Deep Learning Image Augmentation using Inpainting with Partial Convolution and GANs

Abstract

The 21st century has seen the remarkable transformation of machine vision by deep learning. This has enabled intelligent systems like autonomous vehicles and facial recognition software. However, the success of deep learning is largely predicated on the availability of sufficient data; in many instances, data may be scarce and expensive to source. In this thesis, we implemented two deep learning techniques: (1) inpainting using partial convolution and (2) generative adversarial network (GAN) to generate synthetic data to train deep learning image classifiers. We show that the addition of synthetic training images dramatically improved the accuracies of our defect classifiers. Using Gradient- Class Activation Map (Grad-CAM), we also demonstrate that the decision rules learned by the classifiers are significantly enhanced where the classifiers are accurately activating at the specific defect locations upon addition of synthetic training images. The study was performed at Amgen using real images of syringes and vials, indicating the practicality of the technique for industrial applications.