A causality-inspired data augmentation approach to cross-domain burr detection using randomly weighted shallow networks

The challenge of generalizing deep learning models to new, unseen areas or domains is common in material removal applications. This paper focuses on the specific problem of detecting burrs after material removal operations when training data is only available from a single operation. To address this issue, we propose a causality-based data augmentation approach for cross-domain burr image detection. The process involves using a combination of shallow networks to create a deep neural network resilient to differences in image intensity and texture. Additionally, we use various appearance transformations to enhance the training images further. However, we also demonstrate how incorrect correlations between objects in an image can negatively impact the model's robustness. To address this, we intervene in the causal chain of events by resampling the images of potentially linked objects. Hence it eliminates any erroneous correlations and helps the model to make more accurate predictions. To validate our approach, we perform experiments on three cross-domain burr detection tasks: milling, press punching, and casting. The results show that our approach consistently outperforms other methods when tested on unknown domains.