An application of generative adversarial networks to improve automatic inspection in automotive assembly lines

Abstract

In manufacturing systems, quality inspection is a critical issue. This can be performed by humans,or by means of Computer Vision Systems ( CVS), which are trained using representative sets of images, modeling classes of defects that may possibly occur. In practice, the construction of such datasets strongly limits the use of most CVS methods, as the variety of defects is of combinatorial nature. Alternatively, instead of recognizing defects, a system can be trained to detect non-defective cases, becoming appropriate for some application profiles. In flexible automotive manufacturing, for example, parts are assembled within a reduced set of correct combinations, while the number of possible incorrect assembling is enormous. In this paper, we show how a CVS can be extended with a Deep Learning-based approach that exploits a Generative Adversarial Network ( GAN) to detect non-defective production eliminating the need for constructing expensive defect image datasets. The proposal is tested over the assembly line of Renault, in Brazil. Results show that our approach has better accuracy in inspection, compared with the currently used CVS. We also show that the same method can be used in different components inspection, without any modification.