John W. Smutny

John W. Smutny

This paper investigates how rendering engines, like Unreal Engine 4 (UE), can be used to create synthetic images to supplement datasets for deep computer vision (CV) models in image abundant and image limited use cases. Using rendered synthetic images from UE can provide developers and businesses with a method of accessing nearly unlimited, reproducible, agile, and cheap training sets for their customers and applications without the threat of poisoned images from the internet or the cost of collecting them. The validity of these generated images are examined by testing the change in model test accuracy in two different sized CV models across two binary classification cases (Cat vs Dog and Weld Defect Detection). In addition, this paper provides an implementation of how to measure the quality of synthetic images by using pre-trained CV models as auditors. Results imply that for large (VGG16) and small (MobileNetV3-small) parameter deep CV models, adding >60% additional synthetic images to a real image dataset during model training can narrow the test-training accuracy gap to ~1-2% without a conclusive effect on test accuracy compared to using real world images alone. Likewise, adding <10% additional real training images to synthetic only training sets decreased the classification error rate in half, then decreasing further when adding more real training images. For these cases tested, using synthetic images from rendering engines allow researchers to only use 10% of their real images during training, compared to the traditional 50-70%. This research serves as an example of how to create synthetic images, guidelines on how to use the images, potential restrictions and possible performance improvements for data-scarce projects.

John W. Smutny

Gathering enough images to train a deep computer vision model is a constant challenge. Unfortunately, collecting images from unknown sources can leave your model s behavior at risk of being manipulated by a dirty-label or clean-label attack unless the images are properly inspected. Manually inspecting each image-label pair is impractical and common poison-detection methods that involve re-training your model can be time consuming. This research uses GAN discriminators to protect a single class against mislabeled and different levels of modified images. The effect of said perturbation on a basic convolutional neural network classifier is also included for reference. The results suggest that after training on a single class, GAN discriminator s confidence scores can provide a threshold to identify mislabeled images and identify 100% of the tested poison starting at a perturbation epsilon magnitude of 0.20, after decision threshold calibration using in-class samples. Developers can use this report as a basis to train their own discriminators to protect high valued classes in their CV models.