WO2008012241A1 - Pattern classification method - Google Patents
Pattern classification method Download PDFInfo
- Publication number
- WO2008012241A1 WO2008012241A1 PCT/EP2007/057374 EP2007057374W WO2008012241A1 WO 2008012241 A1 WO2008012241 A1 WO 2008012241A1 EP 2007057374 W EP2007057374 W EP 2007057374W WO 2008012241 A1 WO2008012241 A1 WO 2008012241A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- test pattern
- training patterns
- neighbourhood
- class
- function
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
Definitions
- the present invention generally relates to a method of pattern classification or a system implementing such a method.
- Pattern classification is well known in a number of real-world applications, such as, for instance, speech recognition, vehicle seat occupant classification, data mining, risk prediction, diagnosis classification, etc.
- the primary goal of a pattern classifier is to assign a test pattern to one or more classes of a predefined set of classes.
- the test pattern may be considered as a vector of features or, more precisely, numbers quantifying these features.
- a statistical classifier computes the conditional probability of different classes for a given input pattern (hereinafter also referred to as "class membership probability"). The deviation of these class membership probabilities from 1 are often interpreted as a risk of a false classification.
- a challenge in pattern classification is the reduction of misclassifications.
- a classifier may exercise the reject option whenever none of the conditional probabilities of the different classes for a given input pattern exceeds a required minimum threshold. Otherwise, the classifier assigns the input pattern to the class with the highest conditional probability.
- a test pattern close to a decision boarder implicitly defined by the classifier is prone to be rejected, while a test pattern far away from the boarder will be assigned to a class.
- the interested reader is referred to the article "On Optimum Recognition Error and Reject Tradeoff' by C. K. Chow, IEEE Transactions on Information Theory, Vol. IT-16, No. 1 , January 1970.
- a classifier is usually trained, during a training process, by means of training patterns. These training patterns are preferably chosen according to different types (classes) of situations the classifier shall be able to distinguish.
- the class membership probabilities of a test pattern to be classified are based upon the training patterns used in the training process. Ideally, one would prepare the classifier for all types of situations that can occur. In real-world applications, this is most often impossible to achieve, e.g. ⁇ because of "unforeseeable" situations or limited resources. As a result, the feature space, i.e. the space spanned by all possible patterns, is not homogeneously populated with training patterns.
- a classifier that provides the certainty (or the uncertainty) of a class membership probability is attractive in a safety critical context, such as e.g. vehicle seat occupant classification, diagnosis classification, etc., since it allows labelling a test pattern as "unknown” and/or exercise the reject option if the uncertainty of the class membership probability is too high.
- the number of training patterns in the neighbourhood of the test pattern is obtained from computing a convolution of a density function of the training patterns with a Gaussian smoothing function centred on the test pattern, where the density function of the training patterns is represented as a mixture (superposition) of Gaussian functions.
- the number of training patterns in the neighbourhood is not obtained by actually counting the training patterns that lie within a certain distance from the test pattern.
- the training patterns are all stored in a memory.
- the parameters that define the convolution of the density function of the training patterns with a Gaussian smoothing function are stored in memory. Depending on the amount of training patterns, these parameters may require only a small fraction of the memory space necessary for storing the corresponding set of training patterns.
- x represents the test pattern
- N N (x) the number of training patterns in the neighbourhood of x
- K an integer
- ⁇ k a vector in the feature space
- SV a matrix
- N'k a real number
- K represents the number of Gaussian functions in the mixture
- x' a variable in the feature space
- ⁇ (x') the density of training patterns at x'
- ⁇ k the centre of the k-th Gaussian function
- S k a matrix describing the widths of the k-th Gaussian function
- d the dimension of the feature space
- N k represent normalisation factors fulfilling
- N to t is the total number of training patterns in the feature space.
- N N (x) ⁇ p(x')g(x',x,r)dx' , (5)
- the smoothing function may be expressed as:
- x represents the test pattern
- x' a variable in the feature space
- d the dimension of the feature space
- C a symmetric matrix defining a metric on the feature space (such as, for instance, the covariance matrix of the training patterns)
- r a radius of the neighbourhood with respect to this metric.
