科研之友
微信
新浪微博
Facebook
分享链接
摘要
Automated optical inspection (AOI) systems are essential in electronic manufacturing technology. These systems can inspect the products%26apos; quality quickly, accurately and without weariness. However, today employing human operators and engineers is a must to compensate the inflexibility and insufficient intelligence of AOI devices. This cooperation between machines and humans implies special difficulties in the topic of optical inspection. Several AOI systems collect and store the created inspection images. For later usage (e.g., for training the AOI algorithms), the image databases are often separated according to the inspection. Normally, these databases are disjoint and homogeneous, but because of machine and human errors, the imagebases can contain outlier images - e.g., falsely classified data - which do not satisfy the specified requirements. These errors make the optimization and verification process of AOI inspection algorithms challenging, which can decrease the systems%26apos; reliability and accuracy. Currently, filtering the image databases to remove outliers is a very demanding problem. This paper deals with this question. First, we illustrate the difficulties related to AOI image databases, show their background and identify their causes. Next we present our novel outlier detection method. Our technique can find and remove the outlier images from databases of inspection images reliably, by means of image processing, statistical and clustering methods. Our algorithm does not apply training samples (training images) and minimizes - or in some cases totally eliminates - the need for human intervention. It works on several component types without any modification. Our experiments showed that our novel technique can find 99 % of outlier images in average.
- 出版日期2012-4
