Browse Prior Art Database

Inspection of Industrial Products by Technical Neural Networks

IP.com Disclosure Number: IPCOM000108101D
Original Publication Date: 1992-Apr-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 2 page(s) / 80K

Publishing Venue

IBM

Related People

Graulich, M: AUTHOR [+4]

Abstract

Industrial products of increasing integration density, such as circuit boards, semiconductor chips and their substrates, require suitably automated inspection methods. The performance characteristics of up- to-date automatic inspection means are determined not only by throughput values but also by improved defect classification means and adaptability to changing inspection criteria.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Inspection of Industrial Products by Technical Neural Networks

       Industrial products of increasing integration density,
such as circuit boards, semiconductor chips and their substrates,
require suitably automated inspection methods.  The performance
characteristics of up- to-date automatic inspection means are
determined not only by throughput values but also by improved defect
classification means and adaptability to changing inspection
criteria.

      The method described in this article is adaptable to such
changing criteria by a teaching process.  For this purpose, technical
neural networks NNs are used to classify defects in the inspected
products.

      The problem with NNs in automatic inspection means is the
conversion of image data into feature vectors containing the desired
image information and permitting the data reduction required.  The
facet method described in this article resolves this problem.

      For automatic inspection, the product is initially placed below
a TV camera to be digitized.  The digital image thus obtained is
processed by one of two methods (*) such that the pixels representing
the object edges are coded with the edge direction (0 to 360~), so
that there is no information left otherwise.

      Then, the image is subdivided into elementary facets EFs, i.e.,
small square areas of, say, 8 x 8 pixels.  Within each of these EFs,
one feature vector is generated for the image information of the
respective EF.

      The EF feature vectors thus defined are used for two-step
classification by neural networks.

      In a first step, small edge sections are classified as
primitive edge, circle, straight line, and, additionally, a quality
feature good/bad is determined.  For this purpose an edge facet EDF
is defined.  It consist...