JP2006011574A - Support device for registration of teacher data - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To avoid loss of learning processing time based on registration of incorrect teacher data, in a defect type discrimination system having a neural network. <P>SOLUTION: A support device for registration of teacher data is provided with: a collation means which collates a feature quantity string of defect type names inputted by an input means with a feature quantity string corresponding to each of two or more defect type names registered in a database by quantization unit defined for each pictorial feature beforehand; a computing means which when the result of collation by the collation means indicates that there is no registered feature quantity string matching the feature quantity string of the inputted defect type name, computes a matching degree in quantization unit between the feature quantity string of the inputted defect type name and the average value of a feature quantity string previously registered as the feature quantity string of the defect type name thereof; and a presentation means which presents the matching degree computed by the computing means. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention supports registration operation of teacher data necessary for learning processing for optimizing various arithmetic coefficients included in a neural network constituting a defect type discrimination system (for example, applied to a continuous paper making line). The present invention relates to a registration support apparatus for teacher data to be used.

  A binarized image generating means for generating a binarized image from a defect-corresponding image cut out from an image obtained by photographing an inspection object (for example, a strip of paper flowing through a papermaking line); A feature amount extracting means for extracting a feature amount relating to each of various image features from the binarized image generated by the image generating means, and a feature amount relating to each of the various image features extracted by the feature amount extracting means. 2. Description of the Related Art A defect type discrimination system having a neural network that discriminates a defect type (insect, light defect, black spot, hole, thin tear, etc.) based on the conventional type is known (see, for example, Patent Document 1).

  In order to increase the defect type discrimination accuracy in this type of defect type discrimination system, it is necessary to optimize the calculation coefficient (for example, the coupling coefficient W, the determination threshold value of the neuron state S) of each neuron included in the neural network. The learning process for the optimization is realized by executing the learning process based on the teacher data (defect type name and the input feature quantity string corresponding to the teacher data) registered by the operator.

Conventionally, the registration process of teacher data is performed by an operation of inputting a defect type name corresponding to a given input feature amount sequence. At this time, the determination of the defect type name corresponding to the given input feature quantity sequence is left to the operator's visual determination based on the defect-corresponding image displayed on the screen.
Japanese Patent Laid-Open No. 10-302049

  However, since there is not necessarily a strictly unambiguous relationship between the defect-equivalent image and the corresponding defect type name, the defect type name determination method that relies only on the above-mentioned visual judgment is There is a risk that inappropriate teacher data may be registered due to a visual judgment error of the type name. If such erroneous teacher data is registered, in the learning calculation process for optimizing the calculation coefficient of the neuron, the learning calculation process does not converge any time and results in wasted time. In particular, this kind of learning calculation process repeats the specified calculation operation by changing the coupling coefficient and threshold value of each neuron slightly, so even if the teacher data is registered correctly, it is several tens of minutes. The degree of time is such that loss of time due to giving incorrect teacher data can lead to a significant reduction in work efficiency.

  The present invention has been made paying attention to the above-mentioned problems, and the object of the present invention is a learning process based on registering erroneous teacher data in a defect type discrimination system including this type of neural network. An object of the present invention is to provide a teacher data registration support apparatus capable of avoiding time loss.

  Still other objects and operational effects of the present invention will be easily understood by those skilled in the art by referring to the following description of the specification.

