WO2009093525A1 - 識別装置、識別方法及び識別処理プログラム - Google Patents
識別装置、識別方法及び識別処理プログラム Download PDFInfo
- Publication number
- WO2009093525A1 WO2009093525A1 PCT/JP2009/050529 JP2009050529W WO2009093525A1 WO 2009093525 A1 WO2009093525 A1 WO 2009093525A1 JP 2009050529 W JP2009050529 W JP 2009050529W WO 2009093525 A1 WO2009093525 A1 WO 2009093525A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- classification
- observation data
- identification
- pseudo
- normal distribution
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
Definitions
- the present invention relates to an identification device, an identification method, and an identification processing program, and is suitable for application to an identification device that identifies a predetermined target based on, for example, a Bayesian decision rule of a mixed normal distribution.
- the identification target is identified according to the Bayesian decision rule, assuming a probability distribution model.
- Patent Document 1 Japanese Patent Document 1
- the Bayesian decision rule used here is a method in which classification (class) is given in advance and the class to which the observation data (feature vector) belongs is discriminated.
- the occurrence probability of event B is expressed as P (B)
- P (A)> 0 the following equation holds.
- conditional probabilities P [C c 0
- x] and P [C c 1
- x] of each class c 0 , c 1 are compared, and a high probability Is selected as the class to which the observation data belongs. This selection minimizes the error rate.
- x] is expressed by the following equation from the Bayes' theorem described above:
- Equation 5 a conditional probability density function p (x
- C) is modeled as a multi-dimensional (here, D-dimensional, for example) mixed normal distribution (multivariate mixed Gaussian distribution)
- the D-dimensional mixed normal distribution is expressed as shown in Equation 5 below. Is done.
- D is the number of variables
- x is the observed data (feature vector)
- ⁇ is the D ⁇ 1 mean vector
- ⁇ is the D ⁇ D covariance matrix (covariance is how much the two data are related / linked The coefficient indicating whether or not there is).
- ⁇ ) of the D-dimensional mixed normal distribution model is obtained by using the above equation 5.
- Equation 6 ⁇ n ⁇ , ⁇ n ⁇ , ⁇ n ⁇ , N is the number of normal distributions, and ⁇ n is a mixing ratio.
- the mixed normal distribution represented by Equation 6 can be approximated with an arbitrary accuracy by modeling a complex shape distribution by modeling with an appropriate number of mixtures.
- ⁇ k, n ⁇ is a parameter set given to the class c k , and k is 0 or 1 indicating the class (classification).
- variable locations are calculated as variable groups (z k , n (x)) based on the random variable vectors x of the plurality of observation data, and based on the parameters of the mixed normal distribution,
- a constant part is calculated as a constant group K k , n (k is 1 or 0 indicating a classification, and n indicates a distribution number of a normal distribution assumed for each classification).
- the present invention has been made in consideration of the above points, and an object thereof is to propose an identification device, an identification method, and an identification processing program that can significantly reduce the processing burden.
- claim 1 of the present invention is an identification device that classifies the observation data based on a mixed normal distribution parameter on the assumption that the distribution of the observation data follows a mixed normal distribution.
- a variable group z k , n obtained from each feature vector of data (k is 1 or 0 representing a classification, n represents a distribution number of a normal distribution assumed for each classification), and the mixed normal distribution parameter Using the constant group K k , n obtained from the group,
- a cumulative addition unit that calculates the g k upper and the g k lower , and a comparison unit that classifies the observation data by g 1 upper ⁇ g 0 lower and g 0 upper ⁇ g 1 lower It is characterized by providing.
- an identification method for classifying the observation data based on a mixed normal distribution parameter on the assumption that the distribution of the observation data follows a mixed normal distribution Variable group z k , n obtained based on vector (k is 1 or 0 representing classification, n represents distribution number of normal distribution assumed for each classification) and obtained mixed normal distribution parameter Using the constant group K k , n , the following equation
- N k indicates the number of normal distributions of classification k
- h k Is referred to as the h k, characterized in that h k multiplies the n lower, update the n lower comprising a refining process step of calculating the g k lower.
- claim 7 of the present invention is an identification processing program for classifying the observation data based on a mixed normal distribution parameter on the assumption that the distribution of the observation data follows a mixed normal distribution, and each of the plurality of observation data Variable groups z k , n obtained based on feature vectors (k is 1 or 0 representing classification, n represents distribution number of normal distribution assumed for each classification) and obtained mixed normal distribution parameters Using the constant group K k , n
- N k indicates the number of normal distributions of classification k
- h k Is referred to as the h k, characterized in that h k multiplies the n lower, update the n lower comprising a refining process step of calculating the g k lower.
- claim 9 of the present invention is that when the value of i is L, the following equation
- the observation data is obtained from the magnitude relationship between g k upper and g k lower obtained by simple interval calculation. Since it can be identified, it is possible to avoid the calculation processing of a complex exponential function for calculating the occurrence probability value of the classification, and thus the processing load can be significantly reduced.
