.For example, processor can be embodied as at various hardware
Reason component in one or more, the hardware handles component is, for example, coprocessor, microprocessor, controller, data signal
Processor (DSP), the treatment element with or without the DSP that encloses, or various other process circuits, described other treatment electricity
Road includes such as application specific integrated circuit (ASIC), field programmable gate array (FPGA), micro controller unit (MCU), hardware
Such integrated circuit such as accelerator or special-purpose computer chip.In this way, in certain embodiments, the processor can include
One or more are configured as the processing core for independently performing.Polycaryon processor can realize many places in single physical encapsulation
Reason.Additionally or alternatively, the processor can including one or more be configured as via bus connect processor, with
Just independently executing for instruction, streamline and/or multithreading is realized.
In the exemplary embodiment, processor 501 can be configured as performing storage in memory 503 or otherwise
The addressable instruction for the processor 501.Alternatively, or in addition, the processor can be configured as performing hard coded
Function.No matter in this way, being configured by hardware or software approach or by its combination, the processor can be represented
Entity (such as physically body of operation according to an exemplary embodiment of the present invention can be implemented when corresponding configuration has been carried out
Now in circuit).Thus, for example, when the processor is embodied as ASIC or FPGA etc., the processor can be special
Door is configured to guide the hardware of the operation being described herein.Alternatively, as another example, when the processor is embodied as
For software instruction actuator when, processor can be specially configured to implement when executed at this by the instruction
Method and/or operation described in text.The processor can especially include being configured as supporting the operation of the processor
Clock, ALU (ALU) and gate.
It will be understood that the combinations of blocks in each square frame and flow chart of flow chart can be by various device (examples
Such as hardware, firmware, processor, circuit and/or with performing include that the software of one or more computer program instructions is associated
Miscellaneous equipment) realize.For example, said one or multiple processes can be embodied by computer program instructions.Thus, body
The computer program instructions for having showed said process can be entered by the memory 503 of the device 500 using exemplary embodiment of the present
Row storage, and performed by the processor 501 of device 500.As it would be appreciated, any such computer program instructions
To be loaded into computer or other programmable devices (for example, hardware) to produce machine, so that resulting computer
Or other programmable devices realize specified function in flowchart block.These computer program instructions can also be stored
In computer-readable memory, its bootable computer or other programmable devices are operated in a specific way, so that storage
Instruction in computer-readable memory produces product, performs the product and realizes specified function in flowchart block.
Computer program instructions can also be loaded on computer or other programmable devices so that sequence of operations is described
Implement on computer or other programmable devices, so as to produce computer implemented process so that computer or other
The instruction performed on programmable device provides the operation for function specified in flowchart block.
Therefore, the square frame of flow chart is supported to the combination of the device for implementing to specify function and to for implementing to specify
The combination of the operation of function.It is further appreciated that the combination of square frame can be with one or more square frames and flow chart of flow chart
Realized by the combination of the computer system or specialized hardware and computer instruction based on specialized hardware implementing to specify function.
In some exemplary embodiments, some above-mentioned operations can be changed or further enhanced.Additionally, in some exemplary implementations
In example, can include it is additional can selection operation.The modification of aforesaid operations, addition or enhancing can be in any sequence and according to appointing
What combines to implement.
The those skilled in the art in the invention for benefiting from the teaching presented in described above and associated drawings will
Expect many modifications of the invention set forth herein and other embodiments.It will thus be appreciated that the present invention be not limited to it is disclosed
Specific embodiment, and be intended to by it is described modification and other embodiments include within the scope of the appended claims.
Claims (27)
1. a kind of class method for distinguishing for analyzing object of observation, it includes：
Obtain the characteristic information of one or more object of observations；
According to the characteristic information of one or more of object of observations, implemented using the classification policy based on neutral net to institute
State the classification prediction of one or more object of observations；And
The result of the classification prediction is based at least partially on, the respective classes of one or more of object of observations are determined.
2. method according to claim 1, wherein, the classification policy based on neutral net includes being based on supporting vector
The algorithm of machine.
3. method according to claim 1 and 2, wherein, the characteristic information of one or more of object of observations includes institute
State the spectral information of one or more object of observations.
4. method according to claim 1 and 2, wherein, the respective classes of one or more of object of observations are indicated
Below at least one：The source of one or more of object of observations, the material of one or more of object of observations, described one
The size of individual or multiple object of observations, and one or more of object of observations part.
5. method according to claim 1 and 2, wherein, the characteristic information bag for obtaining one or more object of observations
Include：
The positional information of one or more of object of observations is collected, wherein, the positional information indicates one or many
The position of individual object of observation；And
The characteristic information of one or more of object of observations is extracted according to the positional information.
6. method according to claim 1 and 2, wherein, the classification policy based on neutral net passes through network training
The mapping relations that process is come between defined feature information and object type.
