JP,2012-026982,A
`
`
`
`
`Primary document PA}
`* NOTICE *
`3°O and INPIT are not responsible for any darnages caused by the use of this
`transiation.
`1. This docurnent has been translated by cornputer. So the translation may not reflect the origina!
`precisely.
`2. 7% shows a word which cannot be translated.
`3. In the drawings, any words are not translated.
`
`
`
`3
`
`(19) [Publication country] JP
`{12} [Kind of official gazette] A
`(11) [Publication nurnber] 2012026982
`(43) [Date of publication of application] 20120209
`(54) [Title of the invention] INSPECTION DEVICE
`(S51) [International Patent Classification]
`GOIN 21/88
`(2006.01)
`GO6ET 1/00
`(2006.01)
`[FI]
`GOIN 21/88
`300
`GO6T 1/00
`(21) [Application number] 2010168656
`(22) [Filing date] 20100727
`(71) [Applicant]
`[Name] PANASONIC ELECTRIC WORKS SUNX CO LTD
`(72) [Inventor]
`[Full name] NISHIHARA TOSHIMASA
`[Full name] HASHIMOTO YOSHIHITO
`[Full name] IKEDA KAZUTAKA
`[Theme code (reference) ]
`26051
`SBOS?
`[F~term (reference) |
`2G051CA04
`2GOS5TEA12
`2G051TEAL4
`Z2GO5LEA21
`
`

`

`2GO5LECOL
`2GO051ED01
`2G051ED1L1
`2G051£D21
`2GO5iFAOL
`5B057AA02
`5BO57DA03
`5SBO57DAL2
`5B057DB02
`5B057DB09
`5BOS7DCi4
`5SBO57DC22
`5B057DC40
`
`
`
`(57) [Overview]
`PROBLEM TO BE SOLVED: To provide an inspection device which is capable of not only de
`termining the presence or absence of abnormality but also classifying abnormality types
`and is capable of reducing an information amount required for learning for abnormality ci
`assification in comparison with conventional configurations.
`SOLUTION: A first processing unit 1 is provided with a first learnt neural network 101, to
`classify inspection object signals into normal signals and the other signals. A second prac
`assing unit 2 is provided with a second learnt neural network 201, and extracts partial si
`gnals including regions other than normal regions, from inspection object signals classifie
`d to signals other than normal signals and classifies types of abnormality of the partial si
`gnais. An output part 3 outputs classification results of the first processing unit 1 and the
`second processing unit 2. The first neural network 101 is learnt by using only inspection
`object siqnals being normal signals, and the second neural network 201 is learnt so that
`types of abnormality are classified by using inspection object signals including known ab
`normalities.
`
`

`

`
`
`
`
`a3
`
`20——eea
`
`BSsae
`
`
`a2Sh
`
`
`
`|POMSTANNEROTANRROOIRNROIOTROSREYROYREYROYRDNNORTEPEO.
`
`
`
`
`
`£2
`
`TETA
`
`[Patent Claims]
`aim 1]
`A 1 processing part provided with a learned 1 neural network for classifying the signal to
`be inspected into normal and non-normal; A 2 processing part for extracting a partial si
`gnal including an area other than norrnal from an inspection tarqet signal classified by th
`
`