- the accuracy of the comparison operation is improved step by step, and the unknown section is narrowed and simplified.
- the observation data can be classified by the interval calculation, and thus it is possible to avoid the calculation process of the complex exponential function for calculating the occurrence probability value of the classification.
- the identification device of claim 3 the identification method of claim 6, and the identification processing program of claim 9, the observation data is finally obtained even if the magnitudes of g k upper and g k lower are not clear. Can be reliably classified.
- reference numeral 1 denotes an identification device according to the present invention, and it is assumed that the distribution of the observation data D1 follows a mixed normal distribution, and the observation data D1 is classified into 0 or 0 based on the mixed normal distribution parameters. It is possible to identify whether it is any of classification 1.
- classification result is set to classification 0 or classification 1, here, this merely indicates a positive result that the identification target exists or a negative result that the identification target does not exist.
- each pixel is either a background image (class 0) or a person (class 1).
- the identification device 1 inputs image data obtained from the television camera to the feature extraction unit 2 as observation data D1.
- the feature extraction unit 2 performs predetermined image processing, and includes, for example, a feature vector including coordinate axes (x, y) representing the position in the image plane for each pixel and various pieces of pixel information such as the luminance value and color of the pixel. D2 is calculated and sent to the secondary format calculation unit 3.
- the identification device 1 classifies each pixel into a skin color (class 1) and a non-skin color (class 0) in order to identify a person from the background image.
- the feature vector D2 includes coordinate data indicating the position of the pixel and data indicating each value of R (red), G (green), and B (blue) in the pixel represented by the coordinate data. .
- the identification device 1 assumes that the distribution of the observation data D1 follows a mixed normal distribution, and converts the M-dimensional mixed normal distribution parameter group D3 including a plurality of components obtained from the pixel set distribution into a quadratic form.
- the data is sent to the calculation unit 3 and the weight coefficient calculation unit 4.
- the secondary form calculation unit 3 extracts only variables based on the feature vector D2 and the mixed normal distribution parameter group D3, calculates a variable group (z k , n ), and sends this to the identification unit 5.
- the weighting factor calculation unit 4 extracts only constants based on the received mixed normal distribution parameter group D3, calculates a constant group (K k , n ), and sends it to the identification unit 5.
- variable group (z k , n ) and the constant group (K k , n ) as shown in FIG. 2, the variable location ER1 is the variable group (z k , n ) And the constant part ER2 is calculated as a constant group (K k , n ).
- the variable group (z k , n (x)) is represented, but since it is clear that it is a function of the variable x, hereinafter, for convenience of explanation, the variable group (z k , n ).
- the identification unit 5 indicates the probability of becoming a classification 0 (hereinafter referred to as a classification 0 occurrence probability)
- a lower limit value that is smaller than the classification 0 occurrence probability value and does not require calculation of an exponential function (hereinafter referred to as a classification 0 occurrence lower limit value).
- an upper limit value that is larger than the classification 0 occurrence probability value and does not require calculation of an exponential function (hereinafter referred to as a classification 0 occurrence upper limit area).
- the identification unit 5 indicates the probability of becoming a classification 1 (hereinafter referred to as a classification 1 occurrence probability)
- the lower limit value that is smaller than the classification 1 occurrence probability value and does not require the calculation of the exponential function (hereinafter referred to as the classification 1 occurrence lower limit value) is executed without actually calculating the exponential function. )
- an upper limit value that is larger than the classification 1 occurrence probability value and does not require calculation of an exponential function (hereinafter referred to as a classification 1 occurrence upper limit value).
- K k and n are positive constants
- z k and n are positive variables.
- the identification unit 5 compares the classification 0 occurrence upper limit value and the classification 1 occurrence lower limit value, or compares the classification 1 occurrence upper limit value and the classification 0 occurrence lower limit value, thereby comparing the classification 0 occurrence probability value and the classification 1 occurrence value. It is determined which of the occurrence probability values is greater.
- the observation data D1 is identified as the classification 0.
- the classification 1 occurrence probability value is clearly larger than the classification 0 occurrence probability value
- the observation data D1 is identified as the classification 1.
- the identification unit 5 can calculate these classification 0 upper limit, classification 0 occurrence lower limit, class 1 occurrence upper limit, and class 1 occurrence lower limit that can be calculated by simple calculation without performing complicated calculation of the exponential function. Is used to identify whether the observation data D1 is either classification 0 or classification 1 with a simple interval calculation by comparing the magnitudes of the classification 0 occurrence probability value and the classification 1 occurrence probability value. It is made to be able to do.
- the identification unit 5 when the magnitude of the classification 0 occurrence probability value and the classification 1 occurrence probability value is unknown, the identification unit 5 generates the classification 0 occurrence until the magnitude relationship between the classification 0 occurrence probability value and the classification 1 occurrence probability value becomes clear.