  A teacher data registration support apparatus according to the present invention includes a binarized image generating unit that generates a binarized image from a defect-corresponding image cut out from an image obtained by photographing an inspection object, and binarization A feature amount extracting means for extracting a feature amount relating to each of various image features from the binarized image generated by the image generating means, and a feature amount relating to each of the various image features extracted by the feature amount extracting means. In a defect type discrimination system having a neural network for discriminating a defect type based on it, it is used to support a registration operation of teacher data necessary for learning processing for optimizing various calculation coefficients included in the neural network. The

  The teacher data registration support apparatus includes a database in which a defect type name corresponding to already registered teacher data and a feature amount sequence thereof are registered, a defect type name corresponding to teacher data to be newly registered, and a feature thereof. An input means for inputting a quantity sequence, a feature quantity sequence of defect type names input by the input means, and a feature quantity sequence corresponding to each of a plurality of defect type names registered in the database The collation means that collates with the quantization unit defined for each feature, and the result of the collation by the collation means is that there is no registered feature quantity sequence that matches the feature quantity string of the input defect type name. A calculation means for calculating a degree of coincidence in a quantization unit between a feature value string of a defect type name and an average value of a feature value string already registered as a feature value string of the defect type name; And presenting means for presenting the degree of coincidence is the operator, contains.

  According to such a configuration, when there is no registered feature quantity sequence that matches the feature quantity string of the input defect type name as a result of collation by the collation means, new registration is possible. Prior to registration, the certainty of the teacher data (defect type name and its feature quantity string) is presented to the operator as the degree of coincidence, so a feature quantity string that does not match the defect type name is erroneously registered as teacher data. As a result, the problem of impeding the learning process is avoided.

  In the teacher data registration support apparatus according to the present invention, it further corresponds to the feature quantity sequence in which the defect type name corresponding to the registered feature quantity sequence for which collation matching is determined in the quantization unit by the collation means is input. Output means for performing a predetermined output operation when the defect type name does not match.

  According to such a configuration, when the defect type name corresponding to the registered feature value sequence for which the matching match in the quantization unit is determined by the matching unit does not match the defect type name corresponding to the input feature value sequence. In addition, since it includes output means for performing a predetermined output operation, when trying to mistakenly assign different defect types to these two defect-corresponding images even though the feature amount sequences are almost the same, A predetermined output operation indicating that effect is performed, and based on this, the operator can notice an error in his / her judgment.

  In the above-described apparatus, the predetermined output operation may be a warning operation for the operator. According to such a configuration, when different feature types are assigned to the two defect-corresponding images even though the feature amount sequences are substantially the same, a warning operation indicating that effect is performed. Based on this, the operator can notice an error in his judgment. Here, the warning operation can include displaying a warning text on the screen of the image display, in addition to turning on the buzzer and the lamp.

  In the above-described apparatus, the predetermined output operation may be presentation of a chart corresponding to two feature amount sequences related to the mismatch of the defect type name to the operator. According to such a configuration, the operator can determine whether or not to assign the defect type name by collating these two feature amount sequences in more detail.

  In the above-described apparatus, a defect-equivalent image based on the chart may be added to each chart. According to such a configuration, the operator can more strictly determine the suitability of the defect type name assignment by comparing the defect equivalent images.

  In the above apparatus, the presented chart may be a radar chart. According to such a configuration, the operator can determine whether or not to assign the defect type name by more accurately comparing each feature amount constituting the feature amount sequence through vision.

  In each apparatus described above, the inspection object may be a sheet-like object (for example, a steel plate or a film) placed in a continuous manufacturing process. In particular, when the sheet-like object is a belt-like paper placed on a continuous paper making line, there is an advantage that it greatly contributes to quality improvement in paper for food use. In addition, when the defect type name corresponding to the registered special feature column for which the matching match in the quantization unit is determined by the matching unit matches the defect type name corresponding to the input special feature column, the input defect type Calculate the degree of coincidence in the quantization unit between the special quantity column of the name and the average value of the special quantity column already registered as the special quantity column of the defect type name, and present the calculated degree of coincidence to the creator Also good.

  According to the present invention, in the defect type discrimination system having this type of neural network, it is possible to avoid a loss of learning processing time based on registering erroneous teacher data.

  In the following, a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings. A system configuration diagram showing the relationship between the continuous paper making line and the apparatus of the present invention is shown in FIG.

