We have developed a deep learning model that helps to identify important geometrical features of the machine parts from 3D models and images.

Reusability of the manufactured product has been a sustainable practice followed in many industries to save resources. It may be difficult to identify that product manually by going through each part of the product to identify which one can be reused. Machine feature is a process of identifying sematically higher level of geometrical elements such as hole, passage, slots, etc., from an image or a 3D model of the machine parts.

Read The Paper (Kamal et al.)

3D CAD Models

We adopted the curated dataset from CAD models of machanical parts termed FeatureNet which is a novel framework using Deep 3D Convolutional Neural Network (3D-CNNs). These FeatureNet learns the distribution of complex machining feature shapes across a large 3D model data set and discovers distinguishing features that help in recognition process automatically.

Dataset consists of 24000 models belonging to 24 different classes. CNN with 97,272 learnable parameters is used to extract the features from the large database for better deformation from the CAD design. The top 5 accuracy was about 95% and with the final layer as a pyramid pooling layer, it was more efficient during the process.

Machine feature Recognition uses inductive transfer learning.

Geometrical features were extracted from the CAD models using an inductive transfer learning technique using a model pre-trained with fully convolutional geometric features for the purpose of image registration. Point cloud registration is the underlying source task of this process. The number of extracted geometrical feature vectors from this process was varying with respect to the CAD models. In order to get the same number of feature spaces,later an SPP layer was introduced as the target task of inductive transfer learning process. Frobenius norm was computed to measure the similarity between the CAD models. Based on the obtained similarity score, 3D models were assigned to the respective classes.


centered image

Shape search Shape Search deals with creating an automated approach to extracting and identifying different geometrical features of the CAD models and comparing them with the existing models on the database.
Partial shape search Partial shape search recognizes the similarities across the models with partial features by their geometric properties and retrieves the CAD model that has identical partial or complete features to the model given by the user.
Duplicate Part Assessment MFR can recognize the repeating shapes within the CAD models or across CAD models and categorize them into a single group
Model Category Grouping The grouping of CAD models according to their geometric similarity can help us to perform analysis and enables easy access
3D model comparison 3D models can be compared to find their common features through MFR

To learn more, check out our GitHub and read our publication submitted to 9th International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA 2021)


If you have queries about our work, contact us at: