A surface roughness grade recognition model for milled workpieces based on deep transfer learning

Abstract Many roughness measurement methods rely on designed feature indexes that cannot accurately characterize the roughness and are demanding on the workpiece imaging environment. Roughness measurement methods based on deep neural networks require huge numbers of training samples and the same data distribution for training samples and testing samples, which makes it difficult to achieve wide application in the machining industry. Deep AlexCORAL, a surface roughness grade recognition model for milled workpieces based on deep transfer learning, is proposed in this paper to automatically extract more general roughness-related features. It not only reduces the amount of data required by the model but also the difference in data distribution between the source domain (training set) and the target domain (testing set). The experimental results show that Deep AlexCORAL has 99.33% cross-domain recognition accuracy in a variety of cases with inconsistent data distribution due to various lighting environments. This is unmatched by other roughness grade recognition models.