`

`e 1 processing part and classified by the 2 processing part, and classifying the type of ab
`normality for the partial signal ; An inspection apparatus comprising : an output unit that
`outputs a classification result by a 1 processing unit and a 2 processing unit ; a 1 neural
`network that is learned using only an inspection target signal that is normal; and a2 ne
`ural network that learns the type of an abnormality using an inspection target signal inci
`uding a known abnormality.
`[Claim 2]
`The inspection apparatus according to claim 1, wherein the inspection target signal is an
`image signal of an image obtained by imaging an inspection target by an image input uni
`
`t [
`
`Claim 3]
`This device is provided with an abnormal image storage part for storing an image includi
`ng an abnormality, an image superposition part for displaying the image selected from th
`@ abnormal image storage part on a normal image on the screen of the monitor device, a
`nd an adjustment part for adjusting the parameter of the image selected from the abnor
`mal image storage part by an instruction from the input device. The inspection apparatus
`according to claim 2, wherein conditions for classifying the 2 neural network by the para
`meter adjusted by the adjusting unit are adjusted.
`[Claim 4]
`An inspection area dividing part for dividing the image into partial areas based on the re
`quiarity of the image is provided. An inspection apparatus according to claim 2 or 3, whe
`rein, for each partial region divided by said inspection region dividing section, the 1 neur
`al network and the 2 neural network are classified by using the first neural network, and
`the type of abnormality is classified by using the second neural network.
`faim 5]
`The 1 neural network performs an operation of performing learning using the inspection t
`argel signal which is normal, The inspection apparatus according te any one of claims if
`o 4, wherein an operation of classifying the inspection target signal inte a normal state a
`nd a normai state is selected after the learning, and an operation of performing learning
`using the normal inspection target signal is performed after an operation of classifying th
`é inspection target signal,
`[Claim 6]
`The 2 neural network performs learning using the partial signal including a known abnor
`mality. An inspection apparatus according to any one of claims L to 5, wherein an operati
`on of classifying the type of abnormality for the partial signal is selected after learning, a
`nd an operation of performing learning using the partial signal including a known abnorm
`ality is performed again after an operation of classifying the type of abnormality for the p
`artial signal.
`
`
`
`[Detailed description of the invention]
`[Technical field]
`fOOO1)
`The present invention relates to an inspection apparatus for determining the presence or
`absence of an abnormality in a signal to be inspected and classifying the type of abnorm
`ality.
`[Background of the Invention]
`(O002}
`Conventionally, a technique for classifying an inspection target signal using a neural netw
`ork has been proposed. For example, Patent Document 1 describes a technique for deter
`mining whether an inspection target is normal or abnormal by classifying information on
`a color to be Inspected using a neural network. In Patent Document 1, it is described tha
`fikis not necessary to use only a normal sample image and an abnormal sample image |
`
`

`

`s used, so that it is possible to easily realize appearance inspection with high inspection
`accuracy. In addition, deterrnination of whether or not to be norrnal or abnormal by a ne
`ural network (or determination of acceptability) is applied not only to color information b
`ul also to a vibration waveform (for example, refer to Patent Document 2).
`[Prior art reference]
`[Patent document]
`
`(O003]
`{Patent document 1]JP 2010-81594
`[Patent document 2]JP 2006-38478A4
`[Suramary of the invention]
`[Problem to be solved by the invention]
`fO004]
`The technique described in Patent Documents 1 and 2 has a cammon point in that an inp
`ut signal to be inspected (a signal corresponding to an image and 4 signal representing a
`vibration waveform} is classified by a neural network, and each of the configurations is u
`sed for classification between normal and abnormal. Therefore, it is possible te determin
`e whether or not there is an abnormality in the inspection target signal, but it is not poss
`ible to classify the abnormality type when the inspection target signal is present.
`fOoos]
`On the other hand, as in the conventional configuration of Patent Decument 1, an abnor
`mal sarnple image is handled in the same manner as a normal sampie image, and a neur
`al network is learned using a normal sample image and an abnormal sample image. Int
`his case, since learning using an abnormal sample image is performed, it is considered p
`ossible to classify types of abnormalities. However, only by learning using an abnormal s
`ample image having a small number of samples, it is difficult to clearly determine the bo
`undary of each abnormality, and further interference due to a normal sample image larg
`er in number than an abnormal sample image occurs, and it is difficult to clearly determi
`ne the boundary of the type of abnormality.
`fO006]
`AS a result, in the techniques described in Patent Documents 1 and 2, the type of abnor
`mality cannot be classified, and when learning including an abnormal sample image is pe
`rformed, there arises a problem that determination accuracy is lowered due to interferen
`ce of a normal sample image. This occurs in the same way even when not only an image
`bub aiso other signals are used as the signal to be inspected.
`(OOo?)
`To provide an inspection device capable of discriminating not only the presence or absen
`ce of an abnormality but also classification of an abnormality, and reducing an amount of
`information required for learning for classifying an abnormality than a conventional const
`tution.
`[Means for solving the problem]
`[O00]
`To achieve the above object, according to the present invention, there is provided a i pr
`ocessing section which includes a 1 neural network which has been learned and classifies
`a Signal to be inspected into a normal state and a non-normal state. A 2 processing unit i
`Ss provided with a 1 processing unit which extracts a partial signal including an area other
`than normal from an inspection target signal which has been classified by a 1 processing
`unit and has been classified by a 2 processing unit, and classifies the type of abnormality
`for a partial signal, and an output unit which outputs a classification result by the 1 proc
`essing unit and the 2 processing unit, and a second neural network is provided. The 2 ne
`ural network is learned using a signal to be inspected including a known abnormality, an
`d is learned by using only the signal to be inspected which is normal.
`[0009]
`It is preferable that the inspection target signal is an image signal of an image obtained
`by imading an inspection object by an image input unit.
`fOOLO]
`In this case, an abnormal image storage part storing an image containing a defect and a
`n image superposition part displaying an image selected frorn an abnormal image storag
`e part on a normal image on a screen of a monitor device are displayed. It is preferable t
`hat the apparatus includes an adjustment unit that adjusts parameters of an image selec
`ted from an abnormal image storage unit according te an instruction from an input devic
`
`