  In the figure, reference numeral 1 denotes a winder machine provided at the subsequent stage of the continuous paper making line, 2 denotes a winder operation support device, and 3 denotes a sheet inspection device. As is well known to those skilled in the art, the WCS server PC 5 constituting the winder operation support apparatus 2 controls the operation of the winder machine 1 via a programmable controller (PLC) 4. The sheet inspection apparatus 3 includes a camera, a light source, an image processing (defect detection) board, and various processing PCs, and information on detected defects and image data are stored in a database of the data processing PC 6. At the same time, this data is also stored in the WCS server PC5. The type of defect is determined using these image data.

  The WCS server PC5 constitutes the defect type discrimination system of the present invention. That is, the defect type discrimination system includes a binarized image generating unit that generates a binarized image from a defect-corresponding image cut out from an image obtained by photographing the surface of paper that is an inspection object; A feature amount extracting means for extracting feature amounts relating to each of various image features from the binarized image generated by the binarized image generating means, and each of various image features extracted by the feature amount extracting means. And a neural network that discriminates the defect type based on the feature amount.

  More specifically, a block diagram of the defect type discrimination system is shown in FIG. In the figure, a defect image 10 cut out from an image obtained by photographing an inspection object is classified into a bright or dark image and a light / dark mixed image in a defect color discrimination unit 11.

  The bright / dark mixed image is supplied in parallel to the bright side image binarization unit 12 and the dark side image binarization unit 14. The bright side image binarization unit 12 performs a binarization process of the bright side image in the bright and dark mixed image. Similarly, the dark side image binarization unit 14 is responsible for binarization processing of the dark side image in the light / dark mixed image. The binarized images obtained from the bright side and dark side image binarizing units 12 and 14 are synthesized in the binarized image synthesizing unit 15 to obtain a final binarized image that is a basis for feature extraction. . On the other hand, the bright or dark image classified by the defect color discrimination unit 11 is sent to the image binarization unit 13 where it is converted into a binarized image. The binarized images thus obtained by the binarized image composition unit 15 and the image binarization unit 13 are sent to the image preprocessing unit 16.

  The image preprocessing unit 16 performs predetermined preprocessing on the binarized image generated by the previous processing. This preprocessing is for feature quantity extraction. That is, the binarized image preprocessed by the image preprocessing unit 16 is sent to the feature amount extraction unit 17, where the grayscale feature amount and the shape feature amount are extracted.

  The primary neural network discrimination processing unit 18 performs defect type discrimination based on the shape feature value and the light and shade feature value extracted by the feature value extraction unit 17. In this way, the type determination output 21 is generated and output.

  On the other hand, in this embodiment, a secondary neural network discrimination processing unit 20 is further provided. The secondary neural network discrimination processing unit 20 executes a defect type discrimination process from another point of view based on the density histogram obtained from the image preprocessing unit 16 and the processing result data from the primary neural network discrimination processing unit 18. To do.

  Based on the discrimination result of the primary neural network discrimination processing unit 18 and the discrimination result of the secondary neural network discrimination processing unit 20 obtained in this way, a final defect type discrimination output 21 is generated. The exception processing unit 19 generates an exception processing signal based on the preprocessing result of the image preprocessing unit 16. This exception processing signal becomes the type determination output 21. The type discrimination output 21 mentioned here includes insects, light defects, black spots, holes, thin tears, and the like.

  According to such a defect type discriminating system, the operator of the continuous paper making line can use the type discriminating output 21 to place an insect, a light defect, a black spot, a hole, a thin tear, etc. on any part of the belt that flows on the continuous paper making line. It is possible to reliably recognize whether there is a defect. Then, based on the determination result, the defect portion is searched by, for example, reversing the winder machine and removed as appropriate.

