JP2987670B2 - Molded product shape error judgment method by image processing - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for automatically judging a difference between a normal product and an abnormal product of a molded product such as a tooth of a slide fastener by image processing.

[0002]

2. Description of the Related Art In general, a slide fastener is provided with a plurality of teeth which are planted along a pair of tapes, and the teeth which are planted on the pair of tapes are engaged with and disengaged by a slider. The metal slide fasteners
It is stamped and formed one by one by a molding machine. In this manufacturing process, an abnormality of the tooth shape may appear due to an abnormality of the molding machine or the like, and it is necessary to always inspect the tooth shape. Conventionally, for example, by measuring the shape and dimensions with a micrometer, or observing the product shape using a loupe, etc., it is possible to detect processing defects such as irregular dimensions, missing tooth shapes, bending, etc. Failure judgment is being performed.

[0003] This method of determining the tooth shape is easily affected by personal experience such as the skill and experience of the inspecting operator, such as handling of a micrometer and visual observation, and the inspection standards are unstable and unreliable. It has become something.

On the other hand, a CCD camera, a CRT, and a computer for image processing are provided.
A shape determination technology based on image processing that uses a computer to determine various shapes based on the image information is known. By applying this technology to the determination of defective tooth shape, the inspection standard is fixed and reliability is improved. I do.

[0005]

However, in the conventional method of determining a shape defect by image processing, an object is limited, and characteristic amounts (area, moment, dimensions of a specific portion, color, etc.) specific to the object are used. A defect determination device has been created by a program created by a technician familiar with image processing. For this reason, the created shape defect judgment device can be used only for specific objects and is not only inferior in versatility, but also requires a specialized engineer to rewrite the program even if the specification of the object is slightly changed. , Had become a major problem in development. In addition, in order to increase the accuracy rate of the defect determination, it is desirable to use as many feature amounts as possible and to make a comprehensive determination. However, the increase in the feature amounts increases the amount of calculation and slows down the determination result. There was a problem to say.

Accordingly, the present invention solves the above-mentioned problems by applying a hierarchical neural net backpropagation algorithm already known in the fields of voice recognition and character recognition to recognition of a shape defect of a molded product. It is an object of the present invention to provide a molded article shape judging method by versatile image processing for the purpose.

[0007]

In order to realize a versatile judgment method of a molded article, a judgment processing which is not affected by positional deviation and rotation, a common processing method for a plurality of molded articles, and a calculation time are required. It is necessary at least to select a shape determination parameter and to have a shape learning function. As means for realizing these three conditions, a contour shape of a specific molded product is used as a reference shape, three or more points arbitrarily selected from the reference shape are used as reference points, and a gradient detected from the reference point is used. The features that are invariant to position, rotation, etc., such as differences, perimeter differences, distances, etc. are used as the reference features, and the contour shape of the molded product whose shape is determined is used as the input shape, and the input shape is divided into equally spaced parts. And selects three or more points from the input parts corresponding to the reference part, and calculates a gradient difference corresponding to the reference feature,
When a feature that is invariable in position and rotation such as a perimeter difference, a distance, and the like is set as an input feature, three or more sites are detected from the input site so that the input feature and the reference feature are most similar to each other. By performing coordinate transformation such that the part and the reference part overlap, the reference shape and the input shape are aligned, and then, from the aligned input shape, to a plurality of evaluation regions arbitrarily set on the reference shape. Extract the included part, and then
X- and Y-coordinate data near the curve by function series approximation
Similar, the coefficient is used as a parameter to determine the shape.
And extracted, the parameter as input data of the hierarchical neural network, the molded article shape determining method by the image processing and performs the molded article shape determined by performing learning and determination of a molded article shape.

[0008]

[Operation] By applying the present invention, it is possible to judge the shape defect of various molded products by utilizing the shape learning ability without being affected by rotation or positional deviation, so that image processing can be performed even for an object or a slight specification change. A defect determination device can be created without a technician familiar with the method.

[0009]

[Example] As shown in FIG.
A magnifying unit 3 having a magnifying lens 2, a transparent sample holder 4 on which the molded article A is placed, a light source 5 for irradiating light from below the sample holder 4, an image processing computer 6, and a CRT 7. An image processing apparatus is formed, and the molded article A is photographed by the CCD camera 1, and the image information is input to the image processing computer 6, displayed on the CRT 7, and processed at the same time. The molded product A is a tooth of a slide fastener.

