Article Preview
TopIntroduction
Basically, Printed Circuit Board (PCB) is a piece of phenolic or glass epoxy board with copper clad on one or both sides. The portion of copper that are not needed are etched off, leaving 'printed' circuits which connects the components. It is used to mechanically support and electrically connect electronic components using conductive pathways, or traces, etched from copper sheets and laminated onto a non-conductive substrate. PCBs are rugged, inexpensive and highly reliable and so it is used in virtually all but the simplest commercially produced electronic devices.
In recent years, the demand of electronic devices with more compact design and more sophisticated functions has forced the PCBs to become smaller and denser with circuits and components. As it is crucial part of electronic device it needs to be properly investigated before get launched. Automatic inspection systems are used for this purpose but due to more complexity in circuits, PCB inspections are now more problematic. This problem leads to new challenges in developing advanced automatic visual inspection systems for PCB.
Automatic Optical Inspection (AOI) has been commonly used to inspect defects in Printed circuit board during the manufacturing process. An AOI system generally uses methods which detects the defects by scanning the PCB board and analyzing it. AOI uses methods like Local Feature matching, image Skeletonization and morphological image comparison to detect defects and has been very successful in detecting defects in most of the cases but production problems like oxidation, dust, contamination and poor reflecting materials leads to most inevitable false alarms. To reduce the false alarms is the concern of this paper.
Previous approach (Tanaka, Hotta, Iga, & Nakamura, 2007) classifies the defects using neural network and Rau and Wu (2005) proposes a method to classify the defects using the intensity at the pixels around the defects region. These approaches classify the defect under the condition that kinds of the defects are previously known. There are some defects whose recognitions are difficult even with the visual inspection. These defects cause the problem. The problem includes the case of misjudgment where a true defect is recognized as a pseudo defect and it is included in the products as a result. Kondo (Kondo, Kikuchi, Hotta, Shibuya, & Maeda, 2009) has been proposed for the distinction of defect classification by determining the features at random. Kondo, Kikuchi, Hotta, Shibuya, and Maeda (2009) classifies the kinds of defect with selecting the appropriate features with classifiers, but there are still incorrect classification cases where a true defect is classified into a pseudo defect.
Approaches to extract the defect candidate region are proposed in Onishi, Sasa, Nagai, and Tatsumi (2003) Maeda, Ono, Makoto, Kubota, and Nakatani (1997), and Numada, and Koshimizu (2007). Onishi (Onishi, Sasa, Nagai, & Tatsumi, 2003) prepares two images of test image and reference image of mask pattern and takes difference image by logical AND. Maeda (Maeda, Ono, Makoto, Kubota, & Nakatani, 1997) propose IR image matching and Mahalanobis distance, respectively.
Other classification approaches include Wakabayashi (Wakabayashi, Tsuruoka, Kimura, & Miyake, 1995) using PCA, Ishii (Ishii, Ueda, Maeda, & Murase, 1998) using variance inside and outside classes, Amabe (Amabe, & Nagao, 2006) using Genetic Algorithm, Roh (Roh, Yoon, Ryu, & Oh 2001) using Neural Network. Another approach to remove incorrect classification of true defect is proposed in Iwahori (Iwahori, Futamura, & Adachi, 2011; Iwahori, Kumar, Nakagawa, & Bhuyan, 2012), where histogram for each defect and evaluating equation are introduced. It is noted that these approaches generate the reference images which are used to detect defect region manually.