Deep Learning Based Synthetic Image Generation for Defect Detection in Additive Manufacturing Industrial Environments

Abstract

As the amount of additive manufactured parts is rising methods for part defect detection are needed to guarantee good product quality and fulfill quality management requirements. The usage of deep learning methods in industrial environments for artefact detection is growing, therefore, it is crucial to obtain enough training data in order to deploy powerful models for intelligent control systems. We propose a novel approach for synthetic image creation for object defect detection of Fused Deposition Modeling (FDM) manufactured parts based on deep learning methods and demonstrate the capability to enhance deep learning based defect detection with synthetic images. Our approach is based on physical rendering in combination with a generative adversarial network (GAN) for synthetic data generation. We illustrate how the generated synthetic images can be used to enhance deep learning methods by training an auto encoder model which afterwards is used for failure detection. By an evaluation study it is depicted that our approach for synthetic image generation achieves good results where, in comparison to real world images nearly the same amount of images are assumed to be real. Using the advantage of synthetic data, the autoencoder model is able to detect failures in real images. Therefore, the approach is able to generate photo realistic images which can be used to detect defects on parts with limited training material.