- a simple interval calculation that does not calculate the exponential function by gradually narrowing the upper limit, the classification 0 occurrence lower limit, the classification 1 occurrence upper limit, and the classification 1 occurrence lower limit in steps.
- the observation data D1 can be identified as to whether it is a classification 0 or a classification 1.
- FIG. 3 shows a circuit configuration of the identification unit 5 that executes the identification process
- FIG. 4 is a flowchart showing the identification process procedure.
- the identification unit 5 sends the constant group (K k , n ) from the weight coefficient calculation unit 4 to the power-of-two multiplication unit 10.
- k is 0 indicating classification 0 or 1 indicating classification 1
- N 0 and N 1 are the number of normal distributions.
- log 2 e used in the identification process in the identification unit 5 and the following equations 32 and 33 used in (2-2) refinement processing described later are calculated in advance. The result is recorded, and the burden of calculation processing is reduced.
- i 0, 1,...
- the identification unit 5 sends the variable group (z k , n ) from the secondary form calculation unit 3 to the multiplication unit 11 and multiplies the variable group (z k , n ) by the constant log 2 e.
- the plurality of (z k , n log 2 e) obtained in this way are sent to the dividing unit 12.
- the dividing unit 12 divides each (z k , n log 2 e) into an integer part and a decimal part, and the integer part is divided into an integer power group part (“z k , n log 2 e”). Send it out.
- the decimal part obtained from Equation 34 of the following equation is used as a decimal part group ( ⁇ k , n ) for use in the refinement process described later, and the B [i] selection part 14 and the first part of the refinement process part 13. 1 is sent to each selection unit 15.
- the symbol “” in the integer part group described above indicates an integer part.
- the cumulative addition unit 17 cumulatively adds a plurality of h k , n upper calculated by sequentially changing k from 0 to n to 1..., N 0 according to the following expression 37, G 0 upper is calculated (step SP 2) and sent to the comparison unit 18.
- the cumulative addition unit 17 cumulatively adds a plurality of h k , n upper calculated by sequentially changing k from 1 to n 1 to N 1 according to the above-described equation 37, and generating a classification 1
- the upper limit g 1 upper is calculated (step SP 2), and this is sent to the comparison unit 18.
- the cumulative addition unit 17 cumulatively adds a plurality of h k , n lower calculated by sequentially changing k from 1 to n 1 to N 1 by the above-described equation 38, and the lower limit of occurrence of classification 1 G 1 lower is calculated (step SP 2), and this is sent to the comparison unit 18.
- the comparison unit 18 compares g 1 upper and g 0 lower , and determines whether or not g 1 upper ⁇ g 0 lower clearly from the numerical values as shown in FIG. 5A (step SP3). ). When the comparison unit 18 determines that g 1 upper ⁇ g 0 lower indicating that the classification 0 occurrence probability value is clearly larger than the classification 1 occurrence probability value, the comparison unit 18 determines that the classification is 0 (step SP4). The result is sent to a display unit (not shown).
- the comparison unit 18 compares g 0 upper and g 1 lower to obtain the numerical values as shown in FIG. Clearly determines whether or not g 0 upper ⁇ g 1 lower (step SP5). If the comparison unit 18 determines that g 0 upper ⁇ g 1 lower indicating that the classification 1 occurrence probability value is clearly larger than the classification 0 occurrence probability value, the comparison unit 18 determines that the classification is 1 (step SP6). Send the result to the display.
- the display unit can notify the user by displaying the determination result received from the comparison unit 18 via an image.
- skin color identification for example, pixels of skin color (category 1) are set to white based on the result determined for each pixel as skin color (category 1) or non-skin color (category 0). ) Is black, and each pixel is color-coded for each of classification 0 and classification 1, so that an image identifying a background image and a person can be generated and notified to the user.
- pixels that need to be refined which will be described later, are color-coded in gray, and the user can be notified that the refinement process has been performed, or that the basic identification process described above has not been performed.
- the comparison unit 18 compares the numerical values of g 1 upper and g 0 lower or compares the numerical values of g 0 upper and g 1 lower by the basic identification processing described above. As shown in FIG. 6 (A), when these magnitude relationships cannot be determined, a reprocessing signal indicating the determination impossible result is sent to the refinement processing unit 13 so that the refinement process can be executed. Has been made.
- the refinement processing unit 13 includes a storage unit 20 that stores a lookup table (LUT) in which the relationships such as Equations 32 and 33 described above are associated.
- LUT lookup table
- i is incremented by 1 when the magnitude relationship between g 1 upper and g 0 lower and g 0 upper and g 1 lower cannot be determined (when a negative result is obtained in step SP5).
