To find a flaw not larger than a few micrometers on a surface spanning several square meters is a difficult task. A northern German company took up the challenge and constructed an inspection system just with standard components to find flaws on large surfaces in an appropriate time frame. How the company‘s engineers completed the task, what technology has been used to find the flaws and how the data is validated is presented in the following article.

Optical quality inspection is an established field, but assume one needs to produce a 1,500 by 1,000 mm panel that can have various types of defects in the size range of a few micrometers: who is going to inspect that? Obviously, we require an automatic inspection system that can inspect large objects with a high resolution in a reasonable amount of time. The first step in developing an affordable technology for this problem is to use standard mechanical and optical components. To obtain the desired resolution, both mechanical and optical inaccuracies must be compensated for in software, for example image registration is used to compensate for positioning inaccuracies. A standard megapixel camera, a macroscopic zoom lens, and flashed LED illumination define the core of the scanner (see fig.1). In cases where 3D structures are to be inspected, we employ more sophisticated illumination techniques like, for example, photometric stereo. At a frame rate of 20 frames per second and a resolution of 4 micrometers per pixel we scan the above-mentioned panel in about an hour (see table 1 for example calculations). Alternatively, by using a microscope, our customer could maybe manually inspect samples of the panel and spend a few days before making a proper decision.

Estimate the Density
The main challenge however is not only to deliver a large high-resolution image but to detect the defects on the fly and deliver error statistics for the whole panel and the different defect types. In addition to providing error-density maps for the panel, optimization routines are used to estimate the regions of lowest error density in case that only certain parts of the panel may be used. Defects can be of two sorts: (i) more or less well defined defects for which the customer can provide sufficient samples for each error class, and (ii) defects that must be determined as deviations from a regular surface. In both cases we employ proprietary state of the art machine-learning techniques that go well beyond standard tools. For example to detect complex deviations from a regular surface texture, novel one-class classifiers have been developed [1].

Avoid Bad Products
One of the most serious problems with these kinds of applications is validation. Therefore, we have developed techniques that allow to find correspondences between the images taken with our scanner and other more sophisticated imaging techniques like the scanning electron microscope (see fig. 3). In most cases we inspect low-batch expensive panels. These panels are inspected at various stages using comprehensive error statistics to reject defective parts as soon as possible. However, our goal is not only to sort the good and bad, but to deliver a comprehensive error analysis that can help to improve the production process and avoid deficient products. The ease and speed of scanning, the comprehensive error analysis, and the sophisticated validation process can help to better understand and optimize the production pro­cess.

Conclusion
We conclude that state of the art machine learning and pattern recognition techniques can be used to automatically detect complex error types and deviations from defined quality standards in case of high-resolution inspection of large objects. Moreover, comprehensive statistical evaluations can be used to optimize the production process. Based on intelligent software design all this can be achieved with standard components at affordable cost.

Reference
[1] Fabian Timm, Sascha Klement, Thomas Martinetz, and Erhardt Barth. Welding inspection using novel specularity features and a one-class svm. In Proceedings of the Int. Conference on Imaging Theory and Applications, volume 1, pages 146-153, Lisboa, Portugal, 2009. INSTICC.