- N ⁇ N k d Q t(T k S- 1 ) ,
- T k (C- l /r 2 + S k - l y l .
- the confidence interval can be calculated based upon the formula:
- p e represents the (estimate of the) class membership probability for the test pattern (as obtained by evaluating the class membership probability function for the test pattern)
- p+ the upper boundary of the confidence interval and p.
- the lower boundary of the confidence interval and ⁇ is representative of a predefined confidence level.
- the confidence level can be set according to the application.
- the term "confidence interval" shall not be interpreted as limited to the interval [p-,p+]; rather it shall be interpreted as also encompassing the intervals [0,p+] and [p.,1]. Therefore, providing at least one of P- and p+ is regarded as providing a confidence interval in the sense of the present invention.
- a plurality of class membership probabilities for the test patterns are calculated and the confidence interval is calculated for the highest one of the class membership probabilities.
- the test pattern may subsequently be assigned to the class for which the class membership probability is highest only if a lower boundary of the confidence interval exceeds a predefined threshold.
- the test pattern may be classified as unknown or be assigned to another class.
- a sequence of test patterns e.g. in a seat occupancy classification system
- if the classification of a given test pattern based upon the class membership probabilities is deemed unreliable then one could assign this pattern to the same class as the last preceding test pattern.
- the above-described method may also be used for classifying test patterns of a sequence of test patterns.
- each test pattern of the sequence is assigned to a class chosen from the predefined set of classes and the class to which the respective pattern has been assigned is returned as a preliminary classification.
- a quality factor associated to the respective preliminary classification is then determined based upon the confidence interval of the class membership probability or upon the number of training patterns in the neighbourhood of the test pattern considered. This quality factor is used to filter out those preliminary classifications that do not meat certain criteria with regard to the quality factor.
- the quality factor could, for instance, be the logarithm of the reciprocal width of the confidence interval or the number of training patterns, which is in the most straightforward embodiment compared to a threshold value. Alternatively, the quality factor could also be a derived according to another suitable rule.
- the filtering could e.g. be made by Kalman filtering technique, where the quality factor would be used for weighing the different preliminary classifications.
- the method is implemented in a vehicle seat occupant classification system and comprises providing sensor data relating to the occupant and using the sensor data as the test pattern to be assigned to a particular occupant class.
- Fig. 1 is an illustration of a 2D feature space with two clouds of data points representing training patterns belonging to two different classes;
- Fig. 2 is shows the division of the feature space of Fig. 1 into two regions separated by a decision boarder;
- Fig. 3 shows the contour lines of the lower bound of the confidence interval of the class membership probabilities;
- Fig. 4 is a flow diagram of a preferred embodiment of the method according to the invention.
- Fig 5 is an illustration of the classification of a sequence of test patterns.
- Fig. 1 shows an example of a two-dimensional feature space with two clouds 10, 12 of data points 14 representing training patterns belonging to two different classes.
- the patterns of the feature space can be unambiguously expressed as array having as array elements the coordinates of the corresponding data points 14.
- Collection of training patterns can e.g. be achieved by exposing the sensor or sensors whose outputs are to be assigned to different classes to situations whose classification is known.
- the representation of the collected patterns yields two clouds 10, 12 corresponding to a first class and a second class.
- Fig. 2 shows the separation of the feature space by a second order discriminant function that has been trained using the data shown in Figure 1 following the method described by J. Sch ⁇ rmann in "Pattern Classification: Statistical and Neural Network based Approaches", John Wiley and Sons, New York (1990).
- Contour lines 16 indicate a decrease of the class membership probability functions from the clouds 10, 12 towards the decision boarder 18 as Chow has suggested it in the paper referenced above. At the decision boarder 18 itself the probability is 0.5 for either of the classes.
- the class membership probability functions do not account for the fact that there are regions in feature space where no training data were selected.
- test patterns have approximately the same class membership probabilities.