- the maximum value is set to L, it is first determined whether i is currently L (step SP7). When i has not yet reached L, i Is updated to i + 1 (step SP8).
- the decimal part group calculated by the dividing unit 12 by the number 34 as described above (beta k, n)
- the fractional portion group (beta k, n) decimal the i-position of the fractional part constituting the is 1 depending whether or, alternatively 0, respectively different individual processing for each fractional part, each h k updated by performing these different processes, n lower and h k, with n upper, the cumulative addition unit 17 g k upper and g k lower are calculated.
- decimal part when the i-th decimal place is 1 (hereinafter, simply referred to as “1 decimal part”) and the decimal part when the i-th decimal place is 0 (hereinafter, this is simply referred to as “0 decimal number”).
- 1 decimal part the decimal part when the i-th decimal place is 1
- 0 decimal number the decimal part when the i-th decimal place is 0
- the B [i] selection unit 14 of the refinement processing unit 13 calculates the decimal number calculated by the dividing unit 12 using the above-described Expression 34. For each decimal part of the group ( ⁇ k , n ), it is individually determined whether or not the i-th place after the decimal point is 1 (step SP9).
- B [i] selection unit 14 determines that one decimal part exists in the decimal part group ( ⁇ k , n )
- B [i] is read from the storage unit 20 when the one decimal part is processed. To the multiplication unit 21.
- the upper and lower limit value dividing unit 22 of the refinement processing unit 13 corresponds to the corresponding h k , n lower (hereinafter referred to as one corresponding h k , n) calculated by the integer part of the variable for which the one decimal part is obtained. and referred to as lower), also the corresponding h k calculated by the integer part of the variable this one decimal part obtained, n upper (hereinafter, from 1 correspondence h k, called the n upper) and the selector 16 receive.
- the upper and lower limit value dividing unit 22 divides the one corresponding h k and n lower and the one corresponding h k and n upper and sends the one corresponding h k and n upper to the first selecting unit 15 and the second selecting unit 23. At the same time, 1 correspondence h k , n lower is also sent to the first selection unit 15 and the second selection unit 23.
- the first selection unit 15 recognizes that the currently processed decimal part is one decimal part based on the decimal part group ( ⁇ k , n ) received from the dividing part 12, and 1 correspondence h k , n upper is selected and sent to the multiplier 21, and 1-corresponding h k , n lower is discarded.
- the first selection unit 15 sends a selection signal to the second selection unit 23 so as to select one corresponding h k , n lower .
- the second selection unit 23 selects one corresponding h k , n lower and sends it to the update value generation unit 25 to discard the one corresponding h k , n upper .
- the multiplication unit 21 multiplies 1 corresponding h k and n upper by B [i], updates only 1 corresponding h k and n upper (step SP10), and obtains 1 corresponding h k and n upper B obtained as a result. [i] (hereinafter referred to as update h k , n upper ) is sent to the update value generation unit 25.
- Updating value generator 25 updates h k, receives the n upper, update h k, and n upper and 1 corresponding h k, and n lower pair, and sends to the cumulative addition unit 17 via the selector 16 .
- the upper and lower limit value dividing unit 22 of the refinement processing unit 13 corresponds to h k , n lower (hereinafter referred to as 0 corresponding h k , n) calculated by the integer part of the variable from which the 0 decimal part is obtained. and referred to as lower), also the corresponding h k the 0 fractional part calculated by the integer part of a variable obtained, n upper (hereinafter, from 0 corresponding h k, called the n upper) and the selector 16 receive.
- the upper / lower limit value dividing unit 22 divides these 0 corresponding h k and n lower and 0 corresponding h k and n upper and sends the 0 corresponding h k and n lower to the first selecting unit 15 and the second selecting unit 23. At the same time, 0 corresponding h k , n upper is also sent to the first selection unit 15 and the second selection unit 23.
- the first selection unit 15 recognizes that the currently processed decimal part is a 0 decimal part based on the decimal part group ( ⁇ k , n ) received from the upper and lower limit value dividing part 22, The 0 corresponding h k and n lower are selected and sent to the multiplication unit 21 and the 0 corresponding h k and n upper are discarded.
- the first selection unit 15 sends a selection signal to the second selection unit 23 so as to select 0 corresponding h k , n upper .
- the second selection unit 23 selects 0 corresponding h k , n upper and sends it to the update value generation unit 25, and discards the 0 corresponding h k , n lower .
- the multiplication unit 21 multiplies the corresponding h k , n lower by B [i] ⁇ 1 and updates only h k , n lower (step SP11), and the 0 corresponding h k , n lower B [ i] ⁇ 1 (hereinafter referred to as update h k , n lower ) is sent to the update value generation unit 25.
- the update value generation unit 25 When the update value generation unit 25 receives the update h k , n lower , the update h k , n lower and the 0 corresponding h k , n upper are paired and sent to the accumulation addition unit 17 via the selection unit 16. To do.
