CAD-based Data Augmentation enabling Transfer Learning for Part Classification in Manufacturing

Abstract

The stunning success of deep image classification approaches relies heavily on huge, labeled datasets based on real-world images. Nonetheless, the latest generation of neural network architectures provides pre-trained models that allow the training and classification of new classes with only minor example data. In most application use cases, the prediction performance is still too low to meet the high standards of manufacturing systems, e.g., for visual part classification. Furthermore, it is often difficult to acquire sufficient image data of objects in manufacturing systems due to customized and small batch production. While data augmentation techniques like viewport transformations or artificial noise barely increase the data set's overall entropy, we aim to generate highly realistic synthetic images only based on photorealistic rendered CAD. Additionally, this approach can reduce manual effort in data generation and preparation.First, we demonstrate the creation of a parametric rendering pipeline. Then the models are trained for a 30-class classification problem with typical engineering parts from an industrial use case. Finally, we show how the entropy gain, using our proposed method, improves the classification performance in any of the investigated deep image classification models.