  In order to increase the defect type discrimination accuracy in this type of defect type discrimination system, the operation coefficients (for example, the coupling coefficient W, the determination threshold value of the neuron state S, etc.) of each neuron included in the neural networks 18 and 20 are optimized. I have to. The learning process for the optimization is realized by executing the learning process based on the teacher data (defect type name and the input feature quantity string corresponding to the teacher data) registered by the operator. The teacher data registration support apparatus of the present invention supports a teacher data registration operation necessary for a learning process for optimizing various calculation coefficients included in the neural network.

  As described above, this teacher data registration support device corresponds to a database in which defect type names and feature quantity sequences corresponding to already registered teacher data are registered, and teacher data to be newly registered. Input means for inputting the defect type name and the feature quantity sequence thereof, the feature quantity sequence of the defect type name input by the input means, and the feature quantity sequence corresponding to each of the plurality of defect type names registered in the database Are collated in a quantization unit defined in advance for each image feature, and as a result of collation by the collating unit, a registered feature quantity sequence that matches the feature quantity string of the input defect type name is An arithmetic means for calculating a degree of coincidence in a quantization unit between the input feature quantity sequence of the defect type name and the average value of the feature quantity sequence already registered as the feature quantity sequence of the defect type name; Presenting means for presenting the degree of coincidence calculated by the computing means to the operator, and a feature quantity sequence in which the defect type name corresponding to the registered feature quantity sequence for which matching matching in the quantization unit is determined by the matching means is input Output means for performing a predetermined output operation when there is a mismatch with the corresponding defect type name.

  The details of the teacher data registration support processing for realizing such a teacher data registration support device are shown in the flowchart of FIG. The teacher support process is started in response to input of teacher data. The determination of the “defect type name” in the teacher data is left to the operator's visual judgment based on the defect image displayed on the screen.

  That is, when the teacher data D is input, the teacher data D is captured (step 301). Here, newly input teacher data D is represented as a combination of a defect type name [Name] and a feature string [c1, c2,... Cn]. That is, c1 is represented as the first feature, c2 is represented as the second feature, and cn is represented as the nth feature.

  Next, a feature quantity sequence [c1, c2,... Cn] constituting newly input teacher data and a feature quantity sequence [ck1, ck2,... Ckn (k = 1, 2) for each defect type stored in the database. ... Kmax)] are executed one by one. That is, first, the pointer k is set to 1 (step 302), the value of the pointer k is updated by +1 (step 309), the registered data Dk is sequentially read, and the feature data string [c1, c2, c2 ... Cn] and the registered data feature quantity sequence [ck1, ck2,... Ckn] are collated (step 304).

  At this time, if the collation results do not match (NO in step 305), the value of the pointer k is incremented by +1 (step 309), the next registered data is read (step 303), and both are collated (step 309). 304). On the other hand, when a match between the two is determined (YES in step 305), a match between the defect type name [Name] of the teacher data D and the defect type name [Namek] of the registered data Dk is further determined. If the two match (YES in step 306), the value of the pointer k is updated by +1 in the same manner (step 309), and collation processing (step 304) is performed for the next registered data Dk. The process (step 305) is executed until the value of the pointer k reaches the final value kmax (step 308).

  On the other hand, if the collation results match (step 305), and it is determined that the defect type name [Name] of the teacher data D and the defect type name [Namek] of the registered data Dk do not match (NO in step 306). A predetermined output process is executed (step 307).

  An explanatory diagram showing details of the output processing is shown in FIG. In this example, when the output process is started, a predetermined warning display and a warning buzzer are first sounded to the operator (step 401), and then a chart request operation from the operator is waited (step 402). ) If necessary, a process of displaying the teacher data D and the registered data Dk side by side using a radar chart is executed (step 403). Details of the display mode by the radar chart will be described later in detail with reference to FIG.