(Processing procedure for determining tooth shape) Input of reference shape Reference shape for position adjustment and designation of evaluation area (A in FIG. 2)
Is input to the image processing computer 6. The reference shape A is
It is made up of a series of N points indicating the outline of the shape, and has an x coordinate value and a y coordinate value, respectively. The normal tooth shape is usually a tape attachment part 10, a narrow base 11, a narrow middle part 12, a wide tip part 1.
3 has one side straight portion 14, one side portion 15 substantially perpendicular to the one side straight portion 14, and a tip arc portion 1.
6. It is composed of the bent other side surface portion 17 and the other side linear portion 18. On this reference shape, three or more reference portions a, b, and c, which are reference positions for alignment, are designated so as not to be aligned on a straight line. Also, the evaluation areas B, C, and D to be evaluated are designated as quadrangular areas having an appropriate area.

Input Shape Input The input shape (E in FIG. 3) to be evaluated is input to the image processing computer 6 by the CCD camera 1.
Similarly to the reference shape A, the input shape E is also composed of a sequence of N points having X coordinate values and y coordinate values indicating the outline of the shape.

The input shape is registered on the reference shape. The input shape E that has been input is similar in appearance to the reference shape A, but has different x and y coordinate values due to deviations in position, rotation, etc., and finds an evaluation area to be compared directly on the input shape. It is impossible. Therefore, the reference shape and the input shape are aligned, and a point sequence on the input shape included in the evaluation area specified on the reference shape is set as a shape evaluation target.

The registration based on the above-mentioned partial shape is performed by newly inputting the information of a, b, and c in FIG. 2 previously specified in the reference shape A stored in the image processing computer 6. A combination of points that minimizes the evaluation function f determined from the information of an arbitrary point of the input shape B is searched from the input shape, and coordinate conversion is performed such that the found point and the point on the reference shape overlap. This is performed by determining the parameters φ, u, and v and performing coordinate transformation on the input data so that the two partial shapes match.

The evaluation function f includes, as information that does not change due to displacement or rotation on the reference shape, distances between reference portions (FIGS. 4, a, b, and c) on the reference shape (FIG. 4, L1),
Focusing on reference features such as a perimeter difference (FIG. 4, L2) and a gradient difference (FIG. 4, α1), these reference features and an arbitrary input portion on the input shape (FIG. 3, d, e, f) (Fig. 3, L3), circumference difference (Fig. 3, L4), gradient difference (Fig. 3, α
The square of the difference of the input features such as 2) is calculated by combining all the parts and added. This evaluation function f
Is 0 if they match, otherwise it has a value greater than 0. Therefore, change the input part (FIG. 3, d, e, f) appropriately and evaluate the minimum value of the evaluation function f. Thus, a part that matches the condition with the reference part (FIGS. 4, a, b, and c) on the reference shape can be found on the input shape.

When this evaluation function f is expressed by a calculation formula, an evaluation function f = (Σ (distance L between reference parts 1−distance L between input parts L
3) ² × (weight 1) + (Σ (perimeter difference L2 between reference parts−perimeter difference L4 between input parts) ² × (weight 2) + (Σ (gradient difference α1-input part between reference parts) Gradient difference between α2) ² × (weight 3).

By evaluating the minimum value of the above-mentioned evaluation function f, a reference portion (a, b, c in FIG. 4) on the reference shape is obtained.
And the condition on the input shape (FIG. 3, a,
b, c) can be detected. The parameters φ, n, and v of the coordinate transformation shown in Equation 1 are determined by the least square method so that these three points are the best, and the coordinate transformation of Equation 1 is performed on all the points of the input shape. B and the reference shape A can be aligned.

[0017]

(Equation 1)

The coordinate transformation parameters φ, n, and v represent the center coordinates (smx [p _i ], smy [p _i ]) corresponding to the reference parts (a, b, c) on the reference shape, respectively. i
= 1 to 3, p ₁ , p ₂ and p ₃ correspond to a, b and c),
Input sites on the detected input shape (Fig. 3d, e, f) the respective center coordinates corresponding to (mx [k _i], my
[K _i ], i = 1 to 3, k ₁ , k ₂ , and k ₃ are d, e, f
Is calculated by the following equation 2.

[0019]

(Equation 2)

FIG. 5 shows the input shape E and the reference shape A superimposed by these operations. The accuracy of the registration slightly changes depending on the weight of the evaluation function.

Extraction of Evaluation Parameters from Partial Evaluation Area When determining (evaluating) a shape error, it is more efficient to evaluate with as little data as possible within a range where shape identification is possible. Therefore, the area to be evaluated is limited to a partial area where shape abnormality is most likely to appear as shown in B, C, and D of FIG. 5, and the area is evaluated. The x-coordinate and y-coordinate data contained in this evaluation area are represented by x [i] and y [i], respectively, and are approximated by a known least square method using a function series as shown in the following Expression 3, and the coefficients are evaluated. Extract as a parameter.

[0022]

(Equation 3)

By using the evaluation parameter extraction based on this function approximation, parameters for shape evaluation can be greatly reduced.

