李得盛9...

2007 International Conference on Artificial Intelligence and Pattern recognition (AIPR-07)
403
Copper Clad Laminate Defects Classification Using Inverse Wavelets
Transform and Support Vector Machine
Te-Sheng Li
Minghsin University of Science and Technology
Department of Industrial Engineering and Management
1 Hsin-hsin Rd. Hsin-fong, Hsinchu County, Taiwan 30401, R.O.C.
[email protected]
Abstract
In this paper, we present a multi-resolution approach for
the inspection local defects embedded in homogeneous
copper clad laminate (CCL) surfaces. The proposed
method is based mainly on the wavelet transform and
inverse wavelet transform on the subimages by properly
selecting the adequate decomposition levels. The restored
image will remove regular, repetitive texture patterns and
enhance only local anomalies. Based on these local
anomalies, feature extraction methods can then be used to
discriminate between the defective regions and
homogeneous regions in the restored image. Real samples
with five classes of defects have been classified using this
novel multi-classifier, namely, support vector machine.
The experimental results show the efficacy of the
proposed method.
1. Introduction
Visual inspection plays a vital role of quality control in
the manufacturing system. Manual inspection is subjective
and highly dependent on the inspector’s expertise. In this
study, we use machine vision and image processing
techniques combining a novel classifier, support vector
machine, to detect and classify copper clad laminate (CCL)
surface defects. Most CCL surface defects are tiny
involving obvious faulty items such as pine holes, stains,
scratches, strips and other ill-defined defects. These
unanticipated defects are small in size, refractive in light
and cannot be described using explicit measures, making
automatic defect detection difficult.
Because CCL defects are similar to texture patterns, the
inspection task in this study can be classified as texture
detection and classification. Most defect detection
methods for texture surfaces generally involve computing
a set of textural features in a sliding window. The system
searches for significant local deviations in the feature
values. The most difficult task in this approach is feature
extraction and feature selection which completely embody
the texture information in the image. There is no arbitrary
approach for selecting and judging the appropriate
features to use.
Numerous methods have been proposed to extract
textural features either directly from the spatial domain or
from the spectral domain. In the spatial domain, the more
reliable and commonly used features are the second-order
statistics derived from spatial grey-level co-occurrence
matrices [1-2]. They have been applied to wood
inspection [3], carpet wear assessment [4], and roughness
measurements for machined surfaces [5]. Typically, the
spectral-domain features are generally less sensitive to
noise than spatial-domain feature due to the relatively
uniform representation in the spectral domain. Recently,