  Returning to FIG. 3, the collating process in step 304 and the coincidence determining process in step 305 are performed in the quantization unit in the present invention, as shown in FIGS. That is, in the present invention, as shown in FIG. 6A, feature amounts 0 to 0.19, 0.2 to 0.39, 0.4 to 0.59, 0.6 to 0,. 79, 0.8 to 1.00, etc., the feature amount range 0.00 to 1.00 is divided into five levels of quantization, each level being level 1, level 2, level 3, Quantization is performed using indices such as level 4 and level 5. For example, assuming that the defect A has an area of 0.15, an area ratio of 0.55, a circularity of 0.65, a ferret ratio of 0.85, and a roundness of 0.35. By using the previous quantization unit, the area can be expressed as level 1, the area ratio is level 3, the circularity is level 4, the ferret ratio is level 5, the roundness is level 2, and so on.

  Then, using this quantization unit (level 1 to level 5), the matching process (step 304) and the match determination process (step 305) are executed. As a result, the matching process (step 304) and the match determination process (step 305) are performed with a certain degree of redundancy. That is, the collation between the feature value sequence [c1, c2,... Cn] of the teacher data D and the feature value sequence [ck1, ck2,... Ckn] of the registered data Dk is performed with 0.2-width redundancy. .

  According to such a configuration, when the defect type name corresponding to the registered feature value sequence for which the matching match in the quantization unit is determined by the matching unit does not match the defect type name corresponding to the input feature value sequence. Since a predetermined output operation is performed, when a defect type is erroneously assigned to these two defect-equivalent images even though the feature amount sequences are almost the same, this is indicated. A predetermined output operation is performed, and based on this, the operator can notice an error in his / her judgment, and a necessary correction operation is prompted.

  This is explained in more detail in the explanatory diagram of the collation processing of FIG. FIG. 4A shows the registration state in the database. FIG. 5B shows a feature amount sequence relating to a defect to be registered.

  Now, suppose that there are already three types of defects A, B, and C in the database, and the defect A is a feature quantity sequence {(1, 3, 4, 5, 2), (1, 3, 4, 5, 3), (1, 3, 4, 5, 4)} and defect B is a feature quantity sequence {(2, 3, 4, 5, 1), (2, 4, 4, 5, 1). , (2, 5, 4, 5, 1)} and defect C are feature quantity sequences {(5, 5, 1, 3, 2), (5, 5, 2, 2, 3), (5, 5, 1, 2, 3)}.

  At this time, when a new defect C is to be registered, the feature value sequence (2, 3, 4, 5, 1) is already registered as the feature value sequence (2, 3, 4, 5) of the defect B. , 1). At this time, the feature quantity strings of both coincide with each other in the quantization unit, whereas the defect type names of the two differ in that one is the defect B and the other is the defect C.

  In such a case, in the flowchart of FIG. 3 described above, step 305 is determined to be “match” and step 306 is determined to be “not match”, and as a result, a predetermined output process is executed (step 307). Then, as described above with reference to FIG. 4, the teacher data D and the registered data Dk are displayed side by side on the radar chart according to the warning display, the alarm buzzer sounding (step 401), and the operation requested by the operator. Processing (step 403) is executed.

  FIG. 9 shows an example of the feature amount sequence displayed side by side on the radar chart in this way. As is clear from comparison between FIG. 6A and FIG. 6B, the difference between each feature amount is expressed as a graphic difference between the radar chart 902a and the radar chart 902b. The operator can easily recognize the characteristic differences. Moreover, since the target defect images 901a and 901b are further added to the respective radar charts 902a and 902b, the difference between the defect-corresponding image and the value of the feature-specific feature value corresponding thereto can be easily visually confirmed. Can be recognized.

  Next, returning to FIG. 3, when the teacher data to be newly registered does not match any of the already registered teacher data (YES in step 308), the process proceeds to FIG. A process of calculating the degree of coincidence in the quantization unit between the feature amount sequence and the average value of the feature amount sequence already registered as the feature amount sequence of the defect type name is executed.