7. method according to claim 6, wherein, the network training process includes：
The characteristic information of one or more training objects is obtained, wherein, the respective classes of one or more of training objects are
It is known；
The characteristic information and respective classes of one or more of training objects are based at least partially on, are calculated and is closed with the mapping
One or more parameters of the associated training network of system；And
The training network with one or more parameters for being calculated is defined as the neutral net, for predicting
State the classification results of one or more object of observations.
8. method according to claim 7, wherein, the sequence minimum principle of optimality is based at least partially on to calculate described one
Individual or multiple parameters.
9. method according to claim 1 and 2, wherein, methods described is implemented by microscopical at least a portion.
10. a kind of device for analyzing the classification of object of observation, it includes：
Data obtaining module, the characteristic information for obtaining one or more object of observations；
Classification prediction module, for the characteristic information according to one or more of object of observations, using based on neutral net
Classification policy is predicted come the classification implemented to one or more of object of observations；And
Category determination module, the result for being based at least partially on the classification prediction, determines one or more of observations
The respective classes of object.
11. devices according to claim 10, wherein, the classification policy based on neutral net include based on support to
The algorithm of amount machine.
12. device according to claim 10 or 11, wherein, the characteristic information of one or more of object of observations includes
The spectral information of one or more of object of observations.
13. device according to claim 10 or 11, wherein, the respective classes of one or more of object of observations are indicated
It is following at least one：It is the source of one or more of object of observations, the material of one or more of object of observations, described
The size of one or more object of observations, and one or more of object of observations part.
14. device according to claim 10 or 11, wherein, the characteristic information for obtaining one or more object of observations
Including：
The positional information of one or more of object of observations is collected, wherein, the positional information indicates one or many
The position of individual object of observation；And
The characteristic information of one or more of object of observations is extracted according to the positional information.
15. device according to claim 10 or 11, wherein, the classification policy based on neutral net is instructed by network
Practice the mapping relations that process is come between defined feature information and object type.
16. devices according to claim 15, wherein, the network training process includes：
The characteristic information of one or more training objects is obtained, wherein, the respective classes of one or more of training objects are
It is known；
The characteristic information and respective classes of one or more of training objects are based at least partially on, are calculated and is closed with the mapping
One or more parameters of the associated training network of system；And
The training network with one or more parameters for being calculated is defined as the neutral net, for predicting
State the classification results of one or more object of observations.
17. devices according to claim 16, wherein, it is based at least partially on the sequence minimum principle of optimality described to calculate
One or more parameters.
18. device according to claim 10 or 11, wherein, described device is microscopical at least a portion or is leading to
The microscope is coupled on letter.
A kind of 19. devices for analyzing the classification of object of observation, it includes at least one processor and stores computer journey
At least one memory of sequence code, at least one memory and the computer program code be configured as with it is described extremely
A few processor promotes described device at least to implement following operation together：
Obtain the characteristic information of one or more object of observations；
According to the characteristic information of one or more of object of observations, implemented using the classification policy based on neutral net to institute
State the classification prediction of one or more object of observations；And
The result of the classification prediction is based at least partially on, the respective classes of one or more of object of observations are determined.
20. devices according to claim 19, wherein, the classification policy based on neutral net include based on support to
The algorithm of amount machine.
21. device according to claim 19 or 20, wherein, the characteristic information of one or more of object of observations includes
The spectral information of one or more of object of observations.
22. device according to claim 19 or 20, wherein, the respective classes of one or more of object of observations are indicated
It is following at least one：It is the source of one or more of object of observations, the material of one or more of object of observations, described
The size of one or more object of observations, and one or more of object of observations part.
23. device according to claim 19 or 20, wherein, the characteristic information for obtaining one or more object of observations
Including：
The positional information of one or more of object of observations is collected, wherein, the positional information indicates one or many
The position of individual object of observation；And
The characteristic information of one or more of object of observations is extracted according to the positional information.
24. device according to claim 19 or 20, wherein, the classification policy based on neutral net is instructed by network
Practice the mapping relations that process is come between defined feature information and object type.
25. devices according to claim 24, wherein, the network training process includes：
The characteristic information of one or more training objects is obtained, wherein, the respective classes of one or more of training objects are
It is known；
The characteristic information and respective classes of one or more of training objects are based at least partially on, are calculated and is closed with the mapping
One or more parameters of the associated training network of system；And
The training network with one or more parameters for being calculated is defined as the neutral net, for predicting
State the classification results of one or more object of observations.
26. devices according to claim 25, wherein, it is based at least partially on the sequence minimum principle of optimality described to calculate
One or more parameters.