`

`@, and a condition that the 2 neural network classifies the type of the abnormality is adju
`sted by the adjustment unit.
`fOOLL]
`in this case, an inspection region dividing section for dividing an image into partial regio
`ms based on the regularity of an image is provided, and a 1 neural network is used for ea
`ch partial region divided by the inspection region dividing section.
`
`it is preferable to classify a normal and a normal and a non-normal, and classify the type
`of abnorrnality using a 2 neural network.
`(0042)
`In the 1 neural network, it is preferable that an operation of performing Jearning using a
`n inspection target signal and an operation of classifying an inspection target signal into
`a normal state and @ non-normal state after learning are selected, and then, a learning o
`peration is performed by using a nermal inspection target signal after an operation of cla
`ssifying the inspection target signal.
`fOOL3]
`In addition, it is preferable that the 2 neural network performs an operation for performi
`ng learning using a partial signal including a known abnormality and an operation fer cla
`ssifying the type of abnormality for a partial signal after learning, and performs an opera
`tion for performing learning using a partial signal including a known abnormality after an
`operation for classifying the type of abnorrnality for the partial signal.
`fEffect of the Invention]
`fOO14]
`According to the configuration of the present invention, a partial signal including a non-n
`ormal region is extracted from an inspection target signal classified in the 1 precessing s
`ection other than normal, and a type of abnormality is classified in a partial siqna!l of the
`inspection target signal classified in the 2 processing section. Thus, the type of abnormal
`ity is classified inte a range other than normal, and the classification of the type of abnor
`mality is performed locally using the partial signal. As a result, not only is it possible to d
`etermine whether or not there is an abnormality, but also classification of the type of abn
`ormality becomes possible, and moreover, it is possible to reduce an amount of informati
`on necessary for learning for classifying an abnormality.
`[Brief Description of the Drawings]
`[OO15]
`[Fig. LIFIG. 4 is a block diagram illustrating Embodiment 1 :.
`[Fig. ZTE is the figure showing the example of an image same as the above.
`[Fig, SHE is the figure showing the example of a block same as the above,
`[Fig. 4IFiG. 4 is a diagram showing an example of a normal block in the above.
`[Fig. 5IFIG, 4 is a diagram showing an example of a block including a defect in the ahov
`
`Fig. GIFIG. 4 is a Block diagram illustrating Embodiment 2 5.
`[Fig. 7IFIG. 4 is a diagram illustrating the operation of the first ernbodiment ;.
`[Fig. SIFIG, 9 is a diagramillustrating an example of an image to which Embodiment 3 is
`applied ;.
`(Fig. GIFIG. 4 is a diagram showing an example of another image to be applied.
`[Fig. LO]FIG. 4 is a block diagram showing the sarne.
`[Mode for carrying out the invention]
`fOOL6]
`In the embodiment described below, an inspection apparatus in which an image signal re
`lated to an inspection object output by an image input unit is used as an inspection targe
`{ signal and a defect in an appearance of an inspection object is inspected as an abnorm
`ality is exernplified. For example, the Inspection object is a sheet-shaped member, and a
`defect, a flaw, a stain, a color unevenness, or the like is assumed as a defect. However t
`hese defects are merely exarnples, and the technique of the present embodiment is not
`prevented from being applied to the extraction of items which are treated as defects ina
`ppearance inspection such as chipping, cracking, and cracking.
`fOOL?}
`An irnage input unit is assumed to be an imaging unit such as a TV camera, but an imag
`e scanner, a 3 dimensional scanner, or the like may be used. Further, while an example i
`nwhich @ stil image is used is described, a moving image may be used. As an image sig
`nal, a monochrome grayscale image is assumed, bul a color image can be used, Even wh
`
`@ [
`
`

