Evaluation of data augmentation and loss functions in semantic image segmentation for drilling tool wear detection

Tool wear monitoring is crucial for quality control and cost reduction in manufacturing processes, of which drilling applications are one example. Identification of the wear area in images of cutting inserts is important to building a reliable ground truth for the development of indirect monitoring approaches. Therefore, we present a semantic image segmentation pipeline for wear detection on microscopy images of cutting inserts. A broadly used convolutional neural net, namely a U-Net, is trained with different preprocessing and optimisation task configurations: On the one hand the problem is considered as binary problem, and on the other hand as multiclass problem by differentiating the wear into two different types. By comparing these two problem formulations we investigate whether the separation of the two wear structures improves the performance of the recognition of the wear types. For both problem formulations three loss functions, i. e., Cross Entropy, Focal Cross Entropy, and a loss based on the Intersection over Union (IoU), are investigated.The use of different augmentation intensities during training suggests adequate but not too excessive augmentation, and that with optimal augmentation the choice of loss function gets less important. Furthermore, models are trained on image tiles of different sizes, which has an impact on producing artefacts on the whole image predictions performed by the overlap-tile strategy. In summary, the best performing models are binary models, trained on data with moderate augmentation and an IoU-based loss function.