- the cumulative addition unit 17 again calculates g k upper by performing cumulative addition according to the above equation 37 including the updated h k , n upper and 0 corresponding h k , n upper. (Step SP 2), the updated g k upper is sent to the comparison unit 18.
- the cumulative addition unit 17 again calculates g k lower by performing cumulative addition according to the above-described equation 38 including the updated h k , n lower and 1-corresponding h k , n lower (step SP2). g k lower is sent to the comparison unit 18.
- the comparison unit 18 compares the updated g k upper and the updated g k lower , and determines from the numerical values whether or not g 1 upper ⁇ g 0 lower is apparent (step SP3). If the comparison unit can determine that g 1 upper ⁇ g 0 lower by narrowing the section, the comparison unit determines classification 0 (step SP4), and sends the determination result to the display unit.
- the comparison unit 18 compares g 0 upper and g 1 lower and clearly shows g 0 upper ⁇ g 1 lower from the numerical value. Is determined (step SP5). As shown in FIG. 5B, the comparison unit determines that it is possible to determine that g 0 upper ⁇ g 1 lower by narrowing the section (step SP6), and the determination result is displayed on the display unit. To send.
- the refinement processing unit 13 compares the numerical values of g 1 upper and g 0 lower , or compares the numerical values of g 0 upper and g 1 lower to determine the magnitude relationship between these values. Is updated to i + 1, the refinement process is executed again, and the above-described process is repeated.
- the refinement processing unit 13 updates the value of i to i + 1 until the magnitude relationship between the numerical values of g 1 upper and g 0 lower or the numerical values of g 0 upper and g 1 lower can be determined.
- the above-described processing is repeated up to L which is the maximum value of.
- the refinement processing unit 13 updates the value of i to i + 1 and repeats the above-described processing up to L, which is the maximum value, but the numerical values of g 1 upper and g 0 lower or g 0 upper and g
- the section average determination unit 27 can execute section average processing described later so that it can be finally determined whether it is either classification 1 or classification 0. Has been made.
- the section average determination unit 27 calculates g 0 pseudo , which is an updated section average when k is 0, and g 1 pseudo , which is an updated section average when k is 1, by the following equation: 39 (step SP12).
- the section average determination unit 27 determines the magnitude relationship between g 1 pseudo and g 0 pseudo (step SP13), and determines g 1 pseudo ⁇ g 0 pseudo (class SP 0). On the other hand, when determining that the magnitude relationship of g 1 pseudo ⁇ g 0 pseudo does not hold, the section average determination unit 27 determines g 0 pseudo ⁇ g 1 pseudo and determines the observation data D1 as classification 1 (step SP15). To do. In this way, the section average determination unit 27 determines whether the observation data D1 is classified into classification 0 or classification 1, and sends the determination result to the display unit to notify the user via the display unit. It is made to be able to do.
- the identification device 1, and the multiplication processing that multiplies the variable group (z k, n) to a constant log 2 e, a n 1 ..., sequentially changing a plurality of h to N k k , n upper and h k , n lower are multiplied by 2, and a plurality of h k , n upper and h k , n lower are cumulatively added to obtain g k upper and g k lower , respectively.
- the process was executed.
- the identification device 1 does not execute the calculation processing of an accurate exponential function as in the past, but from the magnitude relationship between g k upper and g k lower expressed on a binary integer ⁇ sandwiching the true value.
- the magnitude relationship between the classification 0 occurrence probability value and the classification 1 occurrence probability value it is possible to identify whether the observation data D1 is the classification 0 or the classification 1 by a simple interval calculation. It is possible to avoid the calculation processing of an exponential function having a complicated and large processing load of the occurrence probability value and the classification 1 occurrence probability value, and the processing load can be significantly reduced.
- the identification device 1 can determine at high speed whether the observation data D1 is in the classification 0 or the classification 1 by the amount that the calculation cost of the exponential function can be reduced in this way.
- the identification device 1 can identify whether the observation data D1 is either the classification 0 or the classification 1 by a simple interval calculation by improving the accuracy of the comparison operation step by step, narrowing the interval that becomes unknown, Thus, it is possible to avoid the calculation processing of the exponential function having a complicated and heavy processing load of the classification 0 occurrence probability value and the classification 1 occurrence probability value.
- the identification device 1 if the magnitudes of g k upper and g k lower are not clear even if the refinement process is repeated until i reaches the maximum value L, g 0 pseudo and g 1 which are interval averages are not obtained.
- the observation data D1 is finally obtained. Either classification 0 or classification 1 can be determined with certainty.
- white indicates a skin color
- black indicates a non-skin color
- gray indicates an undetermined region.
- the percentage of completion of the determination only by the basic identification process was 99.692% in the case of 2 mixing, 99.617% in the case of 3 mixing, and 99.613% in the case of 4 mixing. As described above, none of the operations required an exponential function in an area of 99% or more.