- P1 is located within the cloud 12 of training patterns belonging to a common class, while P2 is located substantially outside any cloud of training patterns. Therefore, it is intuitively clear that the class membership probabilities of P2 suffer from higher uncertainty than those of P1.
- the present method thus proposes to compute a confidence interval for a class membership probability of a given test pattern based upon the number of training patterns in the vicinity of the test pattern to be classified.
- the concept of a confidence interval for an estimated probability is well established for a so-called Bernoulli process, which follows statistically a Binomial distribution.
- the Binomial distribution gives the discrete probability distribution P(n
- N) of obtaining exactly n successes out of N Bernoulli trials (where the result of each Bernoulli trial is true with probability p and false with probability q 1 -p).
- the binomial distribution is given by,
- N the limit N>>1/(p(1 -p))
- the Binomial distribution approaches a Gaussian distribution.
- ⁇ is determined by the confidence level that has been chosen.
- the relationship between ⁇ and the confidence level is given by the so-called erf- function, for evaluation of which look-up tables are normally used:
- Equation (10) is now used in the case of a statistical classifier: p is interpreted as the actual class membership probability and N as the number of samples in the neighbourhood of a test point (denoted N N ).
- ⁇ (p e ) [T') l + ⁇ 2 /N ⁇
- Fig. 3 shows the contour lines 20 of the lower bound (p.) of the confidence interval of the estimated class membership probability.
- the density of training patterns has been approximated by two Gaussian functions, one for each cloud 10, 12.
- regions with high training pattern density e.g. around P1
- the size of the confidence interval is very small, so that the resulting lower bound p. is almost equal to the estimated posterior probability p e itself.
- regions a lower density of training patterns, e.g. around P2 the width of the confidence interval is increased and the lower bound of the confidence interval tends to zero.
- the regions outside the outer contour 20.1 correspond to a lower bound of the confidence interval below 0.1. If one requires that the lower bound of the confidence level is at least 0.1 for a test pattern being assigned to a class, test patterns in these regions can be detected as "unknown" and be rejected in order to avoid a false classification.
- Fig. 4 shows a flow diagram of a preferred embodiment of the method for pattern classification.
- the method comprises a certain number of steps that are preferably executed offline and a certain number of steps that are performed online after a test pattern to be classified has been determined.
- the offline steps are represented on the left side of the flow diagram. These steps essentially comprise the training of the classifier, the setting of parameters and the storage of those data that are necessary for the online steps in a memory.
- the test pattern to be classified will be provided in form of a sensor output of one or more sensors.
- the training patterns one may expose the sensor or the sensors to situations that are expected to be encountered during the operation of the classification system and the sensor outputs are collected. Alternatively or additionally the training patterns could also be obtained from simulations of the sensor output.
- the class membership functions are computed (step 44). These class membership functions later take a test pattern as input and output estimated probabilities of this test pattern belonging to the different classes. Additionally, the training patterns are used to generate the function that serves to compute the convolution of the training pattern density with a Gaussian smoothing function (steps 46, 47 and 48). In step 46 the density of training patterns is approximated with a Gaussian mixture model, i.e. a superposition of a finite number of Gaussian functions. Those parameters that determine the neighbourhood of a test pattern and that can be set or computed offline, e.g. the widths and/or the shape of the neighbourhood, are fixed in step 47.
- the convolution can be expressed as in equation (2) and the parameters determining this expression are stored for making them available when the number of training patterns in the neighbourhood of a test pattern is to be computed (step 48). It will be appreciated that the time-consuming computation of the Gaussian mixture model (step 46) and the class membership functions (step 44) does not need being repeated during the classification of a test pattern. It should be said for completeness, that the computation 48 of the class membership functions might also be based upon the Gaussian mixture model of the training pattern density.
- the online steps of this embodiment of the method include first of all the acquisition 50 of a test pattern from the sensor or sensors.
- the class membership functions are retrieved from a memory and the different class membership probabilities of the test pattern are computed (step 52).
- step 54 the parameters defining the equation (2) and thus the convolution to be computed are retrieved from a memory and the number of training data in the neighbourhood of the test pattern is determined.
