Unsupervised Material Classification of Printed Circuit Boards Using Dimension-Reduced Spectral Information

Abstract This paper proposes an unsupervised method for ma-terial classification of raw printed circuit boards (PCBs) based on the dimension-reduced spectral information. First, a spectral imaging system is constructed for ac-quiring precise spectral reflectance images. Only 3-dimensional spectral features are then extracted from the original 31-dimensional spectral image data by PCA for segmentation proposes. Next we apply the normalized cut and k-means clustering algorithms to the material classification using the reduced spectral reflectance data. Experimental results show the feasibility of the proposed method in comparison with the relevant 3-dimensional RGB imaging system. It is shown that the proposed method achieves very high quality in the material classi-fication. 1. Introduction A printed circuit board is one of the most complicated objects to understand from the observed image in a vari-ety of industries. A raw PCB surface layer is composed of various elements, which are a mixture of different mate-rials such as paint, metal, resist, and substrate. The area of each element is very small. These features make the ma-chine inspection difficult by using general color imaging systems based on only three spectral bands RGB [1]. Recently, spectral imaging has drawn great attention due to its potential applications such as object recognition in many fields. A spectral image comprising monochrome images at more than four different wavelengths has a large amount of spectral information which improves the ability to detect object materials or distinguish different areas. Tominaga et al. [2], [3] proposed material classifi-cation algorithms for raw PCBs based on surface-spectral reflectance. However, those algorithms had limitations due to spectral imaging system. Moreover, they required huge data storage and computational complexity for dealing with the high-dimensional spectral components to obtain enough accuracy in classifying PCB elements. The present paper proposes a method for unsupervised material classification for raw PCBs based on the dimen-sion-reduced spectral information. We construct a new spectral imaging system for acquiring precise spectral reflectance images. For solving the problem of the huge volume on the spectral reflectance data for PCBs, we consider extraction of low-dimensional spectral features to be used in the process of material classification. A variety of methods for dimensionality reduction have been proposed [4]. In this paper, PCA is invested for the present problem. As well known, the PCA is a classical technique which reduces dimensionality by forming lin-ear combinations of some statistical features. This analysis leads to effective material classification and image segmentation under the reduced degrees of free-dom, space and time complexities. The well-known normalized cut [5] and k-means clustering [4] algorithms are used for the material classi-fication based on the reduced spectral reflectance data. The performance of the proposed method is examined on experiments using real PCBs. We compare the classifica-tion results to the relevant RGB imaging system. The most effective imaging system and algorithm are deter-mined for improving the accuracy of PCBs material classification.