A Semi-Supervised Learning Framework for Machining Feature Recognition on Small Labeled Sample

Automated machining feature recognition is an essential component linking computer-aided design (CAD) and process planning (CAPP). Deep learning (DL) has recently emerged as a promising method to improve recognition. However, training DL-based models typically require annotating large amounts of data, which time-consuming labor-intensive for researchers. Additionally, DL struggle achieve satisfactory results when presented with small labeled datasets. Furthermore, existing approaches significant memory processing time, thus hindering their real-world application. To address these challenges, this paper presents semi-supervised framework that leverages both unlabeled data learn meaningful visual representations. Specifically, self-supervised utilized extract prior knowledge from dataset without annotations, then transferred downstream tasks. we apply lightweight network techniques two established recognizers, FeatureNet MsvNet, develop reduced-memory, computationally efficient termed FeatureNetLite MsvNetLite, respectively. validate the effectiveness proposed approaches, conducted comparative studies on dataset. With only one sample per class, MsvNetLite outperformed MsvNet by about 19%, whereas approximately 20% in classification. On common X86 CPU, gained 6.68× improvement speed over was 2.49× faster than FeatureNet. The shows while achieving optimal balance between accuracy inference compared other approaches.