Machine vision plays an important role in automation of a large number of industrial processes. In the recent years, considerable efforts have been directed towards the development of automated inspection systems. Surface inspections are required to be automated in many industrial tasks, because they are laborious for human to perform. An ability to inspect specular surface is very valuable for machine vision. Surfaces of many industrial tasks have specular reflection property such as, machine parts, printed circuit boards, solder joint, plastic sheets, and so on. Generally, recognizing specular object is a hard problem for machine vision. Specular reflections of surface appear, disappear, or change their shapes abruptly, due to tiny movements of the viewer. Furthermore, a distant point illumination does not produce smooth shading on specular object, because light is just reflected such that the angle of incident equals to the angle of reflection. It makes difficult to extract stable visual cues enough to infer 3-D shapes of specular object. Thus, the majority of object recognition researches has ignored such specular reflections as useless noise for recognition purposes.
In order to recognize specular object, we must use features arising from specular reflection. Since specular features show very strong, distinct, and saturated brightness, they provide significant information about object which generates them. Specular features can tell us not only something of reflectance properties of the surface, but also some qualitative information about surface shape. To recognize object shape from specular features, it is necessary to establish a particular configuration of the object, the illumination source, and the sensor. Then, we can predict how a specular object exhibits specular reflection depending on the configuration and reflectance properties of its surface. But, a few techniques of illuminating and imaging have been implemented in the specular object inspection for industrial tasks.
In this thesis, a method is presented for obtaining shapes of specular objects using circular illumination to get over such difficulties in recognizing specular objects. The circular illumination system employs a circular light source which is positioned on the axis of camera. We assume that the circular source is distant from a viewed object, and the circular light source is large relative to the object. The assuming configuration is meaningful that the circular light source can be considered as a large number of points sources which are fixed circularly around the object and illuminate the object with the same zenith angle. This configuration restricts specular reflections to the loci of points, called iso-inclination contour, where the surface normal are at iso inclination to a selected reference vector. Iso inclination contour is an important visual cue to infer 3-D shapes. Techniques of measuring 3-D shapes by iso inclination contours are presented and are tested using regular shapes and solder joints on printed circuit board.
This method is applied to inspect 3-D shapes of solder joints on printed circuit board. Solder joint shape inspection, which has been a critical issue for quality control in PCB assembly process, is one of typical specular object recognition task in machine vision. Imaging of a solder joint surface is a difficult task since a solder joint forms a tiny, specular, curved and smooth surface. Besides the imaging problem, another difficulty encountered in this application stems from the great variability of solder joint shapes. To overcome such difficulties, two neural network based methods are proposed, which are called adaptive LVQ and vector correlation neural network, respectively. Practical feasibility of the methods are tested by building a prototype inspection machine. The test results show that the proposed methods are very suitable for recognizing various shapes of tiny specular objects in real situations.
