Detection and Segmentation of Defects in CNC Machine Inserts Using Transfer Learning with Dataset Similarity Evaluation

Abstract

The transfer of knowledge from one product to another product has been a highly demanded technique in industrial domains, since it does not need a large training dataset which is often costly available. However, this technique performance may not be always satisfying, due to several issues such as target dataset size or large difference between the source and target datasets. In this paper, we perform transfer learning for segmentation of CNC machine inserts to detect defective inserts using a pre-trained segmentation network. DeepLabv3 framework is adopted for the segmentation task with a modified loss function to speed up its training. A transfer learning strategy with pre-trained model backbone fixed and classifier fine-tuned is applied and the transfer learning performance is investigated on how it relates to the properties such as dataset size. A similarity measure between datasets is proposed to determine which source dataset is the most appropriate for transfer learning on a target dataset.