`

`en an image signal is processed, an inspection is performed on a digital image, and there
`fore, when an irnage input unit outputs an analog image, the image signal is converted |
`nto a digital image by performing analog-to-digital conversion,
`fOOL8]
`In the following embodiments, a competitive learning type (self-organizing map = SOM)
`is exemplified as a neural network far classifying a signal to be inspected ; however, othe
`r principles are eR
`
`It is also possible to use work. A neural network is implernented by executing a program
`ona Neurnann type computer. However, it may be realized by using a non-Neumann typ
`e computer having a configuration specialized for a neural network. In the following emb
`odiments, a “neural network" is abbreviated as a "neural network".
`fOOL9]
`(Embodiment 1)
`In this embodiment, as shown in FIG. 1, an image signal of a still image output from ani
`mage input unit 4 that captures an image of an inspection target 5 is used as an inspecti
`on target signal. The signal te be Inspected output from the image input unit 4 is input in
`to the 1 processing unit 1 and the 2 processing unit 2, and the presence or absence of a
`n abnormal defect is determined. The determination result of the I processing unit 1 and
`the 2 processing unit 2 is integrated and output in the output unit 3.
`(0020)
`The 1 processing unit 1 includes a 1 neural network 1O1, and the 2 processing unit 2 incl
`udes a 2 neural network 201. A self-organizing map is used for the 1 neural network 101
`and the 2 neural network 201, and at least an input layer and an output layer are provid
`ed, and a plurality of neurons (modes) are connected to the input layer and the output la
`yer respectively. That is, each neuron of the input layer is coupled to all neurons of an ou
`tout layer, and each neuron of the output layer is coupled to all neurons of the input laye
`mn Also, the strength of the binding is represented by a weighting factor.
`fO021]
`A i neural net 1061 provided in the 1 processing part i classifies the signal to be inspecte
`d into normal and non-normal. On the other hand, the 2 processing unit 2 includes a par
`tial signal extracting unit 22 which extracts 4 partial signal including an area other than n
`ormal from an inspection target signal classified as not normal by the 1 neural net 101, a
`nd classifies the type of the abnormality (defect) in the partial signal extracted in the par
`tlal signal extracting unit 22.
`fO0022]
`The 1 processing unit 1 and the 2 precessing unit 2 include a microcomputer, a DSP (Deg
`ital Signal Processor}, an FPGA (Field Programmable Gate Array}, and a memory as
`main components. In addition to temporarily storing the input image signal, the memory
`is used for storing learning data, which will be described later, and storing the determina
`tion conditions of the 1 neural network TO4 and the 2 neural network 201. In addition, th
`@ 1 processing unit 1 and the 2 processing unit 2 include an interface (not shown) for inp
`utting an image signal, outpuiting a determination result, and inputting a determination
`condition, That is, the interface is provided for connection such as an image input unit 4
`consisting of a TV camera, a monitor device 6 for outputting a determination resull, an in
`put device (e.g. a keyboard and a mouse, a touch panel} 7 for inputting a determination
`condition, and the like. Further, an interface may be provided for communicating with ot
`ner devices.
`(0023)
`The 1 processing unit 1 includes a feature extraction unit 13 that extracts a feature amo
`unt of an image signal output from the image input unit 4. As for the feature quantity, th
`@é density can be used if the image represented by the image signal is a grayscale image,
`and the hue, the saturation, the brightness, and the like can be used if the image is a col
`orimage. Itis also possible to use distributions such as density, hue, saturation, and bri
`ghtness as they are, and also use these frequency distributions as feature quantities.
`fo024]
`Further, by performing DCT (Discrete Cosine Transform) or DFT (Discrete Fourier Tran
`sform)} on Ll images, cornponents of spatial frequencies may be extracted as feature qua
`ntities.
`fo025]
`Although the feature amount may be obtained for an entire image, it is desirable to divid
`
`

