Evaluation of Data Augmentation and Loss Functions in Semantic Image Segmentation for Drilling Tool Wear Detection

TL;DR

The paper tackles direct wear detection in drilling tools via a U-Net segmentation pipeline applied to microscopy images. It systematically evaluates binary versus multiclass wear definitions, three loss functions ($CE$, $FCE$, and an $IoU$-based loss), two tile sizes, and three augmentation regimes, using 5-fold cross-validation across 72 configurations. Key findings show that binary models with batch normalization and an $IoU$-based loss trained with moderately augmented data achieve strong performance, with final $IoU$ around $0.886$ for 512 px tiles and $0.904$ for 256 px tiles; multiclass configurations can match or exceed binary performance in some settings, while full augmentation can reduce accuracy. The approach enables robust wear-area segmentation and the extraction of wear metrics directly from segmentation masks, with larger tiles offering resilience to reflections, highlighting practical value for automated tool wear monitoring and process optimization.

Abstract

Tool wear monitoring is crucial for quality control and cost reduction in manufacturing processes, of which drilling applications are one example. In this paper, we present a U-Net based semantic image segmentation pipeline, deployed on microscopy images of cutting inserts, for the purpose of wear detection. The wear area is differentiated in two different types, resulting in a multiclass classification problem. Joining the two wear types in one general wear class, on the other hand, allows the problem to be formulated as a binary classification task. Apart from the comparison of the binary and multiclass problem, also different loss functions, i. e., Cross Entropy, Focal Cross Entropy, and a loss based on the Intersection over Union (IoU), are investigated. Furthermore, models are trained on image tiles of different sizes, and augmentation techniques of varying intensities are deployed. We find, that the best performing models are binary models, trained on data with moderate augmentation and an IoU-based loss function.

Figures (4)

• Figure 1: (\ref{['fig:img']}) Light optical microscopy image of cutting insert, image done by digital microscope (VHX7000, Fa. Keyence); (\ref{['fig:mask']}) associated mask with abrasive wear area A in blue and area of transferred work piece material M in yellow; zoom for better visibility of the two wear types.
• Figure 2: Flowchart of data preprocessing with augmentation.
• Figure 3: Distribution of IoU of 5-fold cross validation of the models trained with loss functions CE, FCE and $L_{\text{IoU}}$, with and without batch normalisation (BN), on data with full, moderate, and no augmentation. The test data are unseen and not augmented tiles with edge length 512 px and 256 px respectively.
• Figure 4: Evaluation of the four test images, each arranged in blocks of two rows: In each block, the first image in the first row shows images with the binary ground truth mask (left), followed by the predictions of the binary model using tiles with edge length 512 px (middle), followed by images of tiles with edge length 256 px (right). The second row shows the same for the multiclass masks and multiclass models. While for the multi class models the predicted categories A and M are shown separately, the reported IoU values are computed using A and M as one joint class.