JP2005214682A - Discrimination system of object to be detected, and discrimination system of image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a deciding and comfirming processing method, constituted so as to enhance the decision and confirmation rate of an object to be detected, enabling automatic processing and which is capable of performing real-time processing and which is a simple algorithm. <P>SOLUTION: This discrimination system of an object to be detected is equipped with a Maharanobis space calculating means for forming a Maharanobis space (reference space), with respect to each of a plurality of the objects to be detected, on the basis of the known characteristic parameters of a plurality of the objects to be detected and applying the characteristic parameters, extracted from an unknown object to be detected, to the Maharanobis space formed by the Maharanobis space calculating means for calculating the Maharanobis space of the unknown object to be detected, a comparison means for performing the comparison with the Maharanobis distance calculated by the Maharanobis distance operating means and a discrimination means for discriminating the object to be detected, on the basis of the comparison result in the comparison means. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a method and apparatus for discriminating / determining / recognizing identity by comparing an object to be detected with a known object, and in particular, a Mahalanobis space created from feature parameters extracted from information on the object to be detected. The present invention relates to a determination / judgment / recognition method and apparatus capable of changing the determination criterion of identity according to the purpose by comparing information between a detection object and a known detection object.

  Various methods for identifying and recognizing detected objects and known detected objects have been proposed. As an example, a method for detecting and distinguishing detected objects from image data using a neural network technique is one example. This will be described below.

  Patent Document 1 includes an imaging unit that captures an original image and a detection processing unit. The detection processing unit further includes a quantization unit that digitizes pixel brightness, a storage unit, a processing region setting unit, and a detection unit. An extraction unit for pixels in the processing region, a classification unit that forms a histogram of each extracted pixel based on lightness, and an operation that quantifies the characteristics of the pixel distribution state in the obtained histogram using a statistical method And a processing unit that performs learning processing using the neural network technique on the numerical values obtained by the calculation unit, and the processing unit finally determines whether or not an object exists in each detection region. The judgment / recognition technique to be performed is disclosed.

  Patent Document 2 includes an imaging unit that captures particle components present in urine and an image processing unit that analyzes the captured particle image and classifies and displays the particles. The imaging unit captures a color particle image, and the image processing unit converts the color particle image into at least one binary signal based on a color reference, and the feature amount of each particle from the binary image signal , And automatically classifying each particle from the feature amount, and displaying each color particle image according to the classification result is disclosed.

  The image processing unit includes a function of storing a color particle image and an automatic classification result, and a function of storing the color particle image and the automatic classification result in association with each other. In addition, a function of storing either the color particle image or the automatic classification result and any data related to the specimen input by the operator is provided.

  The image processing unit has a function of storing both the color particle image and the automatic classification result and arbitrary data related to the specimen input by the operator, and the image processing unit includes at least one or more It has a function of displaying color particle images on the same screen according to a predetermined sequence of automatic classification results. The image processing unit has a function of displaying a color particle image and an automatic classification result in correspondence with each other on the same screen. Based on the display of the color particle image and the automatic classification result, reclassification / learning can be performed. A neural network method is used as the reclassification / learning identification method, and the reclassification / learning can be performed manually.

  Patent Document 3 discloses a method for detecting and classifying a pattern defect on a semiconductor wafer. Means for extracting an image feature of a defect image; means for calculating a coordinate feature amount from a defect coordinate in a coordinate system provided on a substrate of a defect corresponding to the image; and the image for a defect whose classification result is known in advance. Means for learning feature quantities and coordinate feature quantities, and discriminating / classifying unknown defects based on the learning results. As a method for discriminating and classifying, first, teaching data for each defect consisting of multiple variables (features) is learned (created), a boundary line (equal from the center of gravity of each group) is defined between each defect, and a discriminant function Judge and classify at. As this discriminant function, Mahalanobis distance is used, and it is performed by the difference of Mahalanobis distance between the detected object and the known learning data (Patent Document 3, FIG. 4).

  Further, in recent years, as disclosed in the above-mentioned Patent Document 3, based on multivariate data extracted from two or more groups (populations), an unknown sample is discriminated as one of the groups. Therefore, the discriminant function is defined and used by Mahalanobis distance considering the correlation between multivariate. On the other hand, as a new method using this Mahalanobis distance, there is a method disclosed in “Journal of Quality Engineering, Vol. 3, No. 1: {Comprehensive evaluation and multi-dimensional information and SN ratio}”. This method is applied to medical examinations in the medical field and uses the Mahalanobis space as a measure. In this case, the Mahalanobis space is a group of healthy people, and only healthy people make a uniform group, and the Mahalanobis space (group of healthy people's Mahalanobis distance) is the reference point with medically determined measurement items. . The average distance of the people belonging to the Mahalanobis space is 1.0, and the Mahalanobis distance of the people belonging to the space takes only a value of 2 to 4. Then, the Mahalanobis distance with respect to a subject whose health is unknown is examined, and if the value is smaller than a certain threshold value (for example, 4), it is determined to be healthy, otherwise it is determined to be abnormal. Such a method has a wide range of applications and has been published in various fields. For example, Patent Document 4: Environment monitoring system, abnormality detection method, and abnormality detection device.