`

`e 1 images into a pluraity of blocks which are rectangular regions of n x m pixels, andt
`o obtain a feature arnount for each block. This place
`
`By associating the position of each block with the feature quantity, it is possible to deter
`mine the feature quantity for each location in the image, and thus if is possible to deter
`mine the presence or absence and the type of the defect for each location. Further, by fi
`miting the feature amount to a small region of a block unit, itis easy to extract the featu
`re of each defect.
`(0026)
`In the following, a density normalized for each block in the feature extraction unit 13 is u
`sed as &@ feature amount in the monochrome grayscale image. That is, the average value
`of the densities of the blocks is obtained using the density of each pixel in the block, the
`difference between the density of each pixel and the average value is defined as the ner
`malized density, and the distribution of the normalized density in the block is used as the
`feature value, Here, each block is given a label (which may be a number} so that the pos
`ition in the image can be identified.
`(0027)
`The 1 processing unit 1 includes a learning data storage unit 14 that stores the feature a
`mount extracted by the feature extraction unit 13 fromm an image obtained by capturing a
`n image of the normal inspection target 5 as learning data. The learning data includes th
`e position of the Block. In other words, the feature amount per location in the image is e
`xtracted. Further, whether or not the object 5 is a normal inspection object is determined
`visually by a person. In this case, it is desirable that not only the inspection object 5 but
`also the captured image be confirmed on the screen of the monitor device 6.
`(0028)
`Further, as described above, the 1 processing unit 1 includes the 1 neural network 101, a
`nd the 1 neural network 101 performs learning using the learning data stored in the lear
`ning data storage unit 14. Since the 1 neural net 101 performs learning based on learnin
`q data, the state of the output layer is converged when the inspection target 5 is normal,
`and therefore, a determination criterion for determining normal is determined based ont
`his state.
`fo029]
`In the 1 processing unit 1, when an SOM is used as the 1 neural network 101, a clusterin
`g map in which « rm (, mis a natural number} neurons are arranged in a matrix (2 dime
`nsional lattice pattern) as an output layer of the 1 neural network 101 is used.
`,O030]
`When an appropriate data is inputted Into the 1 neural network 104 which has been lear
`ned Qvhen data is given to the input layer}, if a neuron which is not a boundary region a
`mong the clustering maps (neurons adjacent in all directions} is ignited, the data is judg
`ed as normal.
`fO031}
`In addition, when the neurons in the boundary range fire, the Euclidian distances betwee
`n the set of weighting factors (weighted vectors) obtained by learning by the ignited neu
`rons and the weighted vectors of the adjacent neurons are respectively calculated. Furth
`er, the maximum value of the Euclidean distance between the ignited neurons and each
`of the adjacent neurons is set as a threshold value. Here, when the distance between the
`data given toe the 1 neural network 101 and the weighted vector of the ignited neuron ex
`ceeds the threshold value described above, it is deterrnined that the data is norrnal.
`[O032}
`As described above, the I neural net 101 can be selected from an operation (learning op
`eration) of performing learning based on learning data and an operation (inspection oper
`ation} of determining presence or absence and type of a defect in an actual inspection ta
`rget 5 after compistion of learning. The criteria determined by the learning operation are
`used in the inspection operation.
`[0033]
`The i processing unit 1 includes a determination storage unit 15 that stores the determi
`ned criteria, and includes a determination unit 102 that determines an output result of th
`@ 1 neural network 101 using the determination criteria stored in the determination stora
`ge unit 15, Since the criterion for the output of the 1 neural network 1O4 is determined u
`sing only the learning data generated from the normal inspection object 5, the determina
`tion unit £02 is configured to determine the output of the first neural network 10.
`
`

