Improving robustness of image recognition through artificial image augmentation

Abstract

Deep learning based computer vision technologies can offer a number of advantages over manual labour inspection methods such as reduced operational costs and efficiency improvements. However, they are known to be unreliable in certain situations, especially when input images contain augmentations such as occlusion or distortion that computer vision models have not been trained on. While augmentations can be mitigated by controlling some situations, this is not always possible, especially in outdoor environments.

To address this issue, one common approach is supplemental robustness training using augmented training data, which involves training models on images containing the expected augmentations to improve performance. However, this approach requires collection of a substantial volume of augmented images for each expected augmentation, making it time-consuming and costly depending on the difficulty involved in reproducing each augmentation. This thesis explores the viability of using artificially rendered augmentations on unaugmented images as a substitute for the manual collection and preparation of naturally augmented data for image recognition and object detection models. Specifically, this thesis recreates nine environmental augmentations that commonly occur within outdoor environments and evaluates their impact on model performance on three datasets. The findings of this thesis indicate potential for using artificially generated augmentations as substitutes for naturally occurring augmentations. It is anticipated that further research in this area will enable more reliable image recognition and object detection in less controllable environments, thus improving the results of these technologies in uncertain situations.