- step SP3 and step SP5 After determining whether or not 1 lower (step SP5), it may be determined whether or not g 1 upper ⁇ g 0 lower (step SP3).
- the identification device 1 is generally used for devices that require a function of automatically identifying an object based on observation data acquired by a sensor, such as medical / welfare equipment, disaster prevention / monitoring equipment, and automobile / industrial equipment. it can.
- the wireless sensor network is used to construct a network by innumerable small devices called sensor nodes equipped with sensors, signal processing functions, pattern recognition functions, wireless communication functions, batteries (or independent power generation devices). Can be expected.
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Complex Calculations (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
前記hk,n upperと前記hk,n lowerとを用いて、次式
を計算する累積加算部と、前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較部とを備えることを特徴とする。
を記憶した記憶部を備え、前記gk upper及び前記gk lowerを比較して大小関係を判定できないとき、iの値をi+1に更新し、zk,nlog2eの小数部の小数点以下第i位が1であるか、或いは0であるかを判断し、前記小数点以下第i位が1のとき、前記数13のB[i]を前記hk,n upperに乗算しhk,n upperを更新して前記gk upperを計算し、前記小数点以下第i位が0のとき、前記数14のB[i]-1を前記hk,n lowerに乗算しhk,n lowerを更新して前記gk lowerを計算することを特徴とする。
g1 pseudo<g0 pseudoと、g0 pseudo<g1 pseudoとによって前記観測データを分類する平均値処理部を備えることを特徴とする。
を計算する累積加算ステップと、前記累積加算ステップによって算出した前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較ステップとを備えることを特徴とする。
を前記hk,n upperに乗算しhk,n upperを更新して前記gk upperを計算し、前記小数点以下第i位が0のとき、次式
g1 pseudo<g0 pseudoと、g0 pseudo<g1 pseudoとによって前記観測データを分類する平均値処理ステップを備えることを特徴とする。
を計算する累積加算ステップと、前記累積加算ステップによって算出した前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較ステップとをコンピュータに実行させることを特徴とする。
前記小数点以下第i位が1のとき、次式
を前記hk,n upperに乗算しhk,n upperを更新して前記gk upperを計算し、前記小数点以下第i位が0のとき、次式
g1 pseudo<g0 pseudoと、g0 pseudo<g1 pseudoとによって前記観測データを分類する平均値処理ステップを備えることを特徴とする。
図1において、1は本発明による識別装置を示し、観測データD1の分布が混合正規分布に従うと仮定し、混合正規分布パラメータを基に観測データD1を分類0又は分類1のいずれかであるかを識別し得るものである。
(2-1)基本識別処理
ここで識別部5により実行される識別処理は、基本識別処理と洗練化処理とがあり、先ず始めに、図3及び図4を用いて識別処理のうち基本識別処理について以下説明する。
ここで、比較部18は、上述した基本識別処理によって、g1 upper及びg0 lowerの数値を比較し、或いはg0 upper及びg1 lowerの数値を比較し、図6(A)に示すように、これらの大小関係を判定できなかったときには、その判定不可結果を示す再処理信号を洗練化処理部13に送出して、洗練化処理を実行し得るようになされている。
ここで、洗練化処理部13のB[i]選択部14は、上述した数34により分割部12で算出された小数部群(βk,n)の各小数部について、それぞれ個別に小数点以下第i位が1であるか否かを判断する(ステップSP9)。ここで、B[i]選択部14は、小数部群(βk,n)のうち、1小数部が存在すると判断すると、その1小数部の処理時に記憶部20からB[i]を読み出して乗算部21へ送出する。
一方、B[i]選択部14は、小数部群(βk,n)のうち、0小数部が存在すると判断すると、その0小数部の処理時にB[i]-1を乗算部21へ送出する。
累積加算部17は、更新hk,n upper及び0対応hk,n upperも含めて、上述した数37によって累積加算してgk upperを再び算出し(ステップSP2)、この更新したgk upperを比較部18に送出する。
以上の構成において、識別装置1では、変数群(zk,n)に定数log2eを乗算する乗算処理と、nを1…,Nkまで順次変えて複数のhk,n upperとhk,n lowerを算出する2のべき乗倍処理と、複数のhk,n upperとhk,n lowerとをそれぞれ累積加算してgk upper及びgk lowerを求める加算処理とを実行するようにした。
本発明による識別装置1を画像の肌色識別に応用した例を示す。この実施例では、図7(A)に示すような人物が撮像された画像の肌色識別をするために、上述した基本識別処理のみを行った。その結果、正規分布が2混合のときには図7(B)に示すような識別結果を得た。また、正規分布が3混合のときには、図7(C)に示すような識別結果を得、正規分布が4混合のときには、図7(D)に示すような識別結果を得た。
Claims (9)
- 観測データの分布が混合正規分布に従うと仮定し、混合正規分布パラメータを基に前記観測データを分類する識別装置であって、
複数の前記観測データの各特徴ベクトルを基に得た変数群zk,n(kは分類を表す1又は0、nは各分類に仮定される正規分布の分布番号を示す)と、前記混合正規分布パラメータを基に得た定数群Kk,nとを用いて、次式
