|
Quality control by artificial vision is getting more and more widespread within the industry. Indeed, in many cases, industrial applications require a control with high stability performance, satisfying high production rate. For texture control, some major problems may occur: uneasiness to show different textures, segmentation features as well as classification and decision phases requiring still to much computation time. This article presents a comparison between two non-parametric classification methods used for real time control of textured objects moving at a cadence of 10 pieces per second. Four types of flaws have to be indifferently detected: smooth surfaces, bumps, hollow knocked surfaces and lacks of material. These defects generate texture variations which have to be detected and sorted out by our system, each flaw apparition being registered to carry out a survey over the production cycle. We previously presented a search for an optimal lighting system, in this case the acquired images were tremendously improved. On these optimal images, we described a method for selecting the best segmentation features. The third step, which is presented here, is a comparison between two multi-classes classification algorithms: the Parzen's estimator and the so-called 'stressed polytopes' method. These two algorithms which require a learning phase are both based on a non-parametric discrimination method of the flaw classes. In one hand, they are both relatively inexpensive in time calculation but on the other hand they present different assets relative to the easiness of the learning phase or the number of useable segmentation features. They also have a different behavior towards the cut out of the features space, especially on the 'cross-classes' border. Their comparison is made through the aforementioned quoted points which are relevant for the evaluation of the discrimination efficiency. Finally, through an industrial example we present the results of such a comparison. The control, a PC based machine, includes the calculation five classification features (calculations were carried out on the local neighborhood of each pixel), five distinct classes for the classification phase and the decision phase. This led to a 3,63% classification error ratio for the best compromise.
|
