R. Schettini, S. Tominaga, and A. Trémeau (Eds.): CCIW 2011, LNCS 6626, pp. 152–164, 2011. © Springer-Verlag Berlin Heidelberg 2011 Material Classification for Printed Circuit Boards by Kernel Fisher Discriminant Analysis Takahiko Horiuchi, Yuhei Suzuki, and Shoji Tominaga Graduate School of Advanced Integration Science, Chiba University, Japan {horiuchi@faculty,suzukiyuhei@graduate,shoji@faculty}.chiba-u.jp Abstract. This paper proposes an approach to a reliable material classification for printed circuit boards by kernel Fisher discriminant analysis. The proposed approach uses only three dimensional features of the surface-spectral reflectance reduced from the high-dimensional spectral imaging data for effectively classifying the surface material on each pixel point into several elements such as substrate, metal, resist, footprint, and silk-screen paint. We show that a linear classification of these elements does not work well, because the feature distribution is not well separated in the three dimensional feature space. In this paper, a kernel technique is used to constructs a subspace where the class separability is maximized in a high-dimensional feature space. The performance of the proposed method is compared with the previous algorithms using the high-dimensional spectral data. Keywords: Material classification, printed circuit board, spectral reflectance, region segmentation, kernel discriminant analysis. 1 Introduction Automatic visual inspection (AVI) has become crucial to improve quality in printed circuit board (PCB) manufacture. A PCB is one of the most complicated minute objects to understand from the observed image in a variety of industries. The raw PCB, i.e. boards without components, may have defects such as: hairline, pin-hole, wrong size hole, open circuit, and breakout. Many researchers (e.g., see the references in [1]) repeatedly emphasized the importance of developing techniques and algorithms for an automatic inspection system in the electronic industry. Consequently, a wide range of defect detection techniques and algorithms have been reported and implemented in AVI systems [2-6]. Most of them were based on binary or gray-scale images to find board defects. A raw PCB surface layer is composed of various elements, which are a mixture of different materials, and the area of each element is very small. These features make the machine inspection difficult by using binary, gray-scale images, or even by using general color imaging systems based on only three spectral bands of RGB. The segmentation of a PCB image enables us to transform the original problem of inspecting a complex PCB image to a simpler problem of inspecting a well-defined segmented image. The surface of the raw PCB is partitioned into small areas of different materials such as metal, resist, footprint,