  That is, first, a process for obtaining a feature quantity of a defect type set for a defect to be registered is executed from among already registered defects (step 501). Subsequently, a process of leveling (quantizing) the average for each feature amount of the set defect type in five stages is executed (step 502). Subsequently, a process of performing pattern matching between the average of the feature amounts of the set defect types and the defects to be registered is executed (step 503). Subsequently, a collation result determination process is executed (step 504).

  Here, as shown in FIG. 8A, when it is determined that the patterns of the defect type names A match 100% (step 505), all the features are displayed on the predetermined screen. A double circle indicating that the amounts match is displayed (step 508). Further, as shown in FIG. 8B, when it is determined that the pattern matches 80% or more between the defect type names A (step 506), 80% is displayed on the predetermined screen. A circle indicating that the feature amounts match is displayed (step 509). Furthermore, as shown in FIG. 8C, when it is determined that the pattern does not match 80% or more between the defect type names A (step 507), 80% of the pattern is displayed on the predetermined screen. A cross indicating that the feature amounts do not match is displayed (step 509).

  In the example of FIG. 9, not only comparing FIG. 9 (a) and FIG. 9 (b), but also trying to register in each of FIGS. 9 (a) and 9 (b). The feature amount of the defect (filling of the radar chart) is compared with the registered average feature amount (straight line of the radar chart). That is, if the feature quantity to be registered (the radar chart fill) does not match the registered average feature quantity (radar chart feature), FIG. This is a case where the feature amount to be tried (radar chart fill) does not match the registered average feature amount (radar chart straight line).

  As is apparent from the above description, the teacher data registration support apparatus of the present invention is a database in which defect type names corresponding to already registered teacher data and their feature quantity sequences are registered (see FIG. 7A). And an input means for inputting the defect type name corresponding to the teacher data to be newly registered and its feature quantity string (see FIG. 7B), and the feature quantity of the defect type name input by the input means Collation means for collating the column and the feature amount column corresponding to each of the plurality of defect type names registered in the database with a quantization unit (see FIG. 6A) defined in advance for each image feature (See steps 304 and 305 in FIG. 3), and if there is no registered feature value sequence that matches the feature value sequence of the input defect type name as a result of the verification by the verification means (NO in step 308 in FIG. 3), the input Defective species Calculating means (steps 501, 502, and 503 in FIG. 5) for calculating the degree of coincidence in the quantization unit between the feature value sequence of the name and the average value of the feature value sequence already registered as the feature value sequence of the defect type name; And the presenting means (steps 505 to 507 and 508 to 510 in FIG. 5) for presenting the degree of coincidence calculated by the computing means to the operator. Only when there is no registered feature string that matches the feature string of the name, new registration is possible, and in addition to registration, the teacher data (defect type name and its feature string) Since the certainty is presented to the operator as the degree of coincidence, a feature quantity sequence that is not suitable for the defect type name is erroneously registered as teacher data, thereby avoiding the problem of impeding the learning process. .

  In addition, in the teacher data registration support apparatus of the present invention, the defect type name corresponding to the registered feature value sequence for which the matching match in the quantization unit is determined by the matching unit is input to the input feature value sequence. It includes output means (see step 307 in FIG. 3) that performs a predetermined output operation when it does not match the corresponding defect type name.

  Therefore, when the defect type name corresponding to the registered feature quantity sequence for which the matching mismatch in the quantization unit is determined by the matching unit does not match the defect type name corresponding to the input feature quantity sequence, a predetermined output operation is performed. Since the output means (step 307) to be performed is included, when different feature types are assigned to the two defect-corresponding images even though the feature amount sequences are substantially the same, this is indicated. A predetermined output operation is performed, and based on this, the operator can notice an error in his / her judgment.