- step 56 the confidence intervals for the highest class membership probability found in 52 is calculated (step 56).
- the class the test pattern has been assigned to is output in step 58. In case the uncertainty of the class membership probability is deemed too high to name a class, the output may be "unknown".
- Fig. 5 illustrates how the method can be used for classifying test patterns of a sequence of test patterns corresponding to real-life situations 60, 62, 64, 66.
- the classification system has been trained to distinguish a face (class 1 in the example) from an object (class 2), whereby features extracted from camera pictures are used as input data.
- Features indicative of a face may include mouth, eyes, ears, etc.
- Sketches of situations 60, 62, 64, 66 along time axis 68 represent a child making a bubble of chewing gum.
- the class membership probabililties are computed and the class with the highest probability is returned as preliminary classification.
- a quality factor Q associated to the respective preliminary classification is determined based upon the confidence interval of the class membership probability or upon the number of training patterns in the neighbourhood of the test pattern considered.
- a quality factor close to 100% indicates, in the example, that the uncertainty of the preliminary classification is low and that the classification should be used as output 69. If the system encounters a situation that was not expected during training (situation 64), the preliminary classification may be wrong. But in the same time the low quality factor indicates that the preliminary classification is unreliable. In presence of the unknown situation 64, the system thus does not consider the preliminary classification as valid and discards it (shown at reference numeral 70). As output, the system uses in this case the last valid classification, which is in this case the classification of situation 62.
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)
- Complex Calculations (AREA)
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP07787640.7A EP2052351B1 (en) | 2006-07-28 | 2007-07-17 | Pattern classification method |
JP2009521221A JP2009545045A (en) | 2006-07-28 | 2007-07-17 | Pattern classification method |
CN200780028017.6A CN101496035B (en) | 2006-07-28 | 2007-07-17 | Method for classifying modes |
US12/374,076 US8346684B2 (en) | 2006-07-28 | 2007-07-17 | Pattern classification method |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP06118110.3 | 2006-07-28 | ||
EP06118110A EP1883040A1 (en) | 2006-07-28 | 2006-07-28 | Pattern classification method |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2008012241A1 true WO2008012241A1 (en) | 2008-01-31 |
Family
ID=37546914
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/EP2007/057374 WO2008012241A1 (en) | 2006-07-28 | 2007-07-17 | Pattern classification method |
Country Status (5)
Country | Link |
---|---|
US (1) | US8346684B2 (en) |
EP (2) | EP1883040A1 (en) |
JP (1) | JP2009545045A (en) |
CN (1) | CN101496035B (en) |
WO (1) | WO2008012241A1 (en) |
Families Citing this family (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE112010004038T5 (en) | 2009-10-16 | 2012-09-20 | Kesumo, Llc | Foot operated controller |
WO2012078503A2 (en) * | 2010-12-05 | 2012-06-14 | Brown University | Methods for prediction and early detection of neurological events |
US9076419B2 (en) | 2012-03-14 | 2015-07-07 | Bebop Sensors, Inc. | Multi-touch pad controller |
US9460135B2 (en) * | 2012-12-18 | 2016-10-04 | Webtrends Inc. | Methods and automated systems for testing, optimization, and analysis that use robust statistical processing of non-binomial experimental results |
WO2015089115A1 (en) | 2013-12-09 | 2015-06-18 | Nant Holdings Ip, Llc | Feature density object classification, systems and methods |
US9753568B2 (en) | 2014-05-15 | 2017-09-05 | Bebop Sensors, Inc. | Flexible sensors and applications |