Tables 1 and 2 and FIG. 11 show the following equation 4 as the above approximate function.

[0025]

(Equation 4)

Approximation by the 4th and 7th order polynomials with respect to i as shown in the following equation 5

[0027]

(Equation 5)

Shows the correct answer rate of the judgment result of the tooth shape of the slide fastener when approximated by the Fourier series approximation, and shows that these evaluation parameters are effective for the judgment of the molded article shape. .

[0029]

[Table 1]

[0030]

[Table 2]

Judgment of Input Shape Based on Evaluation Parameters Using the evaluation parameters extracted as described above, the computer makes a learning judgment using an algorithm of a hierarchical neural network model having a three-layer structure as shown in FIG. Let In other words, the process is divided into a learning phase and a decision phase,
In the learning stage, the evaluation parameters are given to the neural network, and at the same time, teacher information indicating whether the shape is normal or bad is input. At the judgment stage, the computer will give an appropriate judgment even without teacher information. By utilizing the learning ability of the neural network, a general-purpose shape determination can be realized. (For neural networks, see Neuro Computer Basics, Kaoru Nakano, Corona, 1990, first edition, and many others) On the other hand, based on the evaluation parameters extracted from the shape, set the basic values by the program to determine the shape However, the conventional method is inferior in versatility.

As shown in FIG.
(10) An input layer 20 of 3 units, an intermediate layer 21 of 100 units, and an output layer 22 of 8 units.
The evaluation parameters extracted from the input shape are input to the units of the input layer 20, respectively. In the learning stage, the unit of the output layer 22 assigns a defective pattern, gives an input shape, and sets the corresponding defective unit to 1 and the others to 0. The learning method uses a back propagation algorithm. In the determination stage, the input shape is determined by simply giving an evaluation parameter extracted from the input shape to the input layer 20 and comparing the response of the output layer at that time.

By performing the above-mentioned processes, the shape of a general-purpose molded product can be determined. Table 1, Table 2, and FIG. 11 show the determination results for the normal product sample (FIG. 7), the sample that is too high in the mountain (FIG. 8), the sample that is too low in the mountain (FIG. 9), and the sample with the defective chip (FIG. 10). is there. Although there is a difference in the determination result due to the difference in the approximate function, a shape determination of 90% or more can be realized.

[0034]

According to the present invention, since the contour shape of a specific molded product is used as a reference shape and the positioning is performed based on the partial shape, it is possible to perform the judgment processing relatively fast without depending on the position shift rotation. Since the evaluation target is limited to a specific evaluation area and the evaluation parameters are reduced by function approximation, the processing amount is small and the judgment processing can be performed at high speed. Since the learning ability of the neural network is used, various molded products are stored and determined, so that shape determination with excellent versatility can be performed.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an image processing apparatus.

FIG. 2 is an explanatory view of a basic shape of a correct molded product.

FIG. 3 is an explanatory diagram of an input shape obtained by treating image information of a molded product.

FIG. 4 is an explanatory diagram of a reference point of a basic shape, a distance, a circumference, and a gradient difference.

FIG. 5 is an explanatory diagram of a state in which input shapes of basic shapes are aligned.

FIG. 6 is a diagram illustrating the structure of a neural network.

FIG. 7 is an explanatory view of a normal tooth shape.

FIG. 8 is an explanatory diagram of a defective shape of a working tooth.

FIG. 9 is an explanatory diagram of a defective shape of a working tooth.

FIG. 10 is an explanatory diagram of a defective shape of a working tooth.

FIG. 11 is a table showing a learning rate and a square error amount in four classes.

Claims (1)

(57) [Claims] 1. A contour shape of a specific molded product is set as a reference shape, and three or more points arbitrarily selected from the reference shape are set as reference points, and a gradient difference and a circumferential length difference detected from the reference point are determined. , Features that are invariant to position, rotation, etc.
In addition, the contour shape of the molded product whose shape is determined is set as an input shape, and a portion obtained by dividing the input shape at equal intervals is set as an input portion, and three or more portions are selected from the input portions in correspondence with a reference portion. Then, when a feature such as a gradient difference, a perimeter difference, a distance, and the like corresponding to the reference feature and a feature that is invariable to rotation are input features, three or more points are input so that the input feature and the reference feature are most similar. By detecting from among the parts and performing coordinate transformation such that the detected part and the reference part overlap,
Aligns the reference shape and the input shape, from then aligned input shape, extracts a portion included in the plurality of evaluation areas set arbitrarily on the reference shape, X of the site
Curve approximation of coordinate and Y coordinate data by function series approximation
And use the coefficient as a parameter to determine the shape.
Extracting Te, the parameter as input data of the hierarchical neural network, the molded article shape determining method by the image processing and performs the molded article shape determined by performing learning and determination of a molded article shape.