(「zk,nlog2e」はzk,nlog2eの整数部である)
を計算する2のべき乗倍部と、
前記hk,n upperと前記hk,n lowerとを用いて、次式
(Nkは分類kの正規分布の混合数を示す)
を計算する累積加算部と、
前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較部と
を備えることを特徴とする識別装置。 - 次式
(iは0,1…,L、ここでLは任意に設定した正の整数)
を記憶した記憶部を備え、
前記gk upper及び前記gk lowerを比較して大小関係を判定できないとき、前記iの値をi+1に更新し、zk,nlog2eの小数部の小数点以下第i位が1であるか、或いは0であるかを判断し、
前記小数点以下第i位が1のとき、前記数5のB[i]を前記hk,n upperに乗算しhk,n upperを更新して前記gk upperを計算し、
前記小数点以下第i位が0のとき、前記数6のB[i]-1を前記hk,n lowerに乗算しhk,n lowerを更新して前記gk lowerを計算する
ことを特徴とする請求項1記載の識別装置。 - 観測データの分布が混合正規分布に従うと仮定し、混合正規分布パラメータを基に前記観測データを分類する識別方法であって、
複数の前記観測データの各特徴ベクトルを基に得た変数群zk,n(kは分類を表す1又は0、nは各分類に仮定される正規分布の分布番号を示す)と、前記混合正規分布パラメータを基に得た定数群Kk,nとを用いて、次式
(「zk,nlog2e」はzk,nlog2eの整数部である)
を計算する2のべき乗倍ステップと、
前記2のべき乗倍ステップで算出した前記hk,n upperと前記hk,n lowerとを用いて、次式
(Nkは分類kの正規分布の混合数を示す)
を計算する累積加算ステップと、
前記累積加算ステップによって算出した前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較ステップと
を備えることを特徴とする識別方法。 - 観測データの分布が混合正規分布に従うと仮定し、混合正規分布パラメータを基に前記観測データを分類する識別処理プログラムであって、
複数の前記観測データの各特徴ベクトルを基に得た変数群zk,n(kは分類を表す1又は0、nは各分類に仮定される正規分布の分布番号を示す)と、前記混合正規分布パラメータを基に得た定数群Kk,nとを用いて、次式
(「zk,nlog2e」はzk,nlog2eの整数部である)
を計算する2のべき乗倍ステップと、
前記2のべき乗倍ステップで算出した前記hk,n upperと前記hk,n lowerとを用いて、次式
(Nkは分類kの正規分布の混合数を示す)
を計算する累積加算ステップと、
前記累積加算ステップによって算出した前記gk upper及び前記gk lowerを比較して、g1 upper≦g0 lowerと、g0 upper≦g1 lowerとによって前記観測データを分類する比較ステップと
をコンピュータに実行させることを特徴とする識別処理プログラム。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2009550502A JP5112454B2 (ja) | 2008-01-23 | 2009-01-16 | 識別装置、識別方法及び識別処理プログラム |
US12/811,006 US8321368B2 (en) | 2008-01-23 | 2009-01-16 | Identification device, identification method, and identification processing program |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2008012678 | 2008-01-23 | ||
JP2008-012678 | 2008-01-23 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2009093525A1 true WO2009093525A1 (ja) | 2009-07-30 |
Family
ID=40901033
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2009/050529 WO2009093525A1 (ja) | 2008-01-23 | 2009-01-16 | 識別装置、識別方法及び識別処理プログラム |
Country Status (3)
Country | Link |
---|---|
US (1) | US8321368B2 (ja) |
JP (1) | JP5112454B2 (ja) |
WO (1) | WO2009093525A1 (ja) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2015156894A (ja) * | 2014-02-21 | 2015-09-03 | 学校法人日本大学 | 医用画像処理装置、その医用目的領域抽出方法及び医用目的領域抽出処理プログラム |
JP2016021880A (ja) * | 2014-07-17 | 2016-02-08 | 国立大学法人 新潟大学 | 抗菌効果判定システム、抗菌効果判定方法及び抗菌効果判定プログラム |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2003187229A (ja) * | 2001-12-14 | 2003-07-04 | Nec Corp | 顔メタデータ生成方法および装置、並びに顔類似度算出方法および装置 |
JP2005267570A (ja) * | 2004-03-22 | 2005-09-29 | Mitsubishi Electric Corp | 車両認識方法 |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7080011B2 (en) * | 2000-08-04 | 2006-07-18 | International Business Machines Corporation | Speech label accelerators and techniques for using same |
US6785645B2 (en) * | 2001-11-29 | 2004-08-31 | Microsoft Corporation | Real-time speech and music classifier |
KR100568237B1 (ko) * | 2004-06-10 | 2006-04-07 | 삼성전자주식회사 | 비디오 영상으로부터 이동 물체를 추출하는 장치 및 방법 |
US7756341B2 (en) * | 2005-06-30 | 2010-07-13 | Xerox Corporation | Generic visual categorization method and system |