`

`it is determined whether or not the output of the L neural network 101 indicates normal.
`In other words, the 1 neural network 101 and the determination unit 102 constitute an a
`bnormality detection unit 10 that determines whether or not the vehicle is normal or nor
`mal.
`(0034)
`The determination result of the abnormality detection unit 10 is output from the 1 proces
`sing unit 1 through the result deterrnination unit 16. The result determination unit 16 rec
`eives a determination result of whether the determination result is normal or normal by ¢
`he determination unit 102, and cutouts a normal determination result and a position of a
`block other than normal.
`fO035]
`On the other hand, as described above, the 2 processing unit 2 includes the partial signal
`extracting unit 22. The partial signal extraction unit 22 extracts a block determined to be
`net normal by the 1 processing unit 1 from an image represented by an image signal inp
`ul from the image input unit 4. In other words, a partial signal which is a part of an imag
`e signal is extracted.
`(0036)
`Similarly to the 2 processing unit 1, the T processing unit 2 includes a feature extraction
`unit 23 and a learning data storage unit 24. A feature extraction unit 23 extracts a featur
`e amount for a block extracted by the partial signal extraction unit 22. Here, since the bi
`ock extracted by the feature extraction unit 23 is a block determined to be not normal in
`the 1 processing unit 1, there is a high possibility that the block includes a defect (abnor
`maliby}. Therefore, if a defect (abnormality) included in a block extracted by the partial si
`gnal extraction unit 22 is known, a known defect is associated with the learning data stor
`ed in the learning data storage unit 24 as a category.
`fO037]
`in the 2 neural network 201 provided in the 2 processing unit 2, learning is performed us
`ing learning data in which known defects are associated. Similar toe the 2 neural network
`161, the 1 neural network 201 can select a learning operation and an examination opera
`tion, and learning is performed in the learning operation to determine a criterion for the
`Known defect,
`[0038]
`Here, in the case where an SOM is used as the 2 neural network 201, a clustering map in
`which I x 7 (1, } is a natural number} neurons are arranged in a matrix form is used as a
`nm output layer of the 2 neural network 201. In the 2 neural network 201, the category as
`signed to the neuron that has ignited on the clustering map is defined as the category of
`the input data.
`[0039]
`When learning of the 2 neural network 201 Is performed, learning is performed using a p
`ortion determined to be abnormal by a person in advance by visual inspection and a clas
`sification result of a category of a defect of an abnorrnal portion. Alternatively, an actual j
`nspection is performed, and a portion determined to be non-normal in the 1 neural netbw
`ork IO1 Is learned by visually using the classification result of the category of the defect.
`
`[0040]
`When the 2 neural network 201 performs learning based on learning data whose cateqor
`y is a known defect, a state represented by a neuron of an output layer converges for ea
`ch category. Since neurons of an output layer form clusters corresponding to categories
`of learning data, categories can be classified by assigning categories to each cluster.
`[OO44]
`The 2 processing unit 2 includes a determination storage unit 25 that stores a determina
`tion criterion and a cluster classification unit 202 that associates a cluster of neurons in a
`n output layer of the 2 neural network 201 with a categary. Since the state of the output
`of the 2 neural network 201 is classified into the category of the known defect by the clu
`ster classification unit 202, the abnorrnality classification unit 20 is constituted by the Zn
`éeural network 201 and the cluster classification unit 202. The 2 processing unit 2 include
`S a map storage unlit 26 that stores the types of defects classified by the cluster classifica
`tion unit 202. Therefore, by registering the type of defect in the map storage unit 26, if i
`Ss possible to output the type of defect classified in the cluster classification unit 202.
`
`