Japanese Patent Laid-Open No. 9-102032 JP-A-7-55688 JP 2000-200366 A Japanese Patent No. 3280872

  In the object detection method described in Patent Document 1, the brightness of each pixel in the detection processing area set in the image is histogrammed, and the characteristics of the pixel distribution state are obtained using a statistical method on the histogram obtained thereby. This is a method of digitizing and determining whether or not any object is present in the detection processing region based on the numerical value obtained thereby. The determination method is obtained by using a numerical value obtained by a statistical method as input data, using actual data in a detection processing region from which the input data is obtained as teacher data, and executing a predetermined operation on the input data. The obtained output data is compared with the teacher data, and the learning is performed and the processing is performed based on the neural network method so that the output data matches the teacher data.

  For this reason, when new teacher data is added, the neural network must be reconstructed again, which requires a great amount of man-hours.

  In the particle classification display device described in Patent Document 2, the characteristic parameter of the particle is detected, automatically determined for each particle, the classification result and the particle image are stored and displayed, and the parameter is changed. Is easy, and learning data can be created and re-learned, which greatly contributes to labor saving. However, when new teacher data is added to the processing unit for executing re-learning or automatic classification, a neural network must be reconstructed as in the conventional example. Therefore, the problem that a great number of man-hours is required has not been solved.

  Further, in the detection of the detected object and the classification of particles in urine, the aim is to improve a single function such as a detection function and a classification function, and there is an immediate improvement effect. However, using the results as one piece of information, it is more accurate and more comprehensively captures the “property” of the object and its “state change” more accurately, and the “property” and “ From the viewpoint of judging and recognizing the “state change”, significant improvement is desired.

  Various results have been disclosed in the discriminant function using the Mahalanobis distance with the already learned data described in Patent Document 3 and Patent Document 4, or the discrimination / judgment / recognition based on the Mahalanobis distance from the uniform Mahalanobis space. Wide application range. However, when this method is applied to, for example, classification or discrimination of red blood cells, white blood cells, and glass cell particles in urine, the following problems have occurred.

  There are many known perceived populations. Both populations are candidates for creating the Mahalanobis space that serves as the basis for the measure. For example, the urine particles may be any of red blood cells, white blood cells, and glass particle cells. If the population of red blood cells is a standard Mahalanobis space (standard space), particles such as white blood cells and glass particle cells are assigned to this space, and the white blood cells and glass particle cells are classified and discriminated at the Mahalanobis distance. The At this time, of course, a clear discrimination / classification must be made between the leukocyte group and the vitreous cell group. In addition, when white blood cells are used as a reference space, the same result must be obtained. However, when a particle group having a large variation is used as a reference space, the classification and discrimination results deteriorate and accuracy decreases.

  Despite the fact that there are many known recognized populations, the variance, covariance, and correlation coefficient are assumed to be common to each population (because it is the origin of distance calculation) ). For this reason, in order to improve the ability of classification / discrimination / recognition, the population data of the detected objects has small variations individually, and only the sample whose average value greatly changes becomes the learning data. Therefore, only specific limited particles are discriminated / recognized, but generally the recognition rate is lowered.

  The present invention has been made to solve the problems of the prior art, and the Mahalanobis distance information created from the feature parameters extracted from the information of the detected object is compared with that of a plurality of known detected objects. It is a processing method that can improve the judgment / recognition rate of the detected object by determining / recognizing which object is the same, and can be processed automatically, and is a simple algorithm. An object of the present invention is to provide a determination / recognition processing method and processing apparatus capable of real-time processing.

  In addition, the processing method can easily change the parameters for the determination and recognition, can separately create learning data and re-learn, and can improve the accuracy of determination and recognition of the detected object as needed. And an apparatus.

  It is another object of the present invention to provide a system capable of integrating information of a single detection device to create information with higher accuracy and high added value.

  The configuration of the present invention for achieving the above object is as follows.

  Feature parameter extraction means for extracting multivariate feature parameters from a plurality of known subjects, and a Mahalanobis space (reference space) formed by each of the plurality of subjects based on the feature parameters extracted by the feature parameter extraction means Mahalanobis space computing means, storage means for preliminarily storing information of a plurality of reference spaces formed, feature parameter extraction means for extracting feature parameters from an unknown detector, and feature parameter extraction means The Mahalanobis space calculated by the Mahalanobis space calculation means, the Mahalanobis distance calculation means for calculating the Mahalanobis distance of the unknown detector, and the Mahalanobis distance calculation means Comparison means for comparison and comparison result by the comparison means Detection object discrimination system comprising determining means for discriminating the object to be detected, a based on.

As described in detail above, according to the configuration of the present invention,
(1) The image processing unit captures images of a large number of detected objects including a plurality of known individual differences, creates a large number of feature parameters and the Mahalanobis space from the images, and responds to an input of an unknown detected object. Thus, it is automatically determined / recognized whether the unknown object to be detected is “different” or “same” from the known object at the separation distance from the Mahalanobis space. For this reason, manpower is not required, and the operations of the operator or the specialist can be reduced.
(2) The discrimination / recognition ability of the detected object can be improved and the accuracy is improved, so that more comprehensive discrimination / recognition can be performed.
(3) By storing the determination / recognition result together with the image data, it is possible to easily create teaching data, and to easily change the parameters, and to improve the accuracy of determination / recognition as needed. Can do. For this reason, it is possible to save labor for operators and specialists. Further, it can be configured at low cost.
(4) When a plurality of Mahalanobis spaces and their distances are used for the discrimination / recognition logic, the error can be minimized even for the detection object samples near the boundary between the detection objects. For this reason, accuracy is improved and reliability is improved.
(5) Even when characteristic parameters are changed or new categorization is added, it can be performed simply by adding or modifying the Mahalanobis spatial information, and there is no need to reconstruct the discrimination / recognition network. It is possible to significantly reduce the man-hours for raising.
(6) Since the management device collects results and information from a single device and results and information from other devices via a network, it can collect characteristic parameters necessary for more comprehensive judgment and recognition. Moreover, the Mahalanobis space can be created from these characteristic parameters, and the property of the detected object and its temporal state change are more accurately and subdivided by the difference in the distance between the detected object and the Mahalanobis space. It is also possible to distinguish and recognize property changes. Therefore, it is possible to provide information with high accuracy and high added value that can contribute to comprehensive judgment and recognition, and diagnosis support can be achieved.