  Therefore, according to the present invention, it is possible to avoid a loss of learning processing time based on registering erroneous teacher data. That is, if incorrect teacher data is registered, in the learning calculation process for optimizing the calculation coefficient of the neuron, the learning calculation process does not converge indefinitely, resulting in wasted time. In particular, this kind of learning calculation process repeats the specified calculation operation by changing the coupling coefficient and threshold value of each neuron slightly, so even if the teacher data is registered correctly, it is several tens of minutes. The degree of time is such that loss of time due to giving incorrect teacher data can lead to a significant reduction in work efficiency.

It is a system block diagram which shows the relationship between a continuous paper making line and this invention apparatus. It is a block diagram of a defect classification discrimination system. It is explanatory drawing (the 1) of a registration assistance process. It is explanatory drawing which shows the detail of an output process. It is explanatory drawing (the 2) of a registration assistance process. It is explanatory drawing of the level division | segmentation of the feature-value. It is explanatory drawing of the feature-value row | line collation process of a registered defect and a newly registered defect. It is explanatory drawing of the feature-value row | line collation process of a recorded defect average value and a newly registered defect. It is explanatory drawing which shows an example of a radar chart.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Winder machine 2 Winder operation support apparatus 3 Sheet inspection apparatus 4 Programmable controller (PLC)
5 WC server PC
6 Data processing PC
DESCRIPTION OF SYMBOLS 7 Network 10 Defect image 11 Defect color discrimination | determination part 12 Bright side image binarization part 13 Image binarization part 14 Dark side image binarization part 15 Binary image composition part 16 Image pre-processing part 17 Feature quantity extraction part 18 Primary neural network discrimination processing unit 19 Exception processing unit 20 Secondary neural network discrimination processing unit 21 Type discrimination output 901a, 901b Defect equivalent image 902a, 902b Radar chart

Claims (8)

A binarized image generating unit that generates a binarized image from a defect-corresponding image cut out from an image obtained by photographing the inspection object, and a binarized image generated by the binarized image generating unit A feature quantity extracting means for extracting feature quantities relating to each of various image features from a neural network for determining a defect type based on the feature quantities relating to each of various image features extracted by the feature quantity extracting means. In the defect type identification system that has
A teacher data registration support apparatus used to support a teacher data registration operation necessary for learning processing for optimizing various calculation coefficients included in the neural network,
A database in which defect type names corresponding to already registered teacher data and their feature quantity columns are registered;
An input means for inputting the defect type name corresponding to the teacher data to be newly registered and the feature amount sequence;
A feature amount string of defect type names input by the input means and a feature amount sequence corresponding to each of a plurality of defect type names registered in the database are in a quantization unit defined in advance for each image feature. Matching means for matching;
As a result of the collation by the collation means, when there is no registered feature amount sequence that matches the feature amount sequence of the input defect type name, the feature amount sequence of the input defect type name and the feature amount sequence of the defect type name are A computing means for calculating the degree of coincidence in the quantization unit with the average value of the already registered feature quantity sequence;
Presenting means for presenting the degree of coincidence calculated by the computing means to the operator;
A teacher data registration support apparatus comprising:   Output that performs a predetermined output operation when the defect type name corresponding to the registered feature quantity sequence for which the matching match in the quantization unit is determined by the matching means does not match the defect type name corresponding to the input feature quantity sequence The teacher data registration support apparatus according to claim 1, further comprising: means.   3. The teacher data registration support apparatus according to claim 2, wherein the predetermined output operation is a warning operation for the operator.   The teacher data registration support apparatus according to claim 2, wherein the predetermined output operation is presentation of a chart corresponding to two feature amount sequences related to a mismatch of defect type names to an operator.   5. The teacher data registration support apparatus according to claim 4, wherein a defect equivalent image based on the chart is added to each chart.   The teacher data registration support apparatus according to claim 4, wherein the presented chart is a radar chart.   The teacher data registration support apparatus according to claim 1, wherein the inspection object is a sheet-like object placed in a continuous manufacturing process.   8. The teacher data registration support apparatus according to claim 7, wherein the sheet-like object is a belt-like paper placed on a continuous paper making line.

Images