JPH04142412A - Defect inspecting device - Google Patents

Abstract

PURPOSE:To properly judge the grade for each type of a defect by providing a feature extracting circuit extracting the feature quantity of the defect from the image signal of a detector and neural nets judging and outputting the type and grade of the defect based on the input feature quantity. CONSTITUTION:The surface of an inspected object 5 conveyed in the arrow direction 7 is image-picked-up by a detector 1, and the feature quantity of a defect 6 is extracted by a feature detecting circuit 2 from the image signal. The feature quantity (X1-Xn) from the circuit 2 is inputted to the first neural net 3, the type of a defect is judged to generate an output (Y1-Ym). The feature quantity (X1-Xn) from the circuit 2 and the judged output (Y1-Ym) of the type of the defect from the neural net 3 are inputted to the second neural net 4, and the type of the defect is judged for each type of the defect to generate an output (Z1-Zl) is generated. The neural nets 3, 4 are constituted of multiple neurons, and weight parameters for individual inputs are determined by learning.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Application Field) The present invention is a defect inspection device that photoelectrically scans and images the surface of a strip-shaped object such as a steel plate or an aluminum plate to check for defects on the surface. In particular, the present invention relates to technology that uses neural networks to process information for determining the type and grade (degree) of defects.

(Prior art) In conventional general defect inspection equipment, the type and degree of defects are determined by statistically processing extracted feature quantities of defects or by using branching logic. . When introducing a neural network to information processing for this determination, the system configuration shown in FIG. 4 is easily thought of.

In FIG. 4, the detector 1 photoelectrically scans and images the surface of a strip-shaped inspection object 5 that is conveyed in the direction of an arrow 7. The feature extraction circuit 2 extracts the feature amount of the surface defect 6 of the inspection object 5 from the image signal of the detector 1.

The first neural network 3 determines the type of defect using the extracted feature values as input, and outputs KD, ~KD for each type.
, emit one of the following. The second neural net 4 is
The degree (grade) of the defect is determined using the extracted feature quantity as input, and one of the outputs GD and -G D t for each grade is output. The weight parameters of the first and second neural networks 3.4 are determined by learning.

(Problems to be Solved by the Invention) In the case of the system configuration as shown in FIG. 4, there are problems in that learning takes a very long time and the degree of defects cannot be appropriately determined.

When a skilled inspector performs a visual inspection, the method of grading varies slightly depending on the type of defect, but the system shown in Figure 4 does not have an element that connects type and grading in this way, so it is not possible to properly It is not possible to make accurate grade judgments.

This invention was made in view of the above-mentioned conventional problems, and its purpose is to provide a defect inspection device that allows neural net learning to converge quickly and that can appropriately grade each type of defect. The purpose is to carry.

[Structure of the invention]

(Means for Solving the Problem) Therefore, in the present invention, the feature quantity extracted by the feature extraction circuit is input to the first neural network to determine the type of defect, and the feature quantity and the The second neural network is configured to input the determination output of the first neural network and the second neural network to determine the grade of the defect.

(Operation) In the defect grade determination by the second neural network, not only the feature amount of the defect but also the determination result of the defect type by the first neural network is reflected. In other words, the grade can be determined based on different criteria for each type of defect.

(Example) FIG. 1 shows mt of a basic embodiment of the invention.

The detector 1 images the surface of the inspection object 5 being conveyed in the direction of the arrow 7, and the feature amount of the defect 6 is extracted from the image signal in the feature extraction step N2, which is the same as in the horizontal or conventional method.

The first neural network 3 inputs the feature amounts (XI to X,) extracted by the feature extraction circuit 2, determines the type of defect, and outputs (y+ to Y,).

The second neural network 4 receives as input the feature amount (X + ~ X,) extracted by the feature extraction circuit 2 and the defect type determination output (Y1 ~ Y,) from the first neural network 3 , determines the defect grade for each defect type and outputs it (zl
~Zt).

The first and second neural networks 3.4 are networks consisting of a plurality of neurons, and weight parameters for each input are determined by learning. FIG. 2 shows an embodiment in which the second neural network 4 is the simplest neural network without a feedback mechanism. In FIG. 2, each of the first neuron 4-1 to the first neuron 4-e is a first neural neuron 4-e.
Each of these fires depending on the defect type determination output from the component 3.

FIG. 3 shows another embodiment of the second neural network 4. In FIG. Here, m neural nets 40-1 to 40- are used to judge the grade for each defect type.
Contains m. Each neural net 40-1 to 40
-m performs grade determination using the feature quantity (X, ~X,) as input, and selectively outputs outputs from grade 1 to grade l.

Only one of the grade 1 outputs of each neural net 40-1 to 40-m is selected by the output switch S1 to become the output GD1, and one of the grade 2 outputs is selected by the output switch S2. is selected and becomes the output GD2. Similarly, one of the grade E outputs is output switch S.

is selected and becomes the output GDi.

Then, the output switches Sl to S1 receive the type determination output (Y, ~Y,,) from the first neural network 3, and
Neural networks 40-1 to 40- corresponding to the type
It is possible to switch to select the output of the corresponding one among m. In other words, type determination output Y + (i = 1
□”m) is input, the neural network 40
The output of −1 becomes the grade determination output (GD1 to GD,).

〔Effect of the invention〕

As explained in detail above, the defect inspection device of the present invention can appropriately determine the grade of a defect using appropriate criteria according to the type of defect, just as a skilled person would visually inspect. . Also, the convergence of neural network learning is fast.

[Brief explanation of drawings]

Fig. 1 is a block diagram of the first embodiment of the present invention, Fig. 2 is a block diagram of the second embodiment, Fig. 3 is a block diagram of main parts of the third embodiment, and Fig. 4 is a block diagram of the main part of the third embodiment.
The figure is a configuration diagram of a conventional example. 3...First neural network 4...Second neural network 4-1 to 4-! ! ...neuron

Claims (3)

[Claims] (1) A detector that photoelectrically scans and images the surface of the object to be inspected, a feature extraction circuit that extracts the feature amount of the defect from the image signal of this detector, and a feature extraction circuit that uses the extracted feature amount as input to identify the defect. A defect inspection device comprising: a first neural network that determines and outputs the type; and a second neural network that receives the feature amount and the output of the first neural network as input and determines and outputs the grade of the defect. (2) The defect inspection apparatus according to claim 1, wherein the defect inspection apparatus is configured to distribute the output of the first neural network to each neuron of the first layer of the second neural network. (3) In the defect inspection apparatus according to claim 1, a plurality of neural networks are provided as the second neural net for determining the grade for each type of defect, and the plurality of neural networks are determined by the output of the first neural network. A defect inspection device characterized in that it is configured to selectively operate a net.