Synthetic datasets for Deep Learning in computer-vision assisted tasks in manufacturing

Abstract Artificial Intelligence applications based on Machine Learning methods are widely accepted as promising technologies in manufacturing. Deep Learning (DL) techniques, such as Convolutional Neural Networks (CNN), are successfully used in many computer-vision tasks in manufacturing. These state-of-the-art techniques are requiring large volumes of annotated datasets for training. However, such an approach is expensive, prone to errors and labor as well as time intensive, especially in highly complex and dynamic production environments. Synthetic datasets can be utilized for accelerating the training phase of DL by creating suitable training datasets. This work presents a framework for generating datasets through a chain of simulation tools. The framework is used for generating synthetic images of manufactured parts. States of the parts such as the rotation in different rotation axis need to be recognized by a computer-vision system that assists a manufacturing operation. A number of prior trained CNNs are retrained with the synthetically generated images. The CNNs are tested upon actual images of manufactured parts. The performance of different CNN models is presented, compared and discussed. The results indicate that CNNs trained on synthetically generated datasets may have acceptable performance when used in for assisting tasks in manufacturing.