  In order to facilitate understanding, the above configuration will be described below with an example of recognizing an object to be detected by image processing.

  An imaging unit for imaging the object to be detected; an extraction unit for extracting a number of items (length, area, density, etc.) expressing the characteristics of the object to be detected by digitizing the captured image data; and A plurality of Mahalanobis spaces are formed from a large number of feature parameters of an existing or known learned object group from the feature parameters of the object, and the Mahalanobis distance from the known space is calculated for the input of the detected object, and the separation distance is calculated. Then, the degree of similarity or coincidence (homogeneity) with the detected object is calculated, and as a result, whether or not the unknown detected object belongs to any existing or known group (population) or category. It comprises a processing unit for discriminating and recognizing and an output unit for disclosing or displaying the result. The component is configured as a detection device, and the result or information processed by the processing unit is an upper management device. The detector is provided with a communication unit for transmission, and the result and information thereof can be disclosed. The management device performs the same processing as the detection device from the result and information from the detector and other detection devices. It was set as the structure which comprises a process part.

The procedure of the processing unit of this discrimination / recognition method is shown below.
(1) Collection of learning data forming eigenspaces (reference spaces) of a plurality of known detection objects First, n detection target group samples (1) belonging to the known or existing homogeneous or the same group ˜m groups), n pieces of data are collected for each group with k characteristic parameters (for example, in the case of shape: length, area, density, etc.). As described above, these data groups are known or predetermined recognized objects, and are n data groups that take into account individual differences (variations). Basic statistics such as an average value and standard deviation for each feature parameter (k) are also calculated at the same time, and these pieces of information are stored in the storage unit of the device.
(2) Creation of Mahalanobis space data of “the eigenspace of the detected object” in (1), (1) From the data group, normalize the data group with an average value and standard deviation for each feature parameter; A Mahalanobis distance: D ² of “the eigenspace of the detected object” is calculated. The Mahalanobis distance is calculated according to the correlation coefficient or variance-covariance matrix of each feature element number (k) and the number of data used (n). The average value is D ² ≈1.0, and the standard deviation is a space (Mahalanobis space) exhibiting a χ2 distribution of about 1 to 2, and is created for each group of detected objects. At the same time, basic statistics such as an average value and standard deviation for each Mahalanobis space (m pieces) are calculated, and these pieces of information are stored in the storage unit of the device.
(3) Processing method Data of an unknown detected object or a detected object to be discriminated / recognized is photographed, and feature parameters (x = {x1, x2,..., Xk}) are extracted. The extracted feature parameter data is processed by the same method as the above-mentioned items (1) and (2), and the data is processed by the same method as the above-mentioned items (1) and (2). Address the eigenspace (Mahalanobis space) of the detection object formed in the items 1) and (2), and calculate its isolation distance (Mahalanobis distance): D ² (x). That is, the D ² (x) and the reference space value D ² (0) (Mahalanobis space) which is the “eigenspace of the detection object” are sequentially compared, or one “ “Eigenspace” is fixed (reference), and a comparison is made based on the difference distance from that space.

In the case where one of the above-mentioned “eigenspaces of the detection object” is fixed (reference), when there is no difference in the spatial distance between D ² (x) and D ² (0), that is, D ² (x) = 0. If it is within the range of ˜4, it is determined and recognized that the detected object is the same (“same quality”) as the “detected object”. On the other hand, if there is a difference, the detected object is judged and recognized as being different from the “detected object” (“different”).

Or, when D ² (0) (Mahalanobis space distance) of the reference space of “the eigenspace of the detected object” is m,
These values are sequentially compared, and it is determined and recognized as being “homogeneous” in the reference space of “the eigenspace of the detected object” having the minimum distance. Therefore, an unknown object to be detected is categorized by determining and recognizing whether it belongs to the same space or “heterogeneous” of the eigenspace of the detected object.
(4) Processing method at the time of reconstruction In the above item (3), it corresponds to any object to be detected that cannot be distinguished / recognized as “homogeneous” or “heterogeneous”, that is, the eigenspace of the detected object. For the detection object that does not, the image information and feature parameters are sequentially stored in the storage unit and accumulated, and a data group is collected. Such a data group is re-learned at a predetermined interval or as necessary to create a new group or categorization.

  The processing method is the same as the above-mentioned items (1) and (2). However, since the data group of the feature parameters has already been collected, a Mahalanobis space is newly created, and the statistics and features of the space are created. The process is completed only by the process of storing only the parameter statistics in the storage unit.

  Further, the management device and the device are connected via a network, and the management device has a function of collecting results of the device and results and information of other detection devices. As in the case of the device, the Mahalanobis space, which is the eigenspace of the detected object, is created from various characteristic parameters of a plurality of items, and the processing method similar to the above item (3) is provided. Judging and recognizing changes in body condition and time based on the Mahalanobis distance.