US9442614B2 (en) | 2014-05-15 | 2016-09-13 | Bebop Sensors, Inc. | Two-dimensional sensor arrays |
US9965076B2 (en) | 2014-05-15 | 2018-05-08 | Bebop Sensors, Inc. | Piezoresistive sensors and applications |
US10362989B2 (en) | 2014-06-09 | 2019-07-30 | Bebop Sensors, Inc. | Sensor system integrated with a glove |
US20170316265A1 (en) * | 2014-07-23 | 2017-11-02 | Hrl Laboratories, Llc | Parzen window feature selection algorithm for formal concept analysis (fca) |
US10671917B1 (en) | 2014-07-23 | 2020-06-02 | Hrl Laboratories, Llc | System for mapping extracted Neural activity into Neuroceptual graphs |
CN104180794B (en) * | 2014-09-02 | 2016-03-30 | 西安煤航信息产业有限公司 | The disposal route in digital orthoimage garland region |
US9863823B2 (en) | 2015-02-27 | 2018-01-09 | Bebop Sensors, Inc. | Sensor systems integrated with footwear |
US10082381B2 (en) | 2015-04-30 | 2018-09-25 | Bebop Sensors, Inc. | Sensor systems integrated with vehicle tires |
US9827996B2 (en) | 2015-06-25 | 2017-11-28 | Bebop Sensors, Inc. | Sensor systems integrated with steering wheels |
US10891942B2 (en) | 2016-03-03 | 2021-01-12 | Telefonaktiebolaget Lm Ericsson (Publ) | Uncertainty measure of a mixture-model based pattern classifer |
FR3066296A1 (en) * | 2017-05-10 | 2018-11-16 | Stmicroelectronics (Rousset) Sas | METHOD FOR CONTROLLING DETECTION, WITH FILTERING, OF SCENES BY AN APPARATUS, FOR EXAMPLE A WIRELESS COMMUNICATION APPARATUS, AND CORRESPONDING APPARATUS |
JP6487493B2 (en) * | 2017-05-18 | 2019-03-20 | ファナック株式会社 | Image processing system |
US10884496B2 (en) | 2018-07-05 | 2021-01-05 | Bebop Sensors, Inc. | One-size-fits-all data glove |
US11480481B2 (en) | 2019-03-13 | 2022-10-25 | Bebop Sensors, Inc. | Alignment mechanisms sensor systems employing piezoresistive materials |
US11386636B2 (en) | 2019-04-04 | 2022-07-12 | Datalogic Usa, Inc. | Image preprocessing for optical character recognition |
CN112017273B (en) * | 2019-05-28 | 2022-12-13 | 四川大学 | Multi-resolution core pore fusion method based on mode density function simulation algorithm |
CN111539456B (en) * | 2020-04-02 | 2024-03-01 | 浙江华睿科技股份有限公司 | Target identification method and device |
CN113505064B (en) * | 2021-07-07 | 2022-05-17 | 广东电力信息科技有限公司 | Heterogeneous information flow-based electric power big data service system testing method |
CN113707136B (en) * | 2021-10-28 | 2021-12-31 | 南京南大电子智慧型服务机器人研究院有限公司 | Audio and video mixed voice front-end processing method for voice interaction of service robot |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7415126B2 (en) * | 1992-05-05 | 2008-08-19 | Automotive Technologies International Inc. | Occupant sensing system |
US5335291A (en) * | 1991-09-20 | 1994-08-02 | Massachusetts Institute Of Technology | Method and apparatus for pattern mapping system with self-reliability check |
JP4039708B2 (en) * | 1997-02-14 | 2008-01-30 | ヒューレット・パッカード・カンパニー | Pattern recognition method and pattern recognition apparatus |
US6137911A (en) * | 1997-06-16 | 2000-10-24 | The Dialog Corporation Plc | Test classification system and method |
JP2003323601A (en) * | 2002-05-01 | 2003-11-14 | Fujitsu Ltd | Predicting device with reliability scale |
JP2004341930A (en) * | 2003-05-16 | 2004-12-02 | Nippon Telegr & Teleph Corp <Ntt> | Method and device for recognizing pattern |
JP2004341959A (en) * | 2003-05-16 | 2004-12-02 | Just Syst Corp | Data classification device, data classification method, and program for making computer execute the method |
US7738705B2 (en) * | 2004-06-30 | 2010-06-15 | Stefano Casadei | Hierarchical method and system for pattern recognition and edge detection |
-
2006
- 2006-07-28 EP EP06118110A patent/EP1883040A1/en not_active Withdrawn
-
2007
- 2007-07-17 WO PCT/EP2007/057374 patent/WO2008012241A1/en active Application Filing
- 2007-07-17 JP JP2009521221A patent/JP2009545045A/en active Pending
- 2007-07-17 CN CN200780028017.6A patent/CN101496035B/en active Active
