Synthetic Data Set Generation for the Evaluation of Image Acquisition Strategies Applied to Deep Learning Based Industrial Component Inspection Systems

Abstract

Automated visual inspection is an ongoing machine vision challenge for industry. Faced with increasingly demanding quality standards it is reasonable to address the transition from a manual inspection system to an automatic one using some advanced machine learning approaches such as deep learning models. However, the introduction of neural models in environments such as the manufacturing industry find certain impairments or limitations. Indeed, due to the harsh conditions of manufacturing environments, there is usually the limitation of collecting a high quality database for training neural models. Also, the imbalance between non-defective and defective samples is very common issue in this type of scenarios. To alleviate these problems, this work proposes a pipeline to generate rendered images from CAD models of industrial components, to subsequently feed an anomaly detection model based on Deep Learning. Our approach can simulate the potential geometric and photometric transformations in which the parts could be presented to a real camera to faithfully reproduce the image acquisition behavior of an automatic inspection system. We evaluated the accuracy of several neural models trained with different synthetically generated data set simulating different transformations such as part temperature or part position and orientation with respect to a given camera. The results shows the feasibility of the proposed approach during the design and evaluation process of the image acquisition setup and to guarantee the success of the real future application.