  The functions of the above technical means are as follows.

  According to the configuration of the present invention, the image processing unit captures images of a large number of detected objects including known individual differences, obtains a large number of feature parameters from the images, and what the detected object is based on the feature parameters. Create a Mahalanobis space for each detected object, and automatically determine whether it is “foreign” or “homogeneous” with the known detected object at the separation distance of the Mahalanobis distance from the detected object. ·recognize.

  In addition, even when a new object to be detected other than known data stored in the storage unit is re-learned and added, a Mahalanobis space is created from the image data and feature parameters, and is added to and stored in the storage unit. Group and categorization can be increased in a short time, and discrimination / recognition performance is improved.

  In addition, when the Mahalanobis space and its distance are used for the discrimination / recognition logic, the error can be minimized with respect to an unknown detector sample near the boundary between the detectors. For this reason, the accuracy of discrimination / recognition is improved.

  Furthermore, as described above, even when the characteristic parameter is changed or when a new categorization is added, it is possible to cope with it simply by adding or correcting the information of the Mahalanobis space. There is no need to build.

  Further, since the management device collects results and information from a single device and results and information from other devices via a network, it can collect characteristic parameters necessary for more comprehensive discrimination and recognition. In addition, the Mahalanobis space can be easily created from these characteristic parameters, and the “property” of the detected object and its temporal “state change” can also be determined based on the difference in Mahalanobis distance between the detected object and the Mahalanobis space. ·recognize.

  The present invention will be described below with reference to examples.

  FIG. 1 is a block diagram showing the configuration of a detection device according to an embodiment of the present invention and its configuration diagram, FIG. 2 is an explanatory diagram of a photographing unit of the detection device according to the embodiment of FIG. 1, and FIG. FIG. 4 is a diagram illustrating a basic flow of a processing unit according to the first embodiment, FIG. 4 is a diagram illustrating a processing flow in the detection device according to the embodiment of FIGS. 1 and 3, and FIG. 5 is according to the embodiment of FIG. FIG. 6 is a diagram for explaining in detail the discrimination / recognition processing method in the detection device, and FIG. 6 is an example of discrimination / recognition in the detection device according to the embodiment of FIG.

  1 and 2, the discrimination / recognition method according to an embodiment of the present invention is a device that automatically discriminates / recognizes / classifies particles such as red blood cells, white blood cells, glass cell particles, and cancer cell particles in urine. A case where this is applied to will be described as an example.

  The detection apparatus according to the embodiment of the present invention includes an imaging unit: 1 that captures an object to be detected: 8, and an image that digitalizes (pixels) a video signal from an original image captured by the imaging unit: 1. Extraction unit: 5, feature parameter calculation unit for extracting feature items such as the shape, area, and density of the detection object from the pixel, storage unit for storing the feature parameter: 4, and detection from the feature parameter The processing unit 1: 2 is configured to recognize and recognize the body.

The detected object 8 is captured by the photographing unit 1 by a CCD camera or the like and converted into an electric signal (see FIG. 2). The converted image signal is converted into a digital signal by the AD conversion circuit and stored in the memory of the device. For example, when an 8-bit A / D conversion element is used, each pixel is digitally quantized to 256 (= 2 ⁸ ) gradations.

  From the quantized image of the detected object 8, feature parameters such as area, perimeter, and average RGB density are extracted for each detected object by the feature parameter calculation unit 6 (see FIG. 2). ). The extracted feature parameters are input to the processing unit 1: 2.

The photographing unit: 1, the image extracting unit: 5, the characteristic parameter calculating unit: 6, and the processing unit 1: 2
Controlled by a CPU, a memory (not shown), a storage unit: 4 and an output unit: 3 for outputting or displaying the result are connected to the CPU via a CPU bus.

  Further, the CPU can communicate with a higher-level management device: 110 via a communication unit: 21 and a network: 130, and the result obtained by the detection device: 100 is used as the management device: 110. The result and information can be transmitted.

  Next, a processing method applied to the processing unit 1: 2 will be described with reference to FIG. This figure shows a conventional procedure for discriminating, classifying, and recognizing an object to be detected based on a single reference space. Mahalanobis distance is a measure of distance in multivariate (multidimensional) space. Conventionally, it is used in judgment analysis of multivariate analysis, is a pattern recognition method for prediction and diagnosis, and is an index that can measure the normality of the state of things. Unlike the conventional multivariate analysis, this is a scale indicating how far away or similar to the reference multivariate data group.

  For example, the starting point (reference) of the determination is a normal product (for example, a group of acceptable products in the inspection test, which is G1 in FIG. 3). In general judgment analysis and multiple regression analysis, a scale cannot be created unless information on abnormalities is given. However, as the abnormalities are diverse, not all of them can be considered. There is only. However, by using the Mahalanobis distance, the separation distance can be created just from the reference, so it can be seen how much it differs from the reference.

FIG. 3 shows the process.
(1): The eigenspace (reference space) is selected and feature parameters are selected.

For example, a group of detectors having known properties is defined as the space, and a large amount of data is collected.
(2): Create a database (Mahalanobis space).

  If the number of data is n and the number of feature parameters is Tk, the average, standard deviation, and correlation coefficient for each item are created. The average value for each item is μ (1), μ (2),..., Μ (k), the standard deviation is σ (1), σ (2),. Normalize.

