SYNTHETIC SCENE BASED OBJECT DETECTION: CAN YOU TRAIN NEURAL NETWORKS WITH SYNTHETICALLY GENERATED DATA TO RECOGNIZE REAL OBJECTS FOR MANUFACTURING?

Within the manufacturing industry, Artificial Intelligence (AI) has revolutionized the way manufacturers mass produce goods and has supported a more consumer driven focus. The ability of computers to understand digital images or videos is a key sensing capability that enables connected enterprise systems. Deep learning (DL) techniques used for computer vision tasks such as object detection reliably learn features without explicitly defining said features in the training process. Evaluating deep learning techniques for object detection applications can be broken into three steps: gathering/processing data, model training/evaluation and model deployment. This research focuses on the validity of using synthetically generated data to train convolutional neural networks (CNN) to recognize objects. An object detection pipeline was created to provide an infrastructure for new models to be created using 3D computer aided design (CAD) models of objects. The pipeline was built for research purposes but can be used for application purposes beyond testing models performance. A design of experiments was created to explore significant dataset parameters as well as hyperparameters used to train deep neural networks for object detection. Some significant factors explored were dataset size, inter-class variation, learning rate, number of epochs, and type of ma- chine learning model. Detectron2 was used to train was used to implement most object detection algorithms. Using optimized synthetically generated datasets and optimal hyperparameters, the object detection pipeline was capable of detecting objects bounding box and segmentation with average precision greater than 80%.