Application of Artificial Vision to the Quality Inspection of Surfaces of Luxury Products

This thesis defines a general approach for the automated visual inspection of surfaces of luxury products, the aesthetic quality of which represents a corollary to the technical mastery necessary to manufacture valuable products. Within a manufacturing enterprise, experts' judgment depends on several factors and, since not all these factors are controlled, the practice of visual inspection is affected by an undesired variability. Nevertheless, experts have a great ability to adapt, which enables them to evaluate the parts and to issue a reliable judgment, even if the "defects" appear for the first time on the surfaces of the produced goods. Therefore, experts constitute the reference for the evaluation of the aesthetic surfaces of luxury products. This thesis proposes methods and associated tools for the analysis and qualification of the quality of aesthetic surfaces, in order to reduce the variability of their inspection. The inspection process that we have adopted is based on three phases: the identification of defects, their quantification and the judgment of the quality compliance of the inspected surfaces, on the basis of the criteria defined by the different manufacturers. In explaining the importance of illumination, we draw a parallel with the inspection process carried out by the experts. During the inspection, the direction of the incident illumination and the viewing direction both play an important role in determining surface appearance. In presenting the proposed approach, we have also pointed out the relevant distinction between the analysis carried out under the global point of view and the one focusing on local imperfections. Concerning the global point of view, we have presented some examples of texture analysis, based on the statistical properties calculated over a restrained neighbourhood of pixels. The automatic classification is carried out by using an approach where the reference of classes is provided by the experts. For the local analysis of a surface, the method is based on a model defining the characteristics of the texture. This allows to isolate the irregularities that reveal the presence of a defect. Then, some families of features are proposed in order to bring out the characteristics of the potential anomalies. The estimation of the seriousness of the anomalies is used to assign the inspected parts to different quality classes, depending on whether the anomalies are considered acceptable or not. Two applications of the proposed methods are presented. The first one has been developed for the inspection of local defects appearing on ballbearings manufactured for watch making applications. This application shows the different contributions of the proposed approach: The effective collaboration with human experts in order to define quality classes. The use of illumination enabling the identification of the local defects with a resolution as good as the one of the experts. The acquisition and processing of multiple images in order to achieve a classification of defects based on the references provided by human experts. Another application, developed in a completely different domain, has provided equally encouraging results.