Automated training of convolutional networks by virtual 3D models for parts recognition in assembly process

Abstract

One of the most monotonous activities in using convolutional neural networks for image recognition is preparation of the learning data. It involves creating samples (2D images of object) at different angles of view, different backgrounds/materials and partial overlay of the object. Input data must include a relatively large number of frames, typically about 100 and more images per object to make the learning precision useful. In the paper there is proposed a new approach to creating these data fully automated based on a virtual 3D model of the standardized parts. Automation principle is generating 2D images from the imported 3D construction model, including the following variable parameters: the angle of rotation, background and the material of the component. We used for verification pretrained DNN model Faster RCNN Inception v2 with single shot detection (SSD). The learned convolutional network was next tested by real samples to verify a new approach of learning by virtual models and recognition of real objects (parts).