27. device according to claim 19 or 20, wherein, described device is microscopical at least a portion or is leading to
The microscope is coupled on letter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611006639.6A CN106778801A (en) | 2016-11-16 | 2016-11-16 | Method and apparatus for analyzing the classification of object of observation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611006639.6A CN106778801A (en) | 2016-11-16 | 2016-11-16 | Method and apparatus for analyzing the classification of object of observation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106778801A true CN106778801A (en) | 2017-05-31 |
Family
ID=58969377
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611006639.6A Pending CN106778801A (en) | 2016-11-16 | 2016-11-16 | Method and apparatus for analyzing the classification of object of observation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106778801A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117818A (en) * | 2018-08-30 | 2019-01-01 | 苏州富莱智能科技有限公司 | Material structure characteristic intelligent recognition analysis system and analysis method |
| CN109145846A (en) * | 2018-08-30 | 2019-01-04 | 苏州富莱智能科技有限公司 | Material microstructure intelligent recognition analysis system and analysis method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011008262A2 (en) * | 2009-07-13 | 2011-01-20 | H. Lee Moffitt Cancer Center & Research Institute | Methods and apparatus for diagnosis and/or prognosis of cancer |
| CN105718937A (en) * | 2014-12-03 | 2016-06-29 | 财团法人资讯工业策进会 | Multi-class object classification method and system |
-
2016
- 2016-11-16 CN CN201611006639.6A patent/CN106778801A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011008262A2 (en) * | 2009-07-13 | 2011-01-20 | H. Lee Moffitt Cancer Center & Research Institute | Methods and apparatus for diagnosis and/or prognosis of cancer |
| CN105718937A (en) * | 2014-12-03 | 2016-06-29 | 财团法人资讯工业策进会 | Multi-class object classification method and system |
Non-Patent Citations (2)
| Title |
|---|
| 王岚等: "《金相实验技术 第2版》", 31 May 2010 * |
| 范君: "磨粒图谱识别系统研究", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117818A (en) * | 2018-08-30 | 2019-01-01 | 苏州富莱智能科技有限公司 | Material structure characteristic intelligent recognition analysis system and analysis method |
| CN109145846A (en) * | 2018-08-30 | 2019-01-04 | 苏州富莱智能科技有限公司 | Material microstructure intelligent recognition analysis system and analysis method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110018524B (en) | X-ray security inspection contraband identification method based on vision-attribute | |
| Rocchini et al. | Remotely sensed spectral heterogeneity as a proxy of species diversity: recent advances and open challenges | |
| CN104732240B (en) | A kind of Hyperspectral imaging band selection method using neural network sensitivity analysis | |
| Matthews et al. | The gambin model provides a superior fit to species abundance distributions with a single free parameter: evidence, implementation and interpretation | |
| CN108280440A (en) | A kind of fruit-bearing forest recognition methods and system | |
| CN106709421B (en) | Cell image identification and classification method based on transform domain features and CNN | |
| CN108241901A (en) | A kind of transformer Forewarn evaluation method and device based on prediction data | |
| CN104680185B (en) | Hyperspectral image classification method based on boundary point reclassification | |
| Ramos et al. | Information-theoretical feature selection using data obtained by Scanning Electron Microscopy coupled with and Energy Dispersive X-ray spectrometer for the classification of glass traces | |
| CN107895136B (en) | Coal mine area identification method and system | |
| CN112215696A (en) | Personal credit evaluation and interpretation method, device, equipment and storage medium based on time sequence attribution analysis | |
| Koziarski et al. | DiagSet: a dataset for prostate cancer histopathological image classification | |
| CN114814167B (en) | Soil heavy metal content inversion method fusing multi-source environment variables and spectral information | |
| CN115147615A (en) | Rock image classification method and device based on metric learning network | |
| CN106778801A (en) | Method and apparatus for analyzing the classification of object of observation | |
| CN111401444A (en) | Method and device for predicting origin of red wine, computer equipment and storage medium | |
| CN108663334A (en) | The method for finding soil nutrient spectral signature wavelength based on multiple Classifiers Combination | |
| CN114445356A (en) | Multi-resolution-based full-field pathological section image tumor rapid positioning method | |
| Jang et al. | Deep neural networks for gas concentration estimation and the effect of hyperparameter optimization on the estimation performance | |
| US9773090B2 (en) | Method computer program and system to analyze mass spectra | |
| CN104504399B (en) | A kind of multispectral data supervised classification method of the linear Correlation Information Entropy of combination | |
| CN115186776B (en) | Method, device and storage medium for classifying ruby producing areas | |
| CN107657262A (en) | A kind of computer automatic sorting Accuracy Assessment | |
| CN112529112B (en) | Mineral identification method and device | |
| Grandremy et al. | The ZooScan and the ZooCAM zooplankton imaging systems are intercomparable: A benchmark on the Bay of Biscay zooplankton |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170531 |