Y (i, j) = {X (i, j) -μ (j)} / σ (j) i = 1 to n, j = 1 to k
(3): Next, a correlation matrix between feature parameters: ρ (i, j) (or covariance) is obtained, and its inverse matrix a (i, j) is obtained.
(4): The Mahalanobis distance (D ² ) is obtained.

D ² (i) = (1 / k) * Σy (i, jk) * a (jk, jk) * y (jk, i)
i = 1 to n, jk = 1 to k
This distance shows a distribution with an average of about 1.0 with respect to the total number of data n, and this space (Mahalanobis space) becomes a zero point, which is a unit amount of a scale as a total measurement.
(5): The distance of the detection body is obtained.

D ² (x) is calculated for observation data X (i, j) of an arbitrary unknown object. As shown in FIG. 3, since the Mahalanobis distance of the subject 1 is located in the reference space (defined as G1), its property is discriminated / recognized as “homogeneous” and is the same as the space at the time of creating the reference space. Identify / recognize that it is exhibiting properties. On the other hand, since the objects 2 and 3 are isolated from the reference space, they are determined and recognized as “foreign”. The degree of heterogeneity between the subject 2 and the subject 3 is different from that of the reference space in the subject 3. As this distance increases, the degree of heterogeneity also increases.

  This will be described with reference to this embodiment. G1 is a group of red blood cells among particles in urine. It is determined and recognized that the subject 1 is a red blood cell, the subject 2 is a white blood cell, and the subject 3 is a glass cell particle. At this time, the subjects 2 and 3 have large variation in characteristic parameters as compared with the red blood cells which are the reference space. For this reason, as shown in FIG. 3, it is ideal that the number of white blood cells (G2) and the vitreous cell particles (G3) are categorized as the number increases. However, as described above, since there are many variations in the feature values, it is not always possible to completely separate the G1 and G2 groups. For this reason, a recognition rate falls. This is because, as described above, it is assumed that the pattern constituting the reference space of the reference red blood cells is universal. On the other hand, in order to solve this problem, it is only necessary to use data obtained by reducing the variation of the detected object, but this limits the detected object, resulting in an increase in unknown particles and recognition. The rate will drop. The purpose of this application is to solve the above-mentioned problems when applying the Mahalanobis distance and space, and to improve the recognition rate and the recognition method that is more efficient than the neural network method. .

  FIG. 4 shows the processing flow of the present invention. This will be described with reference to FIG.

S1 (Step 1): Setting of a known detected object space First, as shown in FIG.
90, 91, 92 (1 to m group) or known categorized detector groups: 90 (for example, red blood cells), 91 (for example, white blood cells), 92 (for example, vitreous cell particles) (1 to m groups) Then, m Mahalanobis spaces are created by the procedure described above (see FIG. 3), and the average value, standard deviation, correlation matrix or the characteristic parameter (k) of each detected object group is stored in the storage unit 4. The inverse matrix of the covariance matrix, the average value of the Mahalanobis distance, and its standard deviation are stored (part 7 in FIG. 1). At this point, categorized detected object groups (1 to m groups) of 1 to m groups of detected objects are created.

S2 (Step 2): Data collection of unknown detection object Unknown detection object: When 8 (red blood cells, white blood cells, glass cell particles, m other particles present in urine) is input, as described above Then, one image of the detected object is captured by the CCD camera of the photographing unit: 1 and binarized by the image extraction unit: 5 and the feature parameter calculation unit: 6 to extract feature parameters (k).

S3 (Step 3): The characteristic parameter obtained in the processing method step: 2 is applied to the information constituting the space of each group stored in the storage unit: 4, and the Mahalanobis distance: D ² (x) is calculated. . The detected object is discriminated / recognized at the calculated distance. In the process 1:41, it is discriminated / recognized based on the Mahalanobis distance whether it is the same or different from any of the groups 1 to m, or it is discriminated / recognized whether any of the existing groups is different. In the example shown in the figure, the detected object has a category “◎”.
(For example, white blood cells) "(step 3-1).

  On the other hand, if any of the different categories “? (Does not belong to m known categories)”, that is, if there is no corresponding and discrimination / recognition is impossible, the characteristic parameter information: 93, 94, 95 and image data Is stored in the storage unit 4 (step 3-2).

  When there are a large number of detected objects to be observed, the above steps 2 to 3 are repeated until there are no detected objects. This processing flow stops after confirming the point in time when the object to be detected disappears.

S4 (Step 4): Output processing Based on the result of Step 3, the result is output to the CRT display or the like of the output unit 3 in the device.

  Or you may store in the said memory | storage part: 4.

S5 (Step 5): Disclosure processing Further, the result is transmitted to the upper management device 110 via the communication unit 21 in the order or at predetermined intervals.

  As described above, Step 2 to Step 5 are automatically performed in a series of processing flows. For this reason, it is possible to reduce the work of equipment operators and inspection specialists.

S3-2 (Step 3-2): Re-learning On the other hand, as described above, in the present invention, there is no information on a plurality of known detectors, and feature parameters determined to be heterogeneous and quantized image data are stored in the storage unit: 4 has accumulated. Accordingly, it is possible to create a new group or category from the feature parameters 93, 94, 95 and the quantized image data as needed, during a predetermined period or by looking at the accumulated amount. . That is, as implemented in step 1 described above, the new space information is stored in the storage unit 4 as an average value, a standard deviation, an inverse matrix of a correlation matrix or a covariance matrix, an average value of the Mahalanobis distance, and its Stores the standard deviation. In the example shown in the figure, since there are many images that are the same as the unknown detector 93, and it has been found that the unknown detectors 94 and 95 can be grouped or categorized, redefinition is performed, and the storage unit: 4 is registered.

