Small data-driven semantic segmentation of wear debris in ferrography images

Abstract The segmentation of wear debris images is a prerequisite for ferrographic analysis, and uncertainties and errors in wear debris segmentation will inevitably affect the subsequent analysis. In this work, a small-data semantic segmentation model of wear debris images is constructed based on HRNetv2 for ferrography images acquired by using an online visual ferrography. A major advantage of the current model is that fewer ferrography images are required for training, and fewer iterations are needed. The current work is performed for ferrography images with both clean oil and contaminated oil, and good segmentation results can be found. Specially, the experimental results show that the current model can achieve accurate segmentation of wear debris images with mean intersection over union values of 91.47% and mean pixel accuracy values of 96.48%.