A novel learning-based feature recognition method using multiple sectional view representation

Abstract

Listen

Translate article

In computer-aided design (CAD) and process planning (CAPP), feature recognition is an essential task which identifies the feature type of a 3D model for computer-aided manufacturing (CAM). In general, traditional rule-based feature recognition methods are computationally expensive, and dependent on surface or feature types. In addition, it is quite challenging to design proper rules to recognise intersecting features. Recently, a learning-based method, named FeatureNet, has been proposed for both single and multi-feature recognition. This is a general purpose algorithm which is capable of dealing with any type of features and surfaces. However, thousands of annotated training samples for each feature are required for training to achieve a high single feature recognition accuracy, which makes this technique difficult to use in practice. In addition, experimental results suggest that multi-feature recognition part in this approach works very well on intersecting features with small overlapping areas, but may fail when recognising highly intersecting features. To address the above issues, a deep learning framework based on multiple sectional view (MSV) representation named MsvNet is proposed for feature recognition. In the MsvNet, MSVs of a 3D model are collected as the input of the deep network, and the information achieved from different views are combined via the neural network for recognition. In addition to MSV representation, some advanced learning strategies (e.g. transfer learning, data augmentation) are also employed to minimise the number of training samples and training time. For multi-feature recognition, a novel view-based feature segmentation and recognition algorithm is presented. Experimental results demonstrate that the proposed approach can achieve the state-of-the-art single feature performance on the FeatureNet dataset with only a very small number of training samples (e.g. 8–32 samples for each feature), and outperforms the state-of-the-art learning-based multi-feature recognition method in terms of recognition performances.