JPH04104048A - Image processing apparatus - Google Patents

Abstract

PURPOSE:To make it possible to identify the kinds of the contents of abnormalities by providing a neural network for outputting identification data which are allocated to the characteristic data of the image of a body having the normal appearance, which is closest to the image of a body to be inspected when the characteristic data are inputted, and the image of a body having the various kinds of defects on the appearance. CONSTITUTION:In a neural network 50, the characteristics of the image of a body having the normal appearance and the characteristics of the image having the various kinds of defects are made to learn beforehand. The characteristic data are extracted from the image of the body to be inspected with a characteristic extracting means 15 and inputted into the neural network 50. The pattern matching between the image of the appearance of the body to be inspected and the learned characteristics is executed in the neural network 50. As a result, the identification data indicating the learned image close to the image of the appearance of the body to be inspected are outputted from the neural network 50. Therefore, the contents of the defect in addition to the presence or absence of the defect can be automatically identified. Since the identification for the presence or absence of the defect and the identification of the contents of the defect are performed at the same time in the neural network 50, the time for the visual inspection can be shortened.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Field of Application 1] The present invention relates to an image processing apparatus that uses image processing to identify the presence or absence of a defect such as a scratch or color unevenness on the surface of an object and the content of the defect.

[Prior Art 1] Conventionally, an image processing apparatus is known that performs an external appearance inspection of a product by capturing an image of the surface of the product using a light receiver such as an image sensor and processing an image signal obtained as a result of the image capture.

Conventionally, the visual inspection method often used in this type of device is a method that determines whether there is an abnormality in the external appearance by comparing the brightness level of each pixel indicated by the image signal with a threshold value. be.

[Problems to be Solved by the Invention] However, although conventional devices of this type can detect abnormalities in the appearance of products, they have not been able to identify the contents of the abnormalities. For example, when inspecting the appearance of textile products, the objects of the appearance inspection include thread removal, dirt adhesion,
There are uneven colors in the resin or dyeing.

Conventionally, when an abnormality in appearance is detected by image processing using an image processing device (visual inspection device), an inspector must visually confirm the shape and content of the abnormality in the area where the abnormality has occurred.

SUMMARY OF THE INVENTION In view of the above, an object of the present invention is to provide an image processing device that is capable of identifying the type of abnormality when visual inspection is performed by image processing.

[Means for Solving the Problems] In order to achieve such an object, the present invention provides first feature information indicating characteristics of an image of an object having a normal appearance and an image of the object having various defects in appearance. The second feature information indicating the feature of A neural network outputs identification information corresponding to the first feature information or the second feature information, and extracts the third feature information to be input to the neural network from an image of the object to be inspected. It is characterized by comprising a feature extraction circuit.

[Function j] The present invention focuses on the fact that the neural network has a pattern matching function, and learns in advance the characteristics of images of the normal appearance of objects and the characteristics of images with various defects, and A neural network performs pattern matching between image features and learned features about the appearance of an object. As a result, the neural network outputs identification information indicating a training image that is close to the image of the external appearance of the object to be inspected.
In addition to the presence or absence of a defect, the content of the defect can also be automatically identified. Furthermore, since the neural network simultaneously identifies the presence or absence of a defect and identifies the content of the defect, the time required for visual inspection is reduced compared to the conventional method.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

FIG. 1 shows the circuit configuration of an embodiment of the present invention.

In FIG. 1, the line sensor 10 performs reading scanning perpendicular to the conveyance direction of the product to be inspected, as shown in FIG. In this embodiment, the number of pixels read by the line sensor 10 in the main scanning direction is set to 10 for explanation.

When using a line sensor 1 whose number of read pixels is on the order of several hundred to several thousand, it is preferable to sample a necessary number of image data from the output signal of the line sensor 1O in accordance with the inspection accuracy.

The serial/parallel converter 20 converts the analog image signals serially received from the line sensor IO into parallel image signals for each pixel column.

The prediction filter 30 inputs image data for each column in time series, calculates a predicted value of the image data for each column, and calculates a prediction error between the latest measured image data and the predicted value. .

Calculation processing of predicted values and prediction errors will be explained.

When image reading scanning starts from a certain time T on the product to be inspected while it is being conveyed, and main scanning is performed m times (10 times in this example), the m+1st main scanning and the reading at that time are performed. The predicted brightness level X1 at time 11 pixel position j. , are actually measured luminance levels X1-1. Using i to X + -IIJ, it is expressed by the following formula.

X +, 'r = a + 'L - + J + ax 'X knee z, j ""am'Xi -m, jΣ am'
is the actually measured brightness level x1 at that time and the predicted value X1. are predetermined so that they almost match.

Actual image data X1 of the product to be inspected at time i. The difference between t and the predicted value x1J calculated by the above equation (1), or the square value of the difference, is called a prediction error.

The prediction filter 30 calculates the above-mentioned budget error for each column in response to one reading scan of the line sensor 10, and outputs the calculated prediction errors, respectively. Since a well-known circuit can be used for the prediction filter 30, detailed explanation will be omitted.

The statistical calculation circuit 40 calculates the next feature amount for the prediction error found for each column by the prediction filter 30. In this example, six statistics of prediction error: (1) average, (2) maximum value, (3) minimum value, (4) variance, (5) skewness, and (6) kurtosis are used to represent image features. Use as a feature parameter.

The neural network (also referred to as an associative memory or neural network) 50 includes an input layer 51 . A hierarchical neural network consisting of an output layer 53 and an intermediate layer 52 that performs pack propagation is used.