  In detection of an object to be detected after the completion of such work, the newly registered category operates, so that the determination is automatically made within a series of processing flows from Step 2 to Step 5.

  In this way, according to the above procedure, discrimination or re-learning and reconstruction of recognition information is performed as necessary, and it is possible to increase the number of new categories from the accumulated data in a short time and without trouble. It has become. For this reason, the discrimination / recognition ability of the device is improved and the reliability is also improved.

  On the other hand, in the conventional neural network method in such a case, since it is necessary to re-learn and reconstruct the entire network, a great amount of man-hours are required. However, in the present invention, as described above, data collection has already been completed in step S-2, and since only the processing in step 1 is completed, the number of steps is very small.

  Hereinafter, the discrimination / recognition method used in the present embodiment will be described.

  5 and 6 are diagrams for explaining in detail the determination / recognition method according to the embodiment of FIG.

From the average value, standard deviation, correlation coefficient or inverse coefficient of the covariance matrix created in step 1 above, and the average value and standard deviation of the Mahalanobis space, each category is distinguished in the following steps. ·recognize.
{Ss1; Step}
An average vector: x _μ of each feature parameter of each group is created as the number of groups (m).

x _μ (i) = {x _μ 1, x _μ 2,..., x _μ k} i = 1 to m
{Ss2; Step}
The vector is input to each group, the Mahalanobis distance is calculated, and stored in the storage unit: 4. Examples of calculation results are shown in FIG. 5 and Table 5-1. As shown in Table 5-1, when the input vector is an average vector of the G1 group, the Mahalanobis distance of the G1 group is “0”,
The Mahalanobis distance of the G2 group is calculated as “4”, and the Mahalanobis distance of the Gm group is calculated as “100”. Thereafter, the same calculation is performed m times. In any row, the Mahalanobis distance of the corresponding group is “0” with respect to the average vector of the characteristic parameters of each group, and the other groups have arbitrary Mahalanobis distances. That is, this distance is the distance from the center of gravity of each group, and shows the average rank of the heterogeneity. Further, it is suggested that the detected object belongs to the space (group) indicating the minimum distance among these distances.

  Table 5-2 shows the result of inputting n unknown objects. When an unknown object is input, in this step Ss2, individual Mahalanobis distances are automatically calculated for G1 to Gm, and their ranks are determined. When the rank is determined, as described above, it is determined that the rank belongs to the space indicating the minimum distance.

For example, in the data of the detection object of detection object No. 3, the order of the distance is G1 (for example, red blood cells) <G2 (for example, white blood cells) <... Is determined. That is, in m Mahalanobis distances from G1 to Gm, the group having the smallest distance value and the detected object are “same”.
{Ss3; Step}
When the above steps are completed, two groups (G1: red blood cells, G2: white blood cells) having a high degree of homogeneity of unknown objects to be detected have been selected. It only has to be recognized or confirmed again. Naturally, it is estimated that it belongs to the group with the shortest distance, but in the present invention, the Mahalanobis distance ranking table of Table 5-1 is used to increase the accuracy. If the two groups to be recognized are clearly categorized, the recognition must be the same when one group recognizes the other and vice versa. However, as shown in Table 5-1, the actual individual Mahalanobis spaces are different (covariance matrix of feature parameters, average value of space, standard deviation), and the distances do not match. For this reason, the distance between the two selected groups in Table 5-1, that is, the distance between the centroids, is applied, and the final is determined by whether the Mahalanobis distance of the detected object is located at one of the two boundary lines. Discriminating and recognizing and improving its accuracy and reliability.

  In FIG. 6, the above-described recognition / discrimination method will be described using a specific example.

In FIG. 6, it is assumed that feature parameters are two variables X1 and X2, and three known categorized groups exist.
(1) When an unknown detector sample “Δ” (feature parameter {X1, X2}) is input, as shown in FIG. 5 and Table 5-2, the distance Gx1, to the center of gravity of each group G1, G2, G3 Gx2 and Gx3 are calculated. As a result, the distance becomes Gx2 <Gx1 <Gx3. Therefore, it can be determined that this sample belongs to either G2 or G1, and in other words belongs to G2.
(2) As shown in FIG. 6B, the Mahalanobis spatial distributions of G1 and G2 are separated by a distance √D ² x. For example, the number of feature parameters is different in FIGS. 5 and 6 for explanation, but in the example of FIG. 5 and Table 5-1, the distance between the centers of gravity of G1 and G2 is √D ² x = 2.24. is there. Each standard deviation (σ (g1), σ (g2)) of the space is stored in the storage unit 4 as described above.

From the distance and the standard deviation, the boundary value is obtained from {D ² (g1) −D ² (g2) = 0}.

That is, {√D ² (g1)} = √D ² (g2)}, and when this position is Xd,
(Xd−0.0) / σ (g1) = (√D ² (g2) −Xd) / σ (g2)
Thus, the value can be easily calculated. That is, in the example of the figure, based on the G2, sample position (X _delta) is Xd> X _delta becomes located inside the boundary line can be confirmed or verified as G2 and homogeneous definitely.

  Therefore, finally, this sample is judged and recognized as a specimen in the G2 group. In the present invention, even if the distance is on the boundary line, the probability at such a position is not 50% or less at worst. Therefore, essentially [deletion: type 1], no mistakes are made and the reliability is very high.