US8107541B2 (en) * | 2006-11-07 | 2012-01-31 | Mitsubishi Electric Research Laboratories, Inc. | Method and system for video segmentation |
US7933454B2 (en) * | 2007-06-25 | 2011-04-26 | Xerox Corporation | Class-based image enhancement system |
-
2009
- 2009-01-16 JP JP2009550502A patent/JP5112454B2/ja not_active Expired - Fee Related
- 2009-01-16 US US12/811,006 patent/US8321368B2/en not_active Expired - Fee Related
- 2009-01-16 WO PCT/JP2009/050529 patent/WO2009093525A1/ja active Application Filing
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2003187229A (ja) * | 2001-12-14 | 2003-07-04 | Nec Corp | 顔メタデータ生成方法および装置、並びに顔類似度算出方法および装置 |
JP2005267570A (ja) * | 2004-03-22 | 2005-09-29 | Mitsubishi Electric Corp | 車両認識方法 |
Non-Patent Citations (1)
Title |
---|
SHOGO MURAMATSU ET AL.: "Kukan Keisan ni yoru Kongo Gauss Bunpu Model no Kokateki Hanbetsuho", IEICE TECHNICAL REPORT, vol. 107, no. 475, 24 January 2008 (2008-01-24), pages 55 - 60 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2015156894A (ja) * | 2014-02-21 | 2015-09-03 | 学校法人日本大学 | 医用画像処理装置、その医用目的領域抽出方法及び医用目的領域抽出処理プログラム |
JP2016021880A (ja) * | 2014-07-17 | 2016-02-08 | 国立大学法人 新潟大学 | 抗菌効果判定システム、抗菌効果判定方法及び抗菌効果判定プログラム |
Also Published As
Publication number | Publication date |
---|---|
US8321368B2 (en) | 2012-11-27 |
US20100287133A1 (en) | 2010-11-11 |
JP5112454B2 (ja) | 2013-01-09 |
JPWO2009093525A1 (ja) | 2011-05-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022199583A1 (zh) | 图像处理方法、装置、计算机设备和存储介质 | |
CN114186632B (zh) | 关键点检测模型的训练方法、装置、设备、存储介质 | |
CN110197716B (zh) | 医学影像的处理方法、装置及计算机可读存储介质 | |
CN108230233A (zh) | 数据增强、处理方法和装置、电子设备和计算机存储介质 | |
EP2927864A1 (en) | Image processing device and image processing method | |
US7136540B2 (en) | Picture region extraction method and device | |
CN113988357B (zh) | 基于深度学习的高层建筑风致响应预测方法及装置 | |
CN111583274A (zh) | 图像分割方法、装置、计算机可读存储介质及电子设备 | |
CN111382791B (zh) | 深度学习任务处理方法、图像识别任务处理方法和装置 | |
CN112330671A (zh) | 细胞分布状态的分析方法、装置、计算机设备和存储介质 | |
EP4343616A1 (en) | Image classification method, model training method, device, storage medium, and computer program | |
JP5112454B2 (ja) | 識別装置、識別方法及び識別処理プログラム | |
CN113158860B (zh) | 基于深度学习多维度输出人脸质量评估方法与电子设备 | |
CN112070162A (zh) | 多类别处理任务训练样本构建方法、设备及介质 | |
CN116993577A (zh) | 图像处理方法、装置、终端设备以及存储介质 | |
CN113780555B (zh) | 基于数据增强的模型训练方法、装置、设备及存储介质 | |
CN115439916A (zh) | 面部识别方法、装置、设备及介质 | |
CN115712763A (zh) | 基于人工智能和大数据的网络数据处理方法和平台 | |
CN110889635B (zh) | 一种用于对食品安全事件的处理进行应急演练的方法 | |
CN114693697A (zh) | 图像处理方法、装置、设备及计算机可读存储介质 | |
WO2020040007A1 (ja) | 学習装置、学習方法及び学習プログラム | |
CN113643219B (zh) | 一种基于三光融合的图像成像方法及装置 | |
Merzougui | Multi-measures fusion based on multi-objective genetic programming for full-reference image quality assessment | |
CN115035395B (zh) | 用于机场航站楼场景的安全分析装置方法及装置 | |
CN115408552B (zh) | 显示调整方法、装置、设备及存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 09704788 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2009550502 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 12811006 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 09704788 Country of ref document: EP Kind code of ref document: A1 |