`

`fO042]
`As described above, the 1 processing unit 1 classifies normal and non-normal, and the 2
`processing unit 2 extracts a partial signal (block) which is an area corresponding to a def
`act for an image signal that is determined to be not normal in the 1 processing unit 1, an
`d classifies the type of defect for this block. As a result, it is possible to improve the accu
`racy of classifying the normal and the defect as compared with the case of classifying the
`normal and the defect by one neural net.
`[0043]
`In other words, since the learning data for performing the normal classification can be co
`lected sufficiently more than the learning data for classifying the defects, the 1 process
`ng unit 1 can accurately classify the learning data as normal and normal. Next, in the 2 p
`recessing unlt 2, the types of defects are classified only within a range other than normal
`by excluding the normal inspection target signal, so that even when the learning data is
`relatively small, the boundaries become clear and consequently, the accuracy of classific
`ation is increased. Further, since attention is paid to a block which is not normally deter
`mined as a signal to be inspected other than normal, the type of the defect is classified |
`na jocal area rather than an entire image, and thus, the accuracy of classification can be
`improved by reducing an amount of information of information which is not necessary for
`classification.
`fO044]
`In the above example, in the 1 processing unit 1 and the 2 processing unit 2, the learnin
`g operation is performed only 1 times before the inspection operation.
`(0045)
`For the 1 processing unit 1, learning data may be newly generated from an inspection ta
`rqet signal obtained from a normal inspection target 5, and the number of learning data
`may be increased to perform additional learning of the 1 neural network 101. If the dete
`rmination criterion stored in the determination storage unit 15 is updated by performing
`the additional learning of the 1 neural network 101, the accuracy of separating the norm
`al and the normal is improved, and consequently, the accuracy of separating the iInspecti
`on target 5 including the defect is improved.
`fO046]
`Further, for the 2 processing uri 2, it is possible to newly generate the learning data fro
`m the inspection target signal obtained fromthe defective inspection target 5, increase t
`he number of the learning data of each category, and qenerate the learning data of the n
`ew category. If the additional learning of the 2 neural network 201 is performed by such
`learning data, fh is possible to update the determination criteria stored in the determinati
`on storage unit 25 and the category registered in the map storage unit 26. In other word
`s, it is possible to Increase the accuracy of classifying the types of defects, or to classify
`new types of defects.
`fO047]
`FIG. 2 shows an operation example of the above-described configuration, Each of FIGS.
`2 {a} to 1 (d) shows an image (an image output from the image input unit 4) including a
`block determined to be not normal in the first processing unit 1. In addition, each of FIG
`S. 3 (a) - (d) corresponds to each of FIGS. 2 (a} - (d}, and shows an exarnple of blocks
`Bi-8 6 determined to be non-normal. That is, when an image as shown in FIG. 2 (a) -
`{d} is input to the 1 processing unit 1, the partial signal extracting unit 22 of the 2 proce
`ssing unit 2 extracts an area corresponding to blocks B 1 - B 6 shown in FIGS. 3 {a} -
`{d}. In the extracted blocks B 2 to B 6, the types of defects are classified in the first proc
`essing unit 2.
`[0048]
`Here, a learning data as shown in FIG. 4 may be stored in the learning data storage unit
`14 of the 1 processing unit 1 with respect to a block in which the inspection target 5 is n
`ormal, as shown in FIG. 4. In addition, in the learning data storage unit 24 of the 2 proc
`essing unit 2, the learning data as shown in FIG. 5 ray be stored in the block in the cas
`e where there is a defect in the inspection target 5 (each of the partial diagrams in FIG,
`5 represents an image of a block including a defect).
`(0049)
`In this embodiment, since only the learning data of the defect shown in FIG. 5 is stored ij
`n the learning data storage unit 24 of the 2 processing unit 2, it is possible te extract bio
`cks B5 and 8 6 as shown in FIGS. 3 cand 5d. FIG. 4 shows, on the contrary, normal lea
`rning data as shown in PIG, 5,
`
`

`

`When learning data of a defect defect is used together with learning data, it becornes po
`ssible only to extract blocks B I te B 4 in which defects are blurred as shown in FIGS. 3
`a and 4 b. In other words, it is difficult to extract blocks B 5 and B 6 in which defects suc
`fas shown in FIGS. 3 c and 3 d are blurred.
`fO050]
`In other words, in the present embodiment, classification of normal and non-normal is p
`arformed in the 1 processing unit 1, so that the boundary between normal and normal ca
`n be clearly separated, and extraction of blocks including defects can be accurately perfo
`rmed. In addition, since the 2 processing unit 2 determines the type of defect only for th
`@ block determined to be not normal, the boundary for the type of defect is clarified even
`when the learning data concerning the defect is small beacause the normal range is exclu
`ded from the information to be classified. In other words, it is easy to classify the types
`of defects.
`fOO51]
`in this embodiment, a case in which the inspection tarqet signal is an image signal has b
`een described, but the inspection target signal may be an electric signal including other i
`nformation. For exampie, when the test object 5 generates sound or