The input layer 51 inputs six types of feature amounts, which will be described later. middle layer 5
2, a similarity comparison is made between the learned feature amount and the input feature amount. Output layer 53 outputs the identification information of the feature closest to the human-powered feature.

The neural network 50 has previously learned and stored feature quantities of images of objects with normal appearance, feature quantities of images including various defects, and identification information corresponding to each feature quantity. In this embodiment, an identification code representing a 3-bit numerical value "0" to "7" is used as the identification information.

In this example, defects are identified as (1) normal (2) large defect in the vertical (parallel to the conveyance direction) direction... Thread removal is a small defect in the vertical direction, a large defect in the horizontal direction, a large defect in the horizontal direction, a large circle DefectsSmall circular defects 8)Others: 8 types of defects: falling, thread loss, falling, stains, adhesion of dust, uneven coloring.

The circuit operation of such an image processing device will be explained next.

When the line sensor 10 performs ten reading scans on the product being inspected while it is being transported, the prediction filter 30 calculates the prediction error at the current moment for each column.

The statistical calculation circuit 40 calculates six types of feature quantities regarding the prediction error from the prediction error of each column, and inputs them to the neural network 50. In the neural network 50, a comparison between the input feature amount and the previously learned feature amount is performed in the intermediate layer 52, and the signal representing the input feature amount is converted into an identification signal indicating the type information of the closest feature amount. .

As a result, the IOs sampled from the product under inspection
If the 10 image data include, for example, a large defect image in the vertical direction, the neural network 50 outputs 8 pieces of defect information representing the identification code "2".

Hereinafter, every time an image is read and scanned in the line sensor 10, the prediction filter 30 and the statistical calculation circuit extract the features of the image data group read between the current time and the reading time 10 lines before that time. Done at 40. Furthermore, based on the extracted features, the neural network 50 performs processing to identify the presence or absence of a defect and the content of the defect in the line sensor 1.
This is done every 0 read scan.

In addition to this embodiment, the following examples are given.

1) In this embodiment, the defect contents are identified using various statistical feature amounts regarding prediction errors, but the magnitude of the prediction error may be used as the feature amount indicating the defect contents.

In this case, the prediction errors for each column are summed up by the statistical calculation circuit 40, and as shown in FIG. 3, the neural circuit 50 determines which type of threshold range the total value of the prediction errors falls within.

Note that at this time, if the moving average of the predicted values obtained through multiple reading scans is used as the input value to the neural network 50, the accuracy of identifying the defect content can be further improved.

2) In this embodiment, the multiplication coefficients a1 to a in the prediction calculation formula (1) to be executed in the prediction filter 30 are set in advance, but the values of the multiplication coefficients a and -a can be set variably. When using a prediction filter, the number of multiplication coefficients can also be variably set by performing the following arithmetic processing.

That is, the weighting coefficient ak is expressed by the following equation (2).

am 1. lsw=ai+o+a+C1'CL
, a Xt, j'Xt-mJ (k = 1.2. ・
-m) (2) ak□ indicates the weight after update, and akoll indicates the weight before update. α is an appropriate fixed coefficient and preferably has a value of around 0.2. a. 14 is the initial value “
O'', the weighting coefficients are changed in the following classes so that (X, , -X, , )'' becomes the minimum (a value close to the numerical value "O").

Next, the prediction filter 3 selects N (a finite number) consecutive observed values that do not have any defects from several tens to hundreds or more.
Store it within 0. Then, set an appropriately small value m and use N observed values until all the coefficients meet the convergence condition, for example, until the absolute value of the difference between before and after updating the weighting coefficients becomes 1O-3 or less. Then, the above-mentioned iterative formula (1) and formula (
Execute 2).

After the coefficient values converge, the final prediction error FPE (lIll), which is a function of m, is calculated using the following equation.

FPE (+e) = (N++++÷1)/fN-m-
1)・(1/ (N-11) ) Σ (X, -
X, )" (3) Next, increase the value of m by one and repeat the above. In other words, find the weighting coefficient from N observed values and calculate FPE(m). Repeat this operation one after another. Proceed while increasing m.Then, find m that minimizes FPE.The m at this time is the optimal order m, and the weighting coefficient at this time is also used.Such arithmetic processing is used as a prediction filter. 30, if done automatically, the weighting coefficient a
, ~a, are set.

[Advantageous Effects of the Invention 1] As explained above, according to the present invention, it is possible to simultaneously identify not only the presence or absence of defects on the surface of the object to be inspected, but also the details of the defects, which shortens the visual inspection time. Effects can be obtained.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the circuit configuration of an embodiment of the present invention, Fig. 2 is a perspective view showing the installation position of a line sensor in an embodiment of the invention, and Fig. 3 is identification of defect content in another embodiment of the invention. FIG. 3 is a waveform diagram for explaining processing. 10 Lyso Comp 5... Inspection object, 10... Line sensor, 15... Feature extraction circuit, 20... Serial/parallel converter, 30... Prediction filter, 40... Statistical calculation circuit , 50... Neural network. Eishiki Figure 2

Claims (1)

[Scope of Claims] 1) First feature information indicating features of an image of an object having a normal appearance, second feature information indicating features of an image of the object having various defects in appearance, and the first feature information. and the identification information respectively assigned to the second feature information is learned in advance, and when the third feature information is input, the identification information corresponding to the first feature information or the second feature information closest to the third feature information is learned. An image processing device comprising: a neural network that outputs information; and a feature extraction circuit that extracts the third feature information to be input to the neural network from an image of the object to be inspected. .