The growing interest in automation in industrial production increases the demand for artificial intelligence in the recognition of manufactured components. Sorting tasks require automated classification methods that can be implemented using machine vision and deep learning. One of the main challenges in deep learning applications is the collection of training data, especially when many different objects must be considered. Synthetic training images offer a way to avoid the time consuming acquisition of real images. In the presented method, synthetic images of industrial components are generated automatically within a computer aided design environment. A script creates images of the objects from multiple perspectives directly in the software in which they were designed. This lightweight, CAD integrated workflow explores how artificial data can be produced during the design process. A convolutional neural network is trained using transfer learning with synthetic data only and evaluated on real images to assess the sim to real gap. The method achieved an accuracy of 79.67% on the real world test set. The results show a clear internal improvement in accuracy when increasing the number of synthetic images, even with minor variations. The findings demonstrate the feasibility of the approach while indicating that the sim to real gap remains a challenge. A model trained with this workflow could support automated classification and sorting of industrial components.
Rabeneck et al. (Thu,) studied this question.