- 2007-07-17 US US12/374,076 patent/US8346684B2/en active Active
- 2007-07-17 EP EP07787640.7A patent/EP2052351B1/en active Active
Non-Patent Citations (2)
Title |
---|
CHEE PENG LIM ET AL: "Application of the Gaussian mixture model to drug dissolution profiles prediction", NEURAL COMPUTING & APPLICATIONS SPRINGER-VERLAG UK, vol. 14, no. 4, October 2005 (2005-10-01), pages 345 - 352, XP002413035, ISSN: 0941-0643 * |
PAALANEN ET AL: "Feature representation and discrimination based on Gaussian mixture model probability densities-Practices and algorithms", PATTERN RECOGNITION, ELSEVIER, KIDLINGTON, GB, vol. 39, no. 7, July 2006 (2006-07-01), pages 1346 - 1358, XP005403205, ISSN: 0031-3203 * |
Also Published As
Publication number | Publication date |
---|---|
EP2052351B1 (en) | 2013-08-21 |
CN101496035A (en) | 2009-07-29 |
US20090319451A1 (en) | 2009-12-24 |
EP2052351A1 (en) | 2009-04-29 |
EP1883040A1 (en) | 2008-01-30 |
US8346684B2 (en) | 2013-01-01 |
JP2009545045A (en) | 2009-12-17 |
CN101496035B (en) | 2015-12-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8346684B2 (en) | Pattern classification method | |
CN109934293B (en) | Image recognition method, device, medium and confusion perception convolutional neural network | |
US11669724B2 (en) | Machine learning using informed pseudolabels | |
US10204283B2 (en) | Image recognizing apparatus, image recognizing method, and storage medium | |
US8725660B2 (en) | Applying non-linear transformation of feature values for training a classifier | |
CN109271958B (en) | Face age identification method and device | |
US20070065003A1 (en) | Real-time recognition of mixed source text | |
CN111079639A (en) | Method, device and equipment for constructing garbage image classification model and storage medium | |
US20060088207A1 (en) | Object recognizer and detector for two-dimensional images using bayesian network based classifier | |
US9798955B2 (en) | Image processing apparatus, photographic subject identifying method and program | |
CN112766427A (en) | Training method, device and equipment of image recognition model | |
CN111583146A (en) | Face image deblurring method based on improved multi-scale circulation network | |
US9842280B2 (en) | System and method for evaluating a classifier implemented within an image signal processor | |
CN111340057A (en) | Classification model training method and device | |
Fieldsend et al. | Formulation and comparison of multi-class ROC surfaces | |
KR101590896B1 (en) | Device and method for deep learning structure for high generalization performance, recording medium for performing the method | |
US20220284261A1 (en) | Training-support-based machine learning classification and regression augmentation | |
CN112183336A (en) | Expression recognition model training method and device, terminal equipment and storage medium | |
CN113971442A (en) | Method and system for generating universal countermeasure disturbance based on self-walking learning | |
Gips | Bayesian seafloor characterization from SAS imagery | |
JP6596044B2 (en) | Object recognition device | |
CN113761292A (en) | Object identification method and device, computer equipment and storage medium | |
Lázaro et al. | Classification of binary imbalanced data using a Bayesian ensemble of Bayesian neural networks | |
Yu et al. | Discretization-Induced Dirichlet Posterior for Robust Uncertainty Quantification on Regression | |
Ozawa et al. | An autonomous incremental learning algorithm for radial basis function networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 200780028017.6 Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 07787640 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2009521221 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2007787640 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: RU |
|
WWE | Wipo information: entry into national phase |
Ref document number: 12374076 Country of ref document: US |