  On the other hand, as shown in FIG. 6A, a boundary line G1-2 between G1 and G2 can be obtained functionally and directly calculated. Compared to the former, the calculation time increases somewhat, but the accuracy and reliability do not change, so it is possible to select and select according to the capability of the equipment to be used.

  Furthermore, even when a new group is learned by the procedure of FIG. 4, it is easy to discriminate and recognize a new group simply by reconstructing the above-described ranking table (Table 5-1 of FIG. 5). It is clear that this can be achieved.

  As described above, the determination / recognition method of the present invention has high accuracy of determination / recognition and does not cause the first type of error, so that the reliability can be greatly improved. Furthermore, since complex logical operations are not performed, the real-time property can be sufficiently secured and the productivity is excellent.

  Hereinafter, the difference between the recognition / discrimination method according to the present invention and the neural network method that has been widely applied in this field will be described.

  The neural network method used in the confirmation experiment is a Rammel heart shape and has a three-layer network configuration having one intermediate layer.

  As shown in FIG. 7, the input of learning data to the neural network is a data group composed of variables X1 and X2, and is clearly four categorized data G1 to G4.

  This neural network presents a result that is correct with respect to the input data, that is, presents the teacher data to the output, changes the coupling coefficient of the intermediate layer (hidden layer), and minimizes the error between the input data and the teacher data. So that the coupling coefficient is manipulated. The neural network configured in this way requires prior learning. That is, it is necessary to determine the coupling coefficient through learning in advance. Therefore, as shown in Table 1, 100 input data are prepared as learning data, and four sets of teacher data corresponding to the input data are provided in advance to obtain the coupling coefficient of the intermediate layer (hidden layer). It was. Learning was continued until the cumulative error of the output data with respect to the teacher data was 0.005 or less.

  The number of nodes in the intermediate layer was increased as 2, 3, 4 and 6, and the number of nodes was evaluated and confirmed. As a result, the number of nodes was 2.

  Further, as the data for confirmation, the data (100) used at the time of learning and the maximum, minimum and average values of the data group of each group shown in the above figure are selected and combined to be 5 pieces (1 to 1 in the above figure). Fig. 5). Such data was input to the neural network method and this processing method. The results are shown in Table 1.

  From Table 1, the differences between the conventional method and this method are summarized below.

  In the case of the neural network method, a sample located at the end of each group has a large error and generates false information. In contrast, this treatment method does not occur at all. Therefore, in this processing method, it is clear that the detection target sample located at the end of each group or in the vicinity of the boundary does not generate any false information and the accuracy of discrimination / recognition is improved. Moreover, when the learning time in the neural network method in this case and this processing method is compared with the time in the case of processing with the same computer, in the former case, it took about 120 minutes / 4 groups or more. Completed within 1 minute / 4 groups. This is because in the neural network method, the network itself basically needs to be reconstructed in consideration of the weighting coefficient, the number of intermediate layers, and the like. This is because in the present invention, the formation of a new Mahalanobis space and its statistics need only be reconstructed. Therefore, in this processing method, when adding a new category or re-learning, the processing can be completed in a single time, and it was confirmed that the start-up man-hours can be greatly reduced.

  The method for determining and recognizing the processing unit of the device 100 has been described in detail above. Hereinafter, the functions of the management device 110 will be described again with reference to FIG.

As shown in FIGS. 4 and 1, the management device 110 is connected to the management device 110 via the communication unit 21 of the device 100, and other devices 200 to 202 are also connected to the management device 110 via the network 130. . Further, by installing the terminal device 120 on the network, the management device:
110 can be connected, and the result can be easily seen from a remote place. In such a configuration, the management device 110 includes the individual devices 100,
Information on the detection results 200, 201, and 202 and the determination / recognition results can be collected. Such information is accumulated and stored in the storage unit 114 provided in the management device.

  For example, in this configuration, the instrument 100 is a urine analyzer that recognizes and recognizes tangible particles in urine as in the embodiment. On the other hand, the other device: 200 is a urine qualitative analysis device that measures protein, sugar, occult blood, etc. in urine, and a blood analysis device that analyzes blood as another device: 201. Data collected by such a device group has a data format as shown in FIG. 4 and Table 4-1. This data format is exactly the same as the data format of the determination / recognition method in the aforementioned device 100. But the quality is very different. For example, consider blood pressure measurement, which is a single measurement from individual devices. In blood pressure measurement, the normal range is determined and the subject is judged / recognized as low blood pressure or high blood pressure. However, even if it is determined that the blood pressure is high or low, the specialist takes into account the degree of obesity and recent changes in diet, etc. Not known is well known.

  Therefore, the data format is composed of results determined and recognized with high accuracy and high reliability by individual devices, and these are collected and used as characteristic parameters. This data is suitable for making a judgment. This is equivalent to data collection for the expert to make a final judgment at the time of blood pressure measurement. Furthermore, this data format is advantageous because the processing method applied to the device: 100 can be applied to the processing unit 2: 111 of the management device as it is.

  Furthermore, in the present invention, as described above, a reference space is formed for each of m groups or categories known from the beginning. In the medical field, this category corresponds to an individual. In other words, it is possible to make a comparison with a more subdivided group rather than the large category of healthy people mentioned above. In other words, in the past, only the difference from healthy people can be known, but by applying and expanding this method, the difference from “a group of healthy people and good exercise” or “healthy people and their eating habits” It is possible to recognize differences from “groups that are mainly vegetable”, providing high-value-added information that is more accurate and can contribute to comprehensive judgment and recognition. Can be achieved. In addition, a highly accurate system can be provided at a lower cost.

  In the above description, the detection object detection, the determination / recognition of particles in urine, and diagnosis support in the medical field have been described. However, the present invention is not limited to this. For example, the present invention can be applied to manufacturing process monitoring / inspection systems, plant process monitoring systems, various diagnosis support systems in other medical fields, inspection systems, and the like.

  In the above description, the discrimination / recognition method is described as being provided in the management device 110 and the detection device 100, respectively. However, in the present invention, the processing method and the data format are shared, and therefore, only the internal management device 110 of the device may be provided. In this embodiment, the detection device: 100 simply transmits the characteristic parameter calculation result to the management device via the communication unit 21.

  Note that the output unit: 3 of the device: 100 may display only the operation status, or may transmit the result from the management device: 110 and display it. For this reason, in this embodiment, the speed of each device can be increased, and the real-time property is improved. Furthermore, since individual devices can be manufactured at a lower cost, there is an effect that the economy is improved.

FIG. 2 is a block diagram and a system configuration diagram of a detection device according to the present invention. Detailed explanatory drawing of the imaging | photography part of the detection apparatus which concerns on this invention. The figure explaining the basic flow of the process part of the detection apparatus which concerns on this invention. The figure explaining the processing flow of the system of the detection apparatus which concerns on this invention. The figure explaining the discrimination | determination / recognition processing method concerning this invention. The figure explaining the discrimination / recognition example by Mahalanobis space which concerns on this invention, and its processing method. Comparison between the present invention and a neural network technique.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Imaging | photography part, 2 ... Processing part 1, 3 ... Output part, 4 ... Memory | storage part, 5 ... Image extraction part, 6 ... Feature parameter calculation part, 7 ... Feature parameter group, 8 ... Detected object, 21 ... Communication part 41 ... Process 1,
43 ... Re-learning data group, 90, 91, 92 ... Known detectors 1, 2, 3, 93, 94, 95 ... Detector data for relearning, 100 ... Detection device, 110 ... Management device, 111 ... Management device Processing unit 114... Management device storage unit 120... Terminal device 130 130 network 200 device 2 201 device 3 202 device 4

Claims (6)

Feature parameter extraction means for extracting multivariate feature parameters from a plurality of known subjects;
Mahalanobis space computing means for forming a Mahalanobis space (reference space) with each of the plurality of subjects based on the feature parameters extracted by the feature parameter extracting means;
Storage means for preliminarily storing information on a plurality of formed reference spaces;
Feature parameter extraction means for extracting feature parameters from an unknown detector;
Mahalanobis distance calculation means for applying the feature parameter extracted by the feature parameter extraction means to the Mahalanobis space formed by the Mahalanobis space calculation means, and calculating the Mahalanobis distance of the unknown detector;
A comparison means for comparing the Mahalanobis distance calculated by the Mahalanobis distance calculation means;
A discriminating means for discriminating a detected object based on a comparison result in the comparing means;
A system for identifying an object to be detected.   The comparison means according to claim 1 calculates a separation distance between the Mahalanobis distance of the detected object and the Mahalanobis space of the plurality of known detection objects, and based on the separation distance (Mahalanobis distance), the similarity or A system for determining an object to be detected, wherein a degree of heterogeneity is calculated and it is determined whether the object to be detected belongs to an existing or known group or category. The discriminating means according to claim 2 sequentially compares the Mahalanobis space of a plurality of known detection objects and a plurality of Mahalanobis distances (D ² (x)) calculated from the characteristic parameters of the detected object at those distances. A system for discriminating an object to be detected, characterized in that it is judged and recognized as being of the same quality as a known detecting body having the smallest distance among these distances and being otherwise heterogeneous. In the system for discriminating an object to be detected according to claim 2,
In the storage means, a plurality of reference spaces calculated from feature parameters of a plurality (m) of known detectors are preliminarily converted into two-dimensional maps by their Mahalanobis distances (m D ² (x)). Memorized information (two-dimensional table of rank order of m × m),
The comparison means sequentially compares the Mahalanobis distance calculated from the characteristics of the detected object,
Extracting and recognizing the eigenspace of a known detection body having the first minimum distance and the second minimum distance among the distances in the map stored in the storage means by the determination means to determine and recognize A distinction system for a featured object. In the discrimination system of the detected object according to any one of claims 1 to 4,
When it is impossible to judge / recognize that the detected object is homogeneous or heterogeneous, information on the detected object stored in the storage means and feature parameters extracted from the information are stored in the storage means. A system for discriminating an object to be detected, characterized by creating a new group or category. Imaging means for capturing and capturing the detected object;
Feature parameter extracting means for extracting a plurality of feature parameters from the image captured by the imaging unit;
Mahalanobis space computing means for forming a Mahalanobis space (reference space) with each of the plurality of subjects based on the feature parameters extracted by the feature parameter extracting means;
Storage means for preliminarily storing information on a plurality of formed reference spaces;
Feature parameter extraction means for extracting feature parameters from an unknown detector;
Mahalanobis distance calculation means for applying the feature parameter extracted by the feature parameter extraction means to the Mahalanobis space formed by the Mahalanobis space calculation means, and calculating the Mahalanobis distance of the unknown detector;
A comparison means for comparing the Mahalanobis distance calculated by the Mahalanobis distance calculation means;
A discriminating means for discriminating a detected object based on a comparison result in